| 1 | # Changelog Archive |
| 2 | |
| 3 | Older releases of CodeWhale (v0.8.39 and earlier). Recent releases live in [CHANGELOG.md](../CHANGELOG.md). |
| 4 | |
| 5 | ## [0.8.39] - 2026-05-17 |
| 6 | |
| 7 | ### Fixed |
| 8 | |
| 9 | - **Feishu/Lark bridge startup order is guarded.** The bridge now keeps |
| 10 | `ThreadStore` initialized before startup opens persisted thread state, with a |
| 11 | regression test to prevent moving it below its first use. |
| 12 | - **`/model` picker opens instantly with the curated list again.** Reverted |
| 13 | the v0.8.38 live-catalog rework: the picker no longer makes a blocking |
| 14 | network call on open and once again shows the curated `auto` / |
| 15 | `deepseek-v4-pro` / `deepseek-v4-flash` rows. The `/models` command still |
| 16 | lists the live provider catalog. |
| 17 | - **"Approve for session" groups by command family again.** Session approvals |
| 18 | are keyed by a lossy, arity-aware fingerprint once more, so approving |
| 19 | `cargo build` also covers `cargo build --release`. Denials keep the exact |
| 20 | per-call fingerprint from #1617, so denying one call no longer over-blocks |
| 21 | later, different calls to the same tool. |
| 22 | - **Docker first-run state directories are writable.** The image now |
| 23 | pre-creates `/home/deepseek/.deepseek` with `deepseek` ownership so the |
| 24 | documented named-volume launch can create runtime thread state on first use |
| 25 | (#1684). |
| 26 | - **Runtime API system prompt overrides survive the first turn.** Threads |
| 27 | created with a `system_prompt` override now keep that prompt through |
| 28 | mode/context refreshes before the model request is built (#1688). |
| 29 | - **Compaction keeps a user text query in tool-heavy histories.** Automatic |
| 30 | compaction now pins the latest user text message when the retained tail only |
| 31 | contains tool calls/results, avoiding OpenAI-compatible Jinja template |
| 32 | failures on the next request (#1704). |
| 33 | - **Pager jumps land at the visible bottom.** Pressing `G` or End in the pager |
| 34 | no longer overshoots the render clamp, so `k`/Up scrolls upward immediately |
| 35 | afterward, and mouse wheels now scroll pager overlays directly (#1706, |
| 36 | #1716). |
| 37 | - **Mouse-wheel-as-arrow scrolling preserves composer drafts.** When |
| 38 | `composer_arrows_scroll` is enabled, Up/Down now scroll the transcript even |
| 39 | with text in the composer instead of replacing the draft with input history |
| 40 | (#1677). |
| 41 | - **Multiline composer arrows move between input lines.** Plain Up/Down now |
| 42 | move the cursor within multiline drafts before falling back to input history, |
| 43 | while single-line mouse-wheel-as-arrow scrolling remains unchanged (#1721). |
| 44 | - **Third-party `reasoning_content` streams no longer corrupt text output.** |
| 45 | Generic OpenAI-compatible providers that stream answer text in |
| 46 | `reasoning_content` now render it as normal text unless the selected provider |
| 47 | is one whose reasoning-content semantics are supported (#1673). |
| 48 | - **macOS system theme detection recognizes Light mode.** When `COLORFGBG` is |
| 49 | missing or unusable, `theme = "system"` now falls back to macOS appearance |
| 50 | detection and treats a missing `AppleInterfaceStyle` key as Light mode |
| 51 | (#1670). |
| 52 | - **`rlm_open` accepts schema-filled blank source fields.** Empty `file_path`, |
| 53 | `content`, and `url` strings now count as absent, so calls that provide one |
| 54 | real source no longer fail the exactly-one-source validator (#1712). |
| 55 | - **Resize keeps transcript paging usable immediately.** After a terminal |
| 56 | resize, PageUp/PageDown now use the resized viewport height instead of |
| 57 | falling back to one-line jumps before the next render (#1724). |
| 58 | - **ACP responses stringify JSON-RPC ids.** `serve --acp` now returns string |
| 59 | ids even when clients send numeric ids, matching Zed's stricter ACP client |
| 60 | expectations (#1696). |
| 61 | |
| 62 | ### Thanks |
| 63 | |
| 64 | Thanks to **Matt Van Horn ([@mvanhorn](https://github.com/mvanhorn))** for the |
| 65 | Docker first-run permission fix in #1699 and the runtime system-prompt |
| 66 | regression tests harvested from #1702. Thanks to **Kristopher Clark |
| 67 | ([@krisclarkdev](https://github.com/krisclarkdev))** for the compaction |
| 68 | user-query preservation fix in #1704. Thanks to **Stephen Xu |
| 69 | ([@wlon](https://github.com/wlon))** for the pager jump-bottom fix in #1706. |
| 70 | Thanks to **tdccccc ([@tdccccc](https://github.com/tdccccc))** for the |
| 71 | composer scroll fix in #1715 and pager mouse-wheel support in #1716. |
| 72 | Thanks to **Paulo Aboim Pinto |
| 73 | ([@aboimpinto](https://github.com/aboimpinto))** for the multiline composer |
| 74 | arrow navigation tests harvested from #1719. Thanks to **LittleBlacky |
| 75 | ([@LittleBlacky](https://github.com/LittleBlacky))** for the provider-gated |
| 76 | `reasoning_content` stream fix in #1680. |
| 77 | Thanks to **Eosin Ai ([@Aitensa](https://github.com/Aitensa))** for the macOS |
| 78 | system appearance fallback in #1674. |
| 79 | Thanks to **Anaheim ([@AnaheimEX](https://github.com/AnaheimEX))** for the |
| 80 | `rlm_open` schema validation report in #1712. |
| 81 | Thanks to **THatch26 ([@THatch26](https://github.com/THatch26))** for the |
| 82 | terminal resize paging fix in #1724. |
| 83 | Thanks to **Alvin ([@alvin1](https://github.com/alvin1))** for the Zed ACP id |
| 84 | compatibility report in #1696. |
| 85 | |
| 86 | ## [0.8.38] - 2026-05-15 |
| 87 | |
| 88 | ### Changed |
| 89 | |
| 90 | - **Update guidance is clearer on the website.** The homepage and install page |
| 91 | now surface `deepseek update` while keeping package-manager update paths |
| 92 | visible for Homebrew, npm, and Cargo installs. |
| 93 | - **README setup docs are current again.** The English, Simplified Chinese, |
| 94 | and Japanese READMEs now use the current Docker volume/workspace invocation, |
| 95 | document update paths, list the current provider/model switching surface, and |
| 96 | send release-specific feature notes back to the changelog. |
| 97 | |
| 98 | ### Fixed |
| 99 | |
| 100 | - **OpenAI-compatible providers receive stricter request bodies.** Fireworks |
| 101 | requests now use `reasoning_effort` without the DeepSeek/Anthropic-style |
| 102 | top-level `thinking` field, and chat tool schemas no longer include |
| 103 | Anthropic-only metadata such as `allowed_callers`, `defer_loading`, or |
| 104 | `input_examples` (#1592). |
| 105 | - **pnpm global installs no longer hang in optional postinstall.** pnpm |
| 106 | postinstall now skips install-time binary downloads and leaves the existing |
| 107 | runtime downloader to verify or fetch binaries on first run (#1637). |
| 108 | - **Terminal modes are restored on early TUI exits.** A cleanup guard now |
| 109 | restores raw mode, alternate screen, focus events, mouse capture, bracketed |
| 110 | paste, keyboard flags, and cursor visibility if startup returns early after |
| 111 | terminal initialization (#1593, #1582). |
| 112 | - **Wrapped OSC 8 links keep their full target.** Long clickable URLs now |
| 113 | reopen the original full link target on each wrapped visual chunk instead of |
| 114 | exposing truncated hyperlink targets (#1577). |
| 115 | - **Provider-selected models survive startup and picker reselects.** The |
| 116 | `/model` picker now uses live provider model catalogs when available, saved |
| 117 | default providers sync into the runtime config before the first request, and |
| 118 | reselecting the active provider from the picker keeps the current model |
| 119 | instead of falling back to the provider default (#1632). |
| 120 | - **OpenAI-compatible batch tool calls keep all start events.** Streaming |
| 121 | responses with multiple `tool_calls` in one assistant message now preserve |
| 122 | every tool-use block instead of pairing many tool results with only the last |
| 123 | tool start event (#1686). |
| 124 | - **Diagnostics tool schemas include an empty `required` list.** The built-in |
| 125 | `diagnostics` tool now sends `required: []` with its empty object schema so |
| 126 | DeepSeek no longer rejects it as a null required array (#1685). |
| 127 | - **Windows wheel-as-arrow scrolling works with mouse capture enabled.** |
| 128 | `composer_arrows_scroll` now defaults on for Windows terminals even when |
| 129 | mouse capture is enabled, so wheel events that arrive as arrow keys scroll the |
| 130 | transcript instead of cycling composer history (#1578). |
| 131 | - **Plain Windows PowerShell / ConHost uses calmer rendering.** Unmarked |
| 132 | legacy Windows console hosts now automatically enable low-motion rendering, |
| 133 | disable fancy animations, and resolve `synchronized_output = "auto"` to off |
| 134 | so streaming redraws do not overlap or visibly flicker (#1590). |
| 135 | - **LoopGuard blocks now count as failed tool calls.** Identical tool-call |
| 136 | blocks now return a failed tool result instead of a success, so repeated |
| 137 | blocked checklist/tool retries can trip the existing failure warning and halt |
| 138 | path instead of spinning indefinitely (#1574). |
| 139 | - **Denied tool approvals are scoped to the exact call.** Denying one |
| 140 | write/shell approval now caches the canonical argument fingerprint instead of |
| 141 | a lossy tool/prefix key, so later calls to the same tool with different |
| 142 | arguments can still be reviewed and approved (#1617). |
| 143 | |
| 144 | ### Thanks |
| 145 | |
| 146 | Thanks to **DC ([@duanchao-lab](https://github.com/duanchao-lab))** for the |
| 147 | terminal cleanup-guard idea harvested from #1630, and **imkingjh999 |
| 148 | ([@imkingjh999](https://github.com/imkingjh999))** for the provider/model |
| 149 | switching fixes harvested from #1642. Thanks to **Photo |
| 150 | ([@eng2007](https://github.com/eng2007))** for the provider-aware `/model` |
| 151 | picker catalog work harvested from #1201. Thanks to **hexin |
| 152 | ([@h3c-hexin](https://github.com/h3c-hexin))** for the OpenAI batch tool-call |
| 153 | streaming fix in #1686. Thanks to **chennest |
| 154 | ([@chennest](https://github.com/chennest))** for the diagnostics schema report |
| 155 | in #1685. Thanks to |
| 156 | **[@kunpeng-ai-lab](https://github.com/kunpeng-ai-lab)** for the Windows |
| 157 | composer scroll fix harvested from #1578, and **WuMing |
| 158 | ([@asdfg314284230](https://github.com/asdfg314284230))** for the Windows |
| 159 | PowerShell flicker fix harvested from #1591. Thanks to |
| 160 | **[@maker316](https://github.com/maker316)** for the LoopGuard/checklist loop |
| 161 | report in #1574. Thanks to **lalala |
| 162 | ([@lalala-233](https://github.com/lalala-233))** for the approval denial |
| 163 | regression report in #1617, and **Nightt |
| 164 | ([@nightt5879](https://github.com/nightt5879))** for the exact-call approval |
| 165 | key work harvested from #1624. |
| 166 | |
| 167 | ## [0.8.37] - 2026-05-14 |
| 168 | |
| 169 | ### Added |
| 170 | |
| 171 | - **Tencent Lighthouse + Feishu/Lark bridge setup.** Added a `/opt/whalebro` |
| 172 | Lighthouse runbook, systemd deploy assets, a long-connection Feishu/Lark |
| 173 | bridge, a bridge config validator, and a VPS doctor for runtime, Node, |
| 174 | binaries, env, systemd, and localhost health checks. |
| 175 | - **Tencent Cloud remote-first onboarding.** Documented the CNB + Lighthouse + |
| 176 | Feishu/Lark + optional EdgeOne teaching path and added non-active CNB deploy |
| 177 | templates for a future Lighthouse deploy button. Feishu/Lighthouse branches |
| 178 | are now mirrored to CNB for Tencent-first bootstrap. |
| 179 | - **Homebrew tap automation is release-gated.** The release workflow can update |
| 180 | `Hmbown/homebrew-deepseek-tui` from the checksum manifest when a tap token is |
| 181 | configured, and skips cleanly before downloading release assets when no tap |
| 182 | token exists. |
| 183 | |
| 184 | ### Changed |
| 185 | |
| 186 | - **Bing is the default `web_search` backend.** DuckDuckGo remains selectable |
| 187 | with `[search] provider = "duckduckgo"` and keeps its Bing fallback path. |
| 188 | |
| 189 | ### Fixed |
| 190 | |
| 191 | - **First-run onboarding stays usable without an API key.** Missing-key startup |
| 192 | no longer aborts the TUI before onboarding can collect provider settings. |
| 193 | - **Streamable HTTP MCP sessions keep their server-issued session ID.** Custom |
| 194 | headers also apply to GET preflight requests, fixing authenticated MCP |
| 195 | servers that require both. |
| 196 | - **DeepSeek model completions use canonical IDs.** Alias completions now |
| 197 | resolve to stable DeepSeek model names before being written to config. |
| 198 | - **Terminal and child-process reliability is tighter.** Signal shutdown now |
| 199 | restores the terminal, child tasks preserve proxy environment variables, and |
| 200 | Windows Enter / CSI-u input handling avoids the prior event mismatch. |
| 201 | - **Long terminal text wraps instead of overflowing.** Streaming output, diff |
| 202 | rendering, and the pager now hard-wrap overlong no-whitespace and CJK runs. |
| 203 | - **Release and platform edges are safer.** The TUI no longer trips the Windows |
| 204 | Instant-underflow test path, unsupported desktop targets compile the external |
| 205 | URL opener, and legacy DeepSeek CN provider aliases deserialize to the |
| 206 | canonical DeepSeek provider. |
| 207 | - **Footer diagnostics are less cryptic.** Prefix-cache stability is no longer |
| 208 | shown in the default footer, and the opt-in `/statusline` chip now says |
| 209 | `cache prefix 100%` instead of the ambiguous `P 100%`. |
| 210 | - **Feishu/Lark bridge dependency installs are locked and audited.** The |
| 211 | bridge now ships a package lock, installs with `npm ci` on Lighthouse when |
| 212 | available, and overrides the Lark SDK's transitive `axios` dependency to a |
| 213 | patched line. |
| 214 | - **China-friendly update fallback.** `deepseek update` now supports mirrored |
| 215 | release assets through `DEEPSEEK_TUI_RELEASE_BASE_URL` plus |
| 216 | `DEEPSEEK_TUI_VERSION`, and its network-failure hints point users behind |
| 217 | GitHub-blocking networks to the CNB `cargo install --git` path for both |
| 218 | shipped binaries. |
| 219 | - **CNB is the default Tencent release-candidate mirror.** The CNB sync |
| 220 | workflow now mirrors Feishu/Lighthouse release branches, so Tencent |
| 221 | Lighthouse bootstrap can use CNB before the release branch merges. |
| 222 | |
| 223 | ### Thanks |
| 224 | |
| 225 | Thanks to **ZzzPL ([@Oliver-ZPLiu](https://github.com/Oliver-ZPLiu))** for |
| 226 | the MCP Streamable HTTP and Homebrew automation fixes (#1643, #1631), |
| 227 | **Reid ([@reidliu41](https://github.com/reidliu41))** for CI, streaming wrap, |
| 228 | and model-completion fixes (#1603, #1628, #1601), **MidoriKurage |
| 229 | ([@mdrkrg](https://github.com/mdrkrg))** for the onboarding crash fix (#1598), |
| 230 | **Gordon ([@gordonlu](https://github.com/gordonlu))** for the Windows Enter / |
| 231 | CSI-u fix (#1612), **Aitensa ([@Aitensa](https://github.com/Aitensa))** for |
| 232 | the CJK diff/pager wrap fix (#1622), **qiyan233 |
| 233 | ([@qiyan233](https://github.com/qiyan233))** for legacy DeepSeek CN provider |
| 234 | aliases (#1645), **jieshu666 ([@jieshu666](https://github.com/jieshu666))** |
| 235 | for the repaint-flicker reduction (#1563), **Vishnu |
| 236 | ([@Vishnu1837](https://github.com/Vishnu1837))** for terminal restoration on |
| 237 | signals (#1586), and **axobase001 |
| 238 | ([@axobase001](https://github.com/axobase001))** for proxy environment |
| 239 | preservation in child tasks (#1608). |
| 240 | |
| 241 | ## [0.8.36] - 2026-05-14 |
| 242 | |
| 243 | ### Added |
| 244 | |
| 245 | - **The right sidebar can be hidden for copy-friendly terminals.** |
| 246 | `sidebar_focus = "hidden"` (or `Ctrl+Alt+0` for the current session) removes |
| 247 | the Work/Tasks/Agents/Context rail so raw terminal selection cannot copy |
| 248 | sidebar borders alongside transcript text. |
| 249 | |
| 250 | ### Changed |
| 251 | |
| 252 | - **Sub-agent completion handoffs are leaner and more cache-friendly.** |
| 253 | Internal `<deepseek:subagent.done>` sentinels now point to the preceding |
| 254 | human summary line instead of duplicating the summary, elapsed time, and |
| 255 | step count inside JSON sent to the parent model. |
| 256 | - **Prefix stability is visible beside cache telemetry by default.** The |
| 257 | footer now includes the prefix-stability chip in the default status layout, |
| 258 | and low last-request cache hit rates are no longer colored as hard errors |
| 259 | when the system/tool prefix itself is stable. |
| 260 | - **RLM batch helpers now require an explicit independence assertion.** |
| 261 | `sub_query_batch`, `sub_query_map`, and low-level `*_batched` helpers refuse |
| 262 | dependency-unsafe parallel fanout unless callers pass |
| 263 | `dependency_mode="independent"`, and RLM now exposes `sub_query_sequence` |
| 264 | for A-to-B dependent work. |
| 265 | |
| 266 | ## [0.8.35] - 2026-05-13 |
| 267 | |
| 268 | A post-0.8.34 cleanup release focused on prompt hygiene, context-pressure |
| 269 | guidance, and keeping the next release branch clearly separated from the |
| 270 | already-published v0.8.34 tag. |
| 271 | |
| 272 | > **Note on v0.8.34 contributor credits:** Horace Liu ([@liuhq](https://github.com/liuhq)) |
| 273 | > contributed Nix package support and install documentation in the v0.8.34 |
| 274 | > cycle but was inadvertently omitted from that release's changelog. The |
| 275 | > README contributor list and this note correct the record. |
| 276 | |
| 277 | ### Changed |
| 278 | |
| 279 | - **First-turn prompt context is leaner and easier to audit.** The |
| 280 | generated project context pack now ignores hidden tool/cache state, |
| 281 | balances top-level directories before descending, and `/context` |
| 282 | shows named prompt layers instead of a single opaque system blob. |
| 283 | - **Model-visible prompt policy de-conflicted.** The base and mode |
| 284 | prompts no longer forbid useful `deepseek` CLI diagnostics, no |
| 285 | longer require checklists for simple one-step work, and align |
| 286 | long-session compaction guidance around the 60% suggestion threshold. |
| 287 | - **Context-pressure guidance now has one split rule.** Manual |
| 288 | `/compact` suggestions start around 60% during sustained work, while |
| 289 | automatic replacement compaction remains an opt-in hard guardrail near |
| 290 | 80% so DeepSeek V4 prefix-cache economics stay intact. |
| 291 | - **The Tasks sidebar now ages out stale live-tool noise.** Completed |
| 292 | active tool rows linger briefly and then leave the right rail; very old |
| 293 | running shell rows collapse to a single row instead of occupying the |
| 294 | whole Tasks panel. |
| 295 | |
| 296 | ### Fixed |
| 297 | |
| 298 | - **`auto_compact` settings help now reports the real default**, which |
| 299 | has been off since v0.8.11 to avoid unnecessary cache-prefix rewrites. |
| 300 | |
| 301 | ## [0.8.34] - 2026-05-13 |
| 302 | |
| 303 | A polish, terminal-protocol, and internal-cleanup release. The model-facing |
| 304 | surface is stable; this cycle focused on prefix-cache stability metrics, |
| 305 | broader terminal protocol coverage, bundled skills, and shrinking the |
| 306 | mega-files that had grown around the agent loop and TUI. |
| 307 | |
| 308 | ### Added |
| 309 | |
| 310 | - **Prefix-cache stability tracking.** A footer chip surfaces how stable |
| 311 | the cached prefix has been across recent turns (inspired by Reasonix), |
| 312 | so users can spot cache-busting edits before cost climbs. |
| 313 | - **Bundled DeepSeek-native workflow skills.** A starter set of skills |
| 314 | ships in-binary so a fresh install has a usable `/skills` catalog |
| 315 | without external assets. |
| 316 | - **Native Kitty + Ghostty notification protocols.** `OSC 99` (Kitty) |
| 317 | and `OSC 777` (Ghostty) are now first-class alongside the existing |
| 318 | desktop notification fallback. |
| 319 | - **Theme picker with more presets.** Catppuccin, Tokyo Night, Dracula, |
| 320 | and Gruvbox join the built-in palette set; `/theme` now shows a |
| 321 | live picker. |
| 322 | - **Chunked parallel-safe tool execution.** The engine batches |
| 323 | side-effect-free tool calls into a chunked dispatch so independent |
| 324 | reads/searches finish in one turn instead of serialising round-trip |
| 325 | by round-trip. |
| 326 | - **Cancel-all shell jobs.** A single action stops every running |
| 327 | background shell command instead of cancelling them one-by-one. |
| 328 | - **`edit_file` tolerates typographic punctuation drift.** When the |
| 329 | exact-match and leading-whitespace-fuzzy passes both fail and |
| 330 | `fuzz: true` is set, the tool retries with smart quotes (`"`/`"` → |
| 331 | `"`, `'`/`'` → `'`), en/em-dashes (`–`/`—` → `-`), and non-breaking |
| 332 | spaces (U+00A0 → space) normalized to ASCII. Catches the copy-paste |
| 333 | failure mode where a browser or chat client substituted Unicode |
| 334 | punctuation for the ASCII the file actually contains. |
| 335 | |
| 336 | ### Changed |
| 337 | |
| 338 | - **`crates/tui/src/tui/ui.rs` split into focused modules.** The |
| 339 | former 10k-line single-file TUI dispatcher is decomposed into smaller |
| 340 | modules with clearer responsibilities so reviewing a UI change does |
| 341 | not require holding the entire surface in head. |
| 342 | - **`crates/tui/src/core/engine.rs` reduced.** Helper clusters moved |
| 343 | into the existing `core/engine/` submodule directory next to the |
| 344 | turn loop and tool execution code, making the agent-loop core |
| 345 | easier to read end-to-end. |
| 346 | - **Structured tracing on tool dispatch.** Tool entry, exit, duration, |
| 347 | and result/error are emitted through `tracing` events so |
| 348 | `RUST_LOG=engine.tool_execution=debug` produces a coherent timeline |
| 349 | instead of scattered ad-hoc prints. |
| 350 | - **`/init` updates `AGENTS.md` in place** instead of refusing when |
| 351 | the file already exists, so adding new project guidance does not |
| 352 | require manual stitching. |
| 353 | - **Reasoning tokens included in cost calculations**, and the cost |
| 354 | display auto-switches to CNY when the session locale is `zh-Hans`. |
| 355 | - **Stale repo-root development docs removed.** `TAKEOVER_PROMPT.md` |
| 356 | (v0.8.6 era), `PROMPT_ANALYSIS.md`, and the redundant |
| 357 | `DEPENDENCY_GRAPH.md` no longer ship in releases; `docs/ARCHITECTURE.md` |
| 358 | remains the canonical crate-layout reference. |
| 359 | |
| 360 | ### Fixed |
| 361 | |
| 362 | - **Auth keys checked against the saved provider on startup**, so a |
| 363 | stored DeepSeek key is no longer rejected after switching providers |
| 364 | mid-session. |
| 365 | - **Auto router skipped for decisive local routes**, removing an |
| 366 | extra model round-trip on prompts the dispatcher can route directly. |
| 367 | - **Reasoning content stripped for generic providers** that do not |
| 368 | understand the `reasoning_content` field, preventing HTTP 400s when |
| 369 | pointing at an OpenAI-compatible gateway that lacks DeepSeek |
| 370 | thinking semantics. |
| 371 | - **`FocusGained` debounced** so terminals (Tabby) that emit rapid |
| 372 | focus events no longer trigger a repaint flicker loop. |
| 373 | - **MCP HTTP transport defaults `Accept: application/json, |
| 374 | text/event-stream`** and persists `Mcp-Session-Id` across requests, |
| 375 | matching the spec for resumable streams. |
| 376 | - **Shell output tail preserved when truncating**, so the last lines |
| 377 | of a long command output (usually the error trailer) survive the |
| 378 | in-transcript summary. |
| 379 | - **Prefix cache preserved while pruning tool results.** Old |
| 380 | side-effect tool payloads no longer invalidate the prefix that |
| 381 | the next turn would otherwise reuse. |
| 382 | - **Review sub-agents prevented from spawning further sub-agents** |
| 383 | (#1489), keeping recursive depth bounded. |
| 384 | - **Help overlay closes cleanly** and repaints without a stale frame. |
| 385 | - **Pinyin `/skills` alias dispatched correctly** so Chinese-locale |
| 386 | users reach the same surface. |
| 387 | - **VTE flicker terminals get reduced motion** by default to avoid |
| 388 | thrashing on terminals that mishandle frequent partial redraws. |
| 389 | - **Composer border no longer shows the derived session title**, keeping |
| 390 | the composer chrome reserved for editor and mode state. |
| 391 | |
| 392 | ## [0.8.33] - 2026-05-12 |
| 393 | |
| 394 | A sub-agent and RLM renovation release. The model-facing delegation |
| 395 | surface is now session-oriented instead of one-shot: RLM work happens |
| 396 | through `rlm_open` / `rlm_eval` / `rlm_configure` / `rlm_close`, |
| 397 | sub-agent work happens through `agent_open` / `agent_eval` / |
| 398 | `agent_close`, and large outputs can be parked behind typed handles |
| 399 | that the model reads back explicitly with `handle_read`. |
| 400 | |
| 401 | ### Added |
| 402 | |
| 403 | - **Persistent RLM sessions with bounded REPL helpers.** RLM prompts now |
| 404 | use `peek`, `search`, `chunk`, `context_meta`, `sub_query`, |
| 405 | `sub_query_batch`, `sub_query_map`, `sub_rlm`, and |
| 406 | `finalize(value, confidence)` instead of exposing the full parent |
| 407 | context as an ambient variable. |
| 408 | - **Fork-aware sub-agent sessions.** `agent_open` supports named |
| 409 | sessions, `fork_context`, and bounded recursive depth so the parent can |
| 410 | ask for multiple perspectives while preserving prompt-cache-friendly |
| 411 | prefix context where available. |
| 412 | - **Shared `handle_read` storage.** RLM finals, sub-agent transcripts, |
| 413 | and other large structured results can return `var_handle` references |
| 414 | with slice, range, count, and JSONPath projections. |
| 415 | - **Slash-command routing for the new surface.** `/rlm [N] ...` and |
| 416 | `/agent [N] ...` now prompt the assistant to use the persistent tools |
| 417 | instead of the removed foreground RLM operation. |
| 418 | - **Harness-friendly non-interactive exec sessions.** `deepseek exec` |
| 419 | now supports `--resume`, `--session-id`, `--continue`, and |
| 420 | `--output-format stream-json` so backend wrappers such as ClawBench can |
| 421 | keep conversation state and parse one JSON event per line without running |
| 422 | a long-lived server. |
| 423 | - **`/relay` slash command with CJK aliases** (`/接力`). Hands the |
| 424 | assistant a structured handoff prompt for coordinated multi-turn |
| 425 | continuation across sessions. |
| 426 | - **`checklist_write` sidebar rename.** The sidebar focus tab formerly |
| 427 | known as "Plan" / "Todos" is now "Work" — one panel for the active |
| 428 | checklist and optional plan, consistent across all three modes. |
| 429 | - **Grayscale theme.** `/theme grayscale` and |
| 430 | `/set theme grayscale --save` provide a low-opinion black/white palette |
| 431 | for users who want less brand color in the terminal. |
| 432 | |
| 433 | ### Changed |
| 434 | |
| 435 | - **Prompts and docs now teach only the new tool names.** Legacy |
| 436 | RLM/sub-agent helpers remain internally where needed for durable |
| 437 | transcript compatibility, but the registry exposes the session tools. |
| 438 | - **Large or noisy tool results are easier to keep out of context.** |
| 439 | Tool output summaries, sub-agent results, and transcript snapshots now |
| 440 | point the model toward `handle_read` when it needs raw detail. |
| 441 | - **Tool-surface smoke guidance is explicit.** Release checks now document |
| 442 | the exact version commands and registry-name searches for `handle_read`, |
| 443 | persistent RLM tools, and persistent sub-agent tools. |
| 444 | - **README acknowledgements expanded.** The project thanks OpenWarp and |
| 445 | Open Design for support and collaboration around terminal-agent and |
| 446 | design-forward workflows. |
| 447 | - **Light theme tuned for calmer contrast.** The canvas, panel, elevated, |
| 448 | border, and selection tokens now separate surfaces without the washed-out |
| 449 | white-on-white feel. |
| 450 | - **Session picker is history-first.** `/sessions` and `Ctrl+R` now show |
| 451 | the full selected session history on the left with the session list on |
| 452 | the right; number keys `1`-`9` open visible session histories, `PgUp` / |
| 453 | `PgDn` scroll that history, and `Enter` still resumes. |
| 454 | - **Foreground RLM operation removed.** The old `Op::Rlm` path and its |
| 455 | `handle_rlm` engine method are gone; all RLM work now flows through |
| 456 | the persistent-session tools. |
| 457 | - **Stale competitive-analysis doc removed.** The old cross-agent matrix |
| 458 | had become an unreliable inventory of tool names rather than useful |
| 459 | release guidance. |
| 460 | |
| 461 | ### Fixed |
| 462 | |
| 463 | - **Local/custom endpoints stay prompt-free when auth is optional.** |
| 464 | The dispatcher no longer reads the secret store for SGLang, vLLM, |
| 465 | Ollama, or loopback custom URLs unless API-key auth is explicitly |
| 466 | requested, and the direct TUI treats loopback model endpoints as |
| 467 | no-key by default. This avoids macOS Keychain prompts and stale |
| 468 | DeepSeek keys when users point the app at local OpenAI-compatible |
| 469 | servers. |
| 470 | - **Transcript browsing stays put across resizes.** If the user is reading |
| 471 | older chat history, terminal resize events preserve the current transcript |
| 472 | position instead of jumping back to the live tail; the scrollbar and |
| 473 | jump-to-latest affordance now follow the active theme. |
| 474 | - **Backtrack preview opens near the selected turn.** Pressing Esc twice no |
| 475 | longer opens the live transcript preview at the oldest conversation line; |
| 476 | the highlighted recent user turn is pinned into view, and changing the |
| 477 | backtrack target re-pins only that selection. |
| 478 | - **Completed thinking no longer masquerades as prompt text.** Collapsed |
| 479 | completed reasoning now shows only explicit `Summary:` content inline; raw |
| 480 | reasoning remains available through Ctrl+O/transcript instead of appearing |
| 481 | as assistant self-talk in the main flow. When Ctrl+O starts from a reasoning |
| 482 | block, it opens a full-session reasoning timeline instead of a single |
| 483 | isolated chunk. |
| 484 | - **Transcript selection keeps working while the agent is streaming.** |
| 485 | The loading-state mouse filter now drops inert move events but allows |
| 486 | active transcript and scrollbar drags to continue (reported as a known |
| 487 | issue in v0.8.32). |
| 488 | - **Empty-composer arrow scrolling feels less twitchy.** When configured to |
| 489 | scroll the transcript, plain Up/Down now move by a small wheel-like step |
| 490 | instead of a single-line flick. |
| 491 | - **Mouse and trackpad scrolling feel less sticky in long logs.** Rapid |
| 492 | same-direction transcript scrolls now get bounded acceleration while |
| 493 | direction changes reset to precise single-line movement. |
| 494 | - **RLM smoke-test papercuts fixed.** `rlm_eval` now binds `content` as a |
| 495 | convenience alias for `_context`, tolerates common `timeout_secs` keyword |
| 496 | guesses on child-query helpers while preserving session-level timeout |
| 497 | policy, and stores JSON-serializable `finalize(...)` values as JSON handles |
| 498 | so `handle_read` can project them directly. |
| 499 | - **RLM REPL uses the shared Python resolver.** RLM startup now tries |
| 500 | `python3`, `python`, and `py -3`, matching the dependency resolver used by |
| 501 | code execution and avoiding Windows failures where `python3` is absent |
| 502 | (harvested from PR #1540). |
| 503 | - **Session titles and history previews hide metadata noise.** Saved |
| 504 | session titles and the picker history strip leading `<turn_meta>` envelopes |
| 505 | and thinking-tag blocks so historical conversations read like user-visible |
| 506 | chat rather than prompt plumbing (harvested from PR #1510). |
| 507 | - **Companion binary version smoke is unambiguous.** `deepseek-tui --version` |
| 508 | now reports the `deepseek-tui` binary name instead of the dispatcher label. |
| 509 | - **Vision path boundary test is platform-native.** The absolute-path |
| 510 | rejection smoke uses a Windows absolute path on Windows and `/etc/hosts` |
| 511 | elsewhere (harvested from PR #1526). |
| 512 | - **Tool papercuts:** `file_search` has safer default excludes and an |
| 513 | explicit `exclude` option; `grep_files` returns single-line context as |
| 514 | strings; `fetch_url` can project JSON fields and returns headers; |
| 515 | `edit_file` can opt into leading-indentation fuzz; `exec_shell` can |
| 516 | merge stdout/stderr in chronological order; `revert_turn` rejects |
| 517 | no-op snapshot boundaries. |
| 518 | - **CLI reasoning-effort honoured on non-auto exec routes** (PR #1511 |
| 519 | from **@h3c-hexin**). `deepseek -p "..." --reasoning-effort high` now |
| 520 | applies the flag correctly instead of falling back to the config-file |
| 521 | default. |
| 522 | - **Edit-file replacement boundaries clarified** (PR #1516). The tool |
| 523 | description and error messages now make it unambiguous that |
| 524 | `edit_file` is for one clear replacement in one file. |
| 525 | - **Pandoc output validated before probing** (PR #1523). Binary-format |
| 526 | conversions that produce empty or invalid output now surface a clear |
| 527 | error instead of a confusing pandoc stack trace. |
| 528 | - **Running turns can be steered and repainted** (PR #1533, #1537). |
| 529 | Composer input during an active turn no longer stalls; the TUI |
| 530 | redraws the transcript as the agent streams. |
| 531 | - **Tasks and Activity Detail are calmer under load.** The Tasks panel now |
| 532 | keeps live/background/recent activity from double-counting the same shell |
| 533 | or RLM work, groups repeated read/search/checklist noise, and keeps |
| 534 | failures, status, command summaries, and durations visible. Ctrl+O now |
| 535 | opens Activity Detail for selected/live/recent tool work and the reasoning |
| 536 | timeline for thinking blocks, while Alt+V remains the direct tool-detail |
| 537 | pager; the idle footer now advertises that split for the visible activity. |
| 538 | - **npm retry shows timeout hint on first failure** (PR #1538). |
| 539 | Installations behind slow proxies now see a clear "retrying" message |
| 540 | instead of a silent hang. |
| 541 | - **Issue templates improved** (PR #1525 from **@reidliu41**). Bug and |
| 542 | feature-request templates are clearer and easier for new contributors. |
| 543 | |
| 544 | ### Credits |
| 545 | |
| 546 | Thanks to **@reidliu41** (#1525/#1526), **@h3c-hexin** (#1511), |
| 547 | **@xulongzhe** (#1530/#1544), **@tyouter** (#1510), and |
| 548 | **@Duducoco** (#1540) for community contributions in this release. |
| 549 | |
| 550 | ## [0.8.32] - 2026-05-12 |
| 551 | |
| 552 | A "more useful tools" release. v0.8.31 made the tool surface |
| 553 | reliable on every host; v0.8.32 expands it. Anchor is the question |
| 554 | every new contributor asks: "what does the model actually have to |
| 555 | work with?" — and the answer is now closer to "everything you'd |
| 556 | reach for from a shell, including the document formats the real |
| 557 | world uses." Five new tools (`pdf-extract` swap, `js_execution`, |
| 558 | `pandoc_convert`, `image_ocr`, `image_analyze`), ten community PR |
| 559 | harvests targeting model-protocol bugs (vLLM thinking) and UX |
| 560 | papercuts (Shift+Enter on Windows VSCode, mention truncation |
| 561 | splitting CJK codepoints, approval modal hiding the transcript), |
| 562 | and a snapshot-self-disable on workspaces over 2 GB of |
| 563 | non-excluded content so first-turn `git add -A` no longer hangs |
| 564 | the TUI on multi-hundred-GB project directories. |
| 565 | |
| 566 | ### Performance |
| 567 | |
| 568 | - **Move `instructions = [...]`, user memory, and session goal |
| 569 | below the prompt's volatile-content boundary so DeepSeek's KV |
| 570 | prefix cache survives mid-session edits** (harvested from PR |
| 571 | #1345 by **@Duducoco**). Before this change, the per-workspace |
| 572 | `instructions` block, the user memory file (`/memory`), and the |
| 573 | current session goal (`/goal`) were rendered at position 2.5 |
| 574 | in the system prompt — inside the static prefix layer that the |
| 575 | cache hits. Any edit to those files (or any `# foo` |
| 576 | quick-add to memory) busted the cached prefix from that byte |
| 577 | onwards, forcing the next turn to re-tokenize the rest of the |
| 578 | static layer. Relocating them to position 6 (immediately above |
| 579 | the previous-session handoff block) means the cache hit covers |
| 580 | the entire static prefix — mode, project context, env, skills, |
| 581 | context management, compact template — regardless of how often |
| 582 | the user edits their memory file. Skills, context management, |
| 583 | and the compact template stay always-cacheable in the static |
| 584 | layer where they belong. |
| 585 | |
| 586 | ### Removed |
| 587 | |
| 588 | - **Shift-to-bypass-mouse-capture is gone.** The #376 escape-hatch |
| 589 | feature (hold Shift while moving the mouse → temporarily disable |
| 590 | alt-screen mouse capture so terminal-native text selection works, |
| 591 | then re-enable on release) was causing visible scroll/redraw |
| 592 | thrash: every Shift transition flipped the mouse-capture mode AND |
| 593 | pushed a status toast ("Native selection — release Shift to |
| 594 | return" / "Mouse capture restored"). On modern terminals that |
| 595 | honor `xterm-modifyOtherKeys` the toast cycle fired on stray |
| 596 | Shift events and produced what users described as a "scroll |
| 597 | demon." Removing the bypass path entirely: text selection in |
| 598 | alt-screen sessions now goes through the same path as any other |
| 599 | TUI (your terminal's modifier-bypass — typically Option/Alt on |
| 600 | macOS, Shift in some Linux terminals — still works at the |
| 601 | terminal level, this just stops us from second-guessing it). |
| 602 | |
| 603 | ### Fixed |
| 604 | |
| 605 | - **Tool-result spillover and wire-dedup now share a retrieval |
| 606 | namespace, so `retrieve_tool_result` finds what the model was |
| 607 | pointed at.** Two systems used to mint reference blocks |
| 608 | independently — disk spillover keyed by tool-call id |
| 609 | (`~/.deepseek/tool_outputs/<id>.txt`, only above 100 KB) and the |
| 610 | Chat-Completions wire compactor that replaced repeated tool |
| 611 | results with `<TOOL_RESULT_REF sha="…"/>` (any size, keyed by |
| 612 | SHA256 of the content). The SHA refs were impossible to |
| 613 | retrieve: `retrieve_tool_result ref=<sha>` looked in the |
| 614 | tool-call-id directory and 404'd. Worse, the wire dedup fired |
| 615 | on tiny outputs (a 65-byte `gh run view --json` could turn into |
| 616 | a ref the model then chased through three guesses), and the |
| 617 | `[artifact: …]` block emitted by `apply_spillover_with_artifact` |
| 618 | showed `id:` (with `art_` prefix), `path:` (with a slash), and |
| 619 | `tool_call_id:` separately with no indication of which one |
| 620 | `retrieve_tool_result` accepted. Reported by users hitting |
| 621 | "spilled tool result was not found" 4–5 times per session while |
| 622 | polling CI runs. |
| 623 | - The wire compactor now persists deduped content to |
| 624 | `~/.deepseek/tool_outputs/sha_<sha>.txt` on first sighting, |
| 625 | and only dedupes outputs ≥ 1 KiB — tiny results stay inline |
| 626 | on both occurrences instead of becoming a ref the model has |
| 627 | to chase. The `<TOOL_RESULT_REF>` block grew a |
| 628 | literal `retrieve: retrieve_tool_result ref=sha:<sha>` line. |
| 629 | - `retrieve_tool_result` learns five new ref shapes: |
| 630 | `sha:<64-hex>`, bare 64-hex, `art_<tool_call_id>`, |
| 631 | `artifacts/art_<id>.txt`, and absolute paths anywhere under |
| 632 | the session-artifact root. The lookup tries the legacy |
| 633 | spillover dir, the current session's artifact dir, and the |
| 634 | `art_` → legacy fallback in one call. When everything misses, |
| 635 | the error enumerates every candidate path tried and lists |
| 636 | every accepted ref form so the model can correct on the next |
| 637 | attempt instead of guessing blind. |
| 638 | - The `[artifact: …]` block emitted alongside spilled outputs |
| 639 | now includes a literal `retrieve: retrieve_tool_result |
| 640 | ref=art_<id>` line so the model never has to guess between |
| 641 | `id:`, `path:`, and `tool_call_id:`. |
| 642 | - **`<TURN_META_REF sha="…" original_chars="…" />` is gone.** The |
| 643 | legacy wire-level dedup for identical per-turn metadata blocks |
| 644 | emitted an opaque SHA-tagged reference that periodically leaked |
| 645 | into model reasoning and user-visible debug dumps with no |
| 646 | retrieval mechanism — the SHA was an artifact of the cache |
| 647 | optimization, not a content address the model could resolve. |
| 648 | The compactor now emits a self-explanatory |
| 649 | `<turn_meta_unchanged />` marker instead. Same KV-cache |
| 650 | friendliness (bytes in the same prompt position when nothing |
| 651 | changed), zero ambiguity if the marker ever surfaces. |
| 652 | - **"Request cancelled while awaiting approval" now says |
| 653 | *why*.** The approval and user-input handlers used to emit the |
| 654 | same opaque string regardless of source — user Esc, runtime-API |
| 655 | DELETE, parent-agent cancel, or a torn-down channel all |
| 656 | produced an identical error. A new `CancelReason` enum is |
| 657 | latched alongside the `CancellationToken` and surfaced as a |
| 658 | `(reason: …)` suffix: `user cancelled the request`, `request |
| 659 | cancelled by external caller`, `request was preempted by a new |
| 660 | turn`, `engine torn down before approval resolved`. A closed |
| 661 | approval channel reports the teardown race explicitly instead |
| 662 | of just "Approval channel closed." Remaining non-user |
| 663 | cancellation call sites are tracked in #1541. |
| 664 | - **`pandoc_convert` validates binary-output requests before |
| 665 | resolving the `pandoc` binary** (harvested from PR #1523 by |
| 666 | **@muyuliyan**). Hosts without pandoc installed now still get |
| 667 | the intended validation error for `target_format = "docx"` |
| 668 | without `output_path`, which keeps the CI test independent of |
| 669 | runner packages. |
| 670 | - **Mouse movement no longer leaks xterm tracking bytes into the |
| 671 | composer while the model is streaming** (harvested from PR #1533 |
| 672 | by **@Oliver-ZPLiu**, fixes #1529). When mouse capture is |
| 673 | enabled, move/drag events can arrive faster than the TUI needs |
| 674 | them during a streaming turn; those stale movement events are now |
| 675 | discarded while loading, while click/scroll handling remains |
| 676 | active when the TUI is idle. |
| 677 | - **Resize and successful turn completion avoid blank-frame |
| 678 | flicker** (harvested from PR #1537 by **@czf0718**, addresses |
| 679 | #1515 and #1539). Resize now performs the viewport reset, clear, |
| 680 | and redraw inside a single synchronized-output batch. Successful |
| 681 | `TurnComplete` events use the incremental renderer instead of |
| 682 | forcing a clear+redraw every time; full repaint remains reserved |
| 683 | for interrupted/failed turns and periodic resyncs. |
| 684 | - **The npm wrapper shows binary-download remediation on the first |
| 685 | retryable timeout** (harvested from PR #1538 by **@jieshu666**, |
| 686 | addresses #1532). Users behind networks that time out when |
| 687 | fetching GitHub Release assets now see the mirror/proxy/Cargo |
| 688 | alternatives immediately, and the installer points at the fixed |
| 689 | `docs/INSTALL.md#npm-binary-download-times-out` anchor. |
| 690 | - **`edit_file` states its exact-replacement boundary and warns on |
| 691 | multi-match replacements** (from #1516). The tool description and |
| 692 | base prompt now steer structural or intertwined edits toward |
| 693 | `apply_patch` / `write_file`, while keeping legitimate repeated |
| 694 | replacements compatible with an advisory "verify with read_file" |
| 695 | hint. |
| 696 | - **Shift+Enter explicitly steers a running turn.** Plain Enter |
| 697 | while busy continues to queue (the agreed default since |
| 698 | #1331). Pressing **Shift+Enter** during an in-flight turn |
| 699 | routes the draft directly through `engine.steer()` — the same |
| 700 | path Ctrl+Enter already drove — without going through the |
| 701 | queue at all. When the agent is idle, Shift+Enter still inserts |
| 702 | a composer newline as before. Ctrl+Enter remains bound for |
| 703 | terminals that swallow the Shift modifier on Enter at the |
| 704 | protocol level. (Pattern cross-checked against pi-agent's |
| 705 | `streamingBehavior: "steer"` API: see `agent-session.ts:156` — |
| 706 | same `Shift+Enter → steer / Plain Enter → queue` split, just |
| 707 | surfaced as a keypress instead of a programmatic flag.) |
| 708 | - **`exec_shell_wait` default timeout raised from 5 s → 30 s.** |
| 709 | The 5 s default was fine for "did the command exit yet?" polls |
| 710 | but forced models into hand-rolled `while true; do |
| 711 | exec_shell_wait …; sleep 30; done` loops for `gh run watch`, |
| 712 | long `cargo build`s, and `cargo login`-style interactives. 30 s |
| 713 | covers the common case in a single tool call without burning a |
| 714 | whole turn on the timeout boundary; the model can still pass |
| 715 | any value up to 600 s. |
| 716 | - **Snapshots no longer try to index a multi-hundred-GB workspace |
| 717 | on first turn.** Reported by users running `deepseek-tui` inside |
| 718 | project directories with hundreds of GB of content — datasets, |
| 719 | model weights (`.safetensors`, `.gguf`, `.pt`), Docker image |
| 720 | dumps, parquet / arrow caches — where the side-git snapshot |
| 721 | initialization would hang the TUI for minutes or hours while |
| 722 | `git add -A` walked the workspace. v0.8.32 adds a default |
| 723 | 2 GB ceiling on non-excluded workspace content (measured before |
| 724 | any git work, walking the same excludes the snapshot path |
| 725 | already honors). When the cap is exceeded the side repo isn't |
| 726 | initialized; subsequent snapshots are skipped with a clear |
| 727 | WARN-level log line referencing the new |
| 728 | `[snapshots] max_workspace_gb` config knob users can raise (or |
| 729 | set to `0` to disable the cap entirely and restore v0.8.31 |
| 730 | behaviour). The bounded estimator also early-exits past 200k |
| 731 | file entries, so a workspace full of tiny files trips the cap |
| 732 | before paying for a full walk. Pre-existing v0.8.27 fixes for |
| 733 | the growth-over-time angle (#1112: retention cap, mid-session |
| 734 | prune, expanded built-in excludes) continue to apply; this |
| 735 | closes the orthogonal "snapshots-too-big-to-start" path. |
| 736 | - **Toast stack overlay no longer renders on top of the composer |
| 737 | input** (harvested from PR #1485 by **@MeAiRobot**). When a |
| 738 | deferred tool's schema auto-loaded after the model requested |
| 739 | it, the resulting status toast ("Auto-loaded deferred tool |
| 740 | 'edit_file' after model request.") could render at |
| 741 | `footer_area.y - 1` — which on tight layouts is the bottom row |
| 742 | of the composer area, visibly overwriting the start of the |
| 743 | user's typed text. `render_toast_stack_overlay` now clamps |
| 744 | `max_above` to the gap between `composer_area.y + |
| 745 | composer_area.height` and `footer_area.y`, so when the composer |
| 746 | and footer are adjacent the overlay collapses to zero rows and |
| 747 | the toast is suppressed rather than drawn on top. |
| 748 | - **`/sessions` picker highlights the selected row more strongly |
| 749 | in dark terminals** (harvested from PR #1493 by **@reidliu41**). |
| 750 | Previously the selection background was subtle enough to lose |
| 751 | in low-contrast dark themes; keyboard navigation up/down didn't |
| 752 | obviously change which row was active. The selected row now |
| 753 | uses a bolded label on a stronger background so the focused row |
| 754 | reads cleanly across the dark palettes the TUI ships with. |
| 755 | - **TUI input no longer freezes while long-running shell jobs |
| 756 | flood stdout** (#1299, harvested from PR #1494 by |
| 757 | **@CrepuscularIRIS / autoghclaw**). The job-panel refresh path |
| 758 | was calling `full_output()` from inside the `ShellManager` |
| 759 | mutex, which cloned the entire accumulated stdout/stderr buffer |
| 760 | every 2.5 seconds. For browser-automation or large-build jobs |
| 761 | the buffer grew unboundedly; cloning held the mutex for |
| 762 | O(total_bytes) time, starving the `crossterm::event::poll` loop |
| 763 | and dropping keystrokes. The refresh now reads only the last |
| 764 | `max_tail_chars * 4` bytes under the lock (lock hold time is |
| 765 | O(1) regardless of total output volume) and decodes those into |
| 766 | a tail string for display. `stdout_len` / `stderr_len` still |
| 767 | report the true total byte counts so no caller invariant |
| 768 | breaks. Also tightens `take_delta_from_buffer` to slice |
| 769 | `[cursor..total]` inside the lock guard instead of cloning the |
| 770 | whole buffer first, and skips UTF-8 continuation bytes at |
| 771 | `tail_start` so `from_utf8_lossy` never emits a leading U+FFFD |
| 772 | in the job panel. |
| 773 | - **`@`-mention truncation no longer splits multi-byte UTF-8 |
| 774 | sequences** (#1441, harvested from PR #1495 by |
| 775 | **@CrepuscularIRIS / autoghclaw**). When `@`-mentioning a file |
| 776 | larger than 128 KB the composer truncated the buffer at exactly |
| 777 | `MAX_MENTION_FILE_BYTES`, which on CJK / emoji content landed |
| 778 | mid-codepoint and produced a stray U+FFFD at the cut point. The |
| 779 | truncator now uses `str::from_utf8(...).error_len()` to detect |
| 780 | the incomplete-tail case and rounds down to the last valid |
| 781 | codepoint boundary before decoding. Genuinely invalid UTF-8 |
| 782 | files still surface the "file is not UTF-8" error (the rounding |
| 783 | is only applied when the error is an incomplete tail, not a |
| 784 | real decoding failure mid-buffer). |
| 785 | - **vLLM provider: `reasoning_effort = "off"` now actually |
| 786 | disables thinking on Qwen3 / DeepSeek-R1 servers, cutting |
| 787 | TTFT from ~13s to ~270ms** (harvested from PR #1480 by |
| 788 | **@h3c-hexin**). The vLLM branch of `apply_reasoning_effort` |
| 789 | was injecting `thinking: {type: "disabled"}` at the top of |
| 790 | the request body — but vLLM speaks OpenAI's chat-completions |
| 791 | protocol, not Anthropic-native fields, and silently ignored |
| 792 | the directive. The model then emitted a full hidden reasoning |
| 793 | trace into the non-standard `reasoning` field (which this |
| 794 | client doesn't surface), so users saw a multi-second freeze |
| 795 | before any content token arrived. The vLLM branch now emits |
| 796 | the OpenAI extension `chat_template_kwargs.enable_thinking` |
| 797 | (which vLLM forwards into the model's chat template — the |
| 798 | canonical way to toggle Qwen3's `<think>...</think>` mode). |
| 799 | Measurement against vLLM + Qwen3.6-35B-A3B-FP8: TTFT |
| 800 | 13039ms → 274ms, total LLM call 13s → 5.7s. The `high` / |
| 801 | `max` effort levels likewise switch to the OpenAI extension. |
| 802 | No change for non-vLLM providers. |
| 803 | - **`/sessions` picker no longer shows `<turn_meta>` as the |
| 804 | session title** (harvested from PR #1498 by **@wdw8276**). |
| 805 | `session_manager::create_saved_session_with_id_and_mode` |
| 806 | picked the first text content block off the user message via |
| 807 | `find_map`; the engine prepends an internal `<turn_meta>` block |
| 808 | ahead of the real user text, so the picker rendered that |
| 809 | metadata blob as the session name. Guard added so titles fall |
| 810 | through to the actual user input. Existing sessions without |
| 811 | the prefix block are unaffected. |
| 812 | - **Kitty keyboard protocol now activates on Windows (VSCode + |
| 813 | Windows Terminal), so `Shift+Enter` inserts a newline instead |
| 814 | of submitting** (#1359, harvested from PR #1483 by |
| 815 | **@CrepuscularIRIS / autoghclaw**). Root cause: crossterm's |
| 816 | `PushKeyboardEnhancementFlags` gates the escape sequence on |
| 817 | `is_ansi_code_supported()`, which on Windows queries the |
| 818 | console mode rather than the VT capability and unconditionally |
| 819 | returns false — so the Kitty push (`\x1b[>1u`) was never |
| 820 | written, leaving xterm.js in legacy mode where `Shift+Enter` |
| 821 | and `Enter` both produce `\r` and are indistinguishable. |
| 822 | `Alt+Enter` / `Ctrl+J` were affected the same way. The fix |
| 823 | writes the push and pop escapes directly under `#[cfg(windows)]`, |
| 824 | bypassing the capability gate; terminals that don't speak the |
| 825 | protocol silently discard the sequences. Also extends the |
| 826 | pop-on-exit path to two missed call sites (the `main.rs` panic |
| 827 | hook and `external_editor.rs::spawn_editor_for_input`) so a |
| 828 | crash or `$EDITOR` invocation can no longer leave the parent |
| 829 | shell's keyboard state corrupted. |
| 830 | - **Approval modal can be collapsed to a one-line banner with |
| 831 | Tab** (harvested from PR #1455 by **@tiger-dog**). Previously the |
| 832 | approval prompt rendered as a full-screen takeover that hid the |
| 833 | transcript behind it, so users had to dismiss the modal just to |
| 834 | remember which tool call they were being asked to approve. Tab |
| 835 | now toggles between the takeover card and a single-line bottom |
| 836 | banner — the rest of the conversation stays visible while the |
| 837 | decision is pending. Tab again restores the full card; the |
| 838 | selection state is preserved across the toggle. |
| 839 | - **Markdown renderer no longer eats underscores inside |
| 840 | identifiers** (harvested from PR #1455 by **@tiger-dog**). The |
| 841 | inline parser was matching `_italic_` against the underscore in |
| 842 | `deepseek_tui` / `foo_bar_baz` and rendering the second half of |
| 843 | the identifier in italic, which made transcript snippets that |
| 844 | named code symbols read as garbled prose. Both `_italic_` and |
| 845 | `*italic*` now apply a CommonMark-style boundary check on the |
| 846 | closing delimiter — when the next character is a letter, digit, |
| 847 | or underscore, the delimiter is treated as literal text instead |
| 848 | of markup. Regression-pinned with cases like `crate deepseek_tui |
| 849 | handles approvals` and `look at *not_emphasised*tail`. |
| 850 | |
| 851 | ### Added |
| 852 | |
| 853 | - **npm wrapper installs cleanly on OpenHarmony / HarmonyPC** |
| 854 | (#1072, harvested from PR #1499 by **@CrepuscularIRIS / |
| 855 | autoghclaw**). `os.platform()` returns `openharmony` on |
| 856 | HarmonyPC and on OpenHarmony's Linux ABI-compatible userspace, |
| 857 | but the npm wrapper's platform-asset matrix only covered |
| 858 | `linux` / `darwin` / `win32`, so `npm i -g deepseek-tui` would |
| 859 | abort with `Unsupported platform: openharmony` even though the |
| 860 | Linux x64 / arm64 binaries run unchanged on that environment. |
| 861 | Added a `PLATFORM_ALIASES` mapping that resolves `openharmony` |
| 862 | to the `linux` asset family before lookup so install succeeds |
| 863 | on those hosts. The error message for genuinely unsupported |
| 864 | platforms still reports the raw platform name (`freebsd`, |
| 865 | etc.) so OS-mismatch reports stay diagnostic. |
| 866 | - **Startup empty-state shows useful context instead of |
| 867 | repeating the header** (harvested from PR #1444 by |
| 868 | **@reidliu41**). The center of the welcome view used to repeat |
| 869 | information already displayed in the header and footer. It now |
| 870 | shows the build version, the active model with a `/model` |
| 871 | hint, and the current working directory so first-time users |
| 872 | have somewhere to look while they decide what to type. |
| 873 | - **Opt-in `v4-best-practices` bundled skill** (harvested from |
| 874 | PR #1448 by **@SamhandsomeLee**). A single 50-line `SKILL.md` |
| 875 | encoding three V4-specific workflow rules for multi-step |
| 876 | thinking-mode tasks. Each rule maps to a concrete observable |
| 877 | failure class. Discovered through the existing |
| 878 | `crates/tui/assets/skills/...` mechanism alongside the |
| 879 | `skill-creator` skill — not enabled by default; users opt in |
| 880 | via the standard `/skills` UI. |
| 881 | - **`image_analyze` tool — vision-model image understanding** |
| 882 | (harvested from PR #1467 by **@MMMarcinho**). Sends an image |
| 883 | file to an OpenAI-compatible vision endpoint and returns the |
| 884 | model's natural-language description. Complements `image_ocr`: |
| 885 | use `image_ocr` for "what text is on this image", `image_analyze` |
| 886 | for "what is this image about". **Opt-in only** — gated by both |
| 887 | the `[features] vision_model = true` flag and a `[vision_model]` |
| 888 | config block specifying `model` (and optionally `api_key` / |
| 889 | `base_url`). Default configuration ships the feature flag at |
| 890 | `false`, so no install sees vision API calls fire without an |
| 891 | explicit two-step opt-in. **Billing**: each call hits the |
| 892 | configured vision endpoint (OpenAI by default), so usage is |
| 893 | billed by the third-party provider; calls are stateless (no |
| 894 | conversation context attached). Workspace-boundary check: the |
| 895 | tool rejects absolute paths and any `..` parent-dir traversal |
| 896 | before any base64 encoding or API call. To disable later: set |
| 897 | `[features] vision_model = false` (or omit `[vision_model]`). |
| 898 | Supports PNG, JPEG, GIF, WebP, and BMP inputs. |
| 899 | - **`image_ocr` tool — extract text from images via local |
| 900 | tesseract.** Lets the model OCR a screenshot, scanned receipt, |
| 901 | whiteboard photo, or image-only PDF the user drops into the |
| 902 | workspace, without bouncing through `exec_shell`. Spawns |
| 903 | `tesseract <image> -` and returns the recognised text inline; |
| 904 | no file is written. PNG / JPEG / TIFF inputs supported. |
| 905 | Registration is gated on `dependencies::resolve_tesseract()`; |
| 906 | when tesseract is missing the tool isn't advertised, so the |
| 907 | model never tries to call an OCR engine the host can't run. |
| 908 | `deepseek doctor` reports tesseract status alongside the other |
| 909 | external-binary dependencies with platform-aware install hints |
| 910 | (`brew install tesseract` / `apt install tesseract-ocr` / |
| 911 | `winget install UB-Mannheim.TesseractOCR`). For non-default |
| 912 | language packs or PSM modes, users can still drop into |
| 913 | `exec_shell` with the full tesseract CLI surface. |
| 914 | - **`pandoc_convert` tool — convert documents between formats via |
| 915 | the local pandoc binary.** Pandoc is the Swiss Army knife the |
| 916 | real world uses for moving prose around — Markdown to HTML, |
| 917 | HTML to Markdown, reST to anything, anything to DOCX / EPUB / |
| 918 | LaTeX — and surfacing it as a model-callable tool unblocks |
| 919 | "rewrite this report as ..." / "publish this changelog as ..." |
| 920 | workflows that previously needed the user to drop into a |
| 921 | terminal between turns. Curated target whitelist of 11 formats |
| 922 | (markdown, gfm, commonmark, html, rst, latex, docx, odt, epub, |
| 923 | plain, asciidoc) so the model can't ask for `pdf` (would need |
| 924 | LaTeX) or typos like `markown`. Binary targets (docx, odt, |
| 925 | epub) require an `output_path`; text targets can return the |
| 926 | converted text inline. Approval routes through the WritesFiles |
| 927 | / Suggest tier on every call. Registration is gated on |
| 928 | `dependencies::resolve_pandoc()`; `deepseek doctor` surfaces |
| 929 | the binary's status with platform-aware install hints. |
| 930 | - **`js_execution` tool — execute model-provided JavaScript via a |
| 931 | local Node.js runtime.** Mirrors `code_execution` (Python) so |
| 932 | the model has a single consistent surface for "run this snippet |
| 933 | locally and tell me what it printed" across both interpreters. |
| 934 | Same tempfile-spawn pattern, same 120-second timeout, same |
| 935 | stdout/stderr/return_code result shape — so prompt-cache |
| 936 | layouts that cover one tool also cover the other. Registration |
| 937 | is gated on `crate::dependencies::resolve_node()`: when Node is |
| 938 | missing the tool is simply not advertised, so the model never |
| 939 | sees a runtime it can't actually use. `deepseek doctor` reports |
| 940 | Node availability under "Tool Dependencies" with platform-aware |
| 941 | install hints (`brew install node` / `apt install nodejs` / |
| 942 | `winget install OpenJS.NodeJS`). Approval routes through the |
| 943 | same Suggest tier as `code_execution`. |
| 944 | - **`/translate` opt-in: respond in the user's UI locale, with a |
| 945 | post-hoc fallback for English that leaks through** (harvested |
| 946 | from PR #1462 by **@YaYII**). Two-layer design: when the user |
| 947 | enables translation via the `/translate` slash command, a |
| 948 | `## Language Output Requirement` block is appended to the |
| 949 | system prompt instructing the model to reply in the resolved |
| 950 | session locale (Simplified Chinese, Traditional Chinese, |
| 951 | Japanese, or Brazilian Portuguese — code identifiers and |
| 952 | user-requested English code blocks are exempt). For replies |
| 953 | that still surface English despite the directive, a heuristic |
| 954 | in `tui::translation` (Latin-vs-CJK character ratio with |
| 955 | weighting for CJK information density) detects the leak and |
| 956 | invokes a focused per-message translation API call to render |
| 957 | the localised version before display. Both layers are off by |
| 958 | default and have no effect on installs that don't enable them. |
| 959 | Trust-boundary scope: opt-in only, system prompt addition is |
| 960 | conditional on the runtime flag, no model behaviour change for |
| 961 | English-locale users. |
| 962 | - **AtlasCloud is now a first-class provider** (harvested from |
| 963 | PR #1436 by **@lucaszhu-hue**). AtlasCloud hosts the V4 family |
| 964 | (and other DeepSeek-compatible models) on its own endpoint at |
| 965 | `https://api.atlascloud.ai/v1`, and several contributors had |
| 966 | been running it through the OpenAI-compatible passthrough with |
| 967 | manual `base_url` / model overrides. Selecting |
| 968 | `provider = "atlascloud"` in `~/.deepseek/config.toml` (or via |
| 969 | `DEEPSEEK_PROVIDER=atlascloud`) now wires up the documented |
| 970 | defaults, a `[providers.atlascloud]` config block for per-user |
| 971 | api_key / base_url / model / http_headers overrides, the |
| 972 | `ATLASCLOUD_API_KEY` env var path, and the |
| 973 | provider-picker / `/provider` slash command entries — same |
| 974 | shape as the existing NVIDIA NIM / Fireworks / OpenAI provider |
| 975 | rows. Default remains DeepSeek; nothing changes for installs |
| 976 | that don't opt in. |
| 977 | - **`web_search` supports Tavily and Bocha as configurable |
| 978 | backends** (harvested from PR #1294 by **@sandofree**). DuckDuckGo |
| 979 | with Bing fallback remains the default — no API key required — |
| 980 | but users in regions where those scrapers are unreliable can now |
| 981 | set `[search] provider = "tavily" | "bocha"` plus |
| 982 | `api_key = "..."` in `config.toml` (or via the |
| 983 | `DEEPSEEK_SEARCH_PROVIDER` / `DEEPSEEK_SEARCH_API_KEY` env vars) |
| 984 | to route every `web_search` call through the chosen API. Tavily |
| 985 | is an AI-search API targeted at general use; Bocha is the |
| 986 | mainland-China-friendly equivalent. Trust-boundary pins: an |
| 987 | unset `api_key` on an opted-in provider surfaces a clear |
| 988 | `ToolError` naming the missing key rather than silently falling |
| 989 | through to a different provider, the network policy gate |
| 990 | (`[network]`) is consulted for the provider host on every call, |
| 991 | and the default path is unchanged so no install sees provider |
| 992 | behaviour change unless they explicitly opt in. |
| 993 | - **`/change` slash command** displays the most recent |
| 994 | CHANGELOG.md version section from inside the TUI, so users can |
| 995 | see what they just upgraded into without leaving the chat |
| 996 | (harvested from PR #1416 by **@zhuangbiaowei**). The command |
| 997 | works against the bundled release-notes copy when no workspace |
| 998 | CHANGELOG is available, and on non-English locales it requests |
| 999 | a model-side translation of the section so localised users see |
| 1000 | the changelog in their UI language. Pure offline fallback when |
| 1001 | no API key is configured. |
| 1002 | |
| 1003 | ### Fixed |
| 1004 | |
| 1005 | - **`deepseek update` now refreshes the companion TUI binary |
| 1006 | alongside the dispatcher** (harvested from PR #1492 by |
| 1007 | **@NorethSea**). Closes the documented two-binary footgun: |
| 1008 | `~/.cargo/bin/deepseek` would update to the latest dispatcher, |
| 1009 | but `~/.cargo/bin/deepseek-tui` would stay at the previously |
| 1010 | installed version, so users saw the dispatcher report a new |
| 1011 | release while the TUI runtime they actually interacted with |
| 1012 | reported the old version. Most painful for Volta-managed npm |
| 1013 | installs and any maintainer flow that calls `update` instead of |
| 1014 | re-running both `cargo install --path crates/{cli,tui}`. The |
| 1015 | updater now enumerates colocated binaries up front, downloads |
| 1016 | and verifies every release asset before replacing anything, |
| 1017 | then swaps the sibling first and the running dispatcher last so |
| 1018 | a partial network failure cannot leave the launcher updated |
| 1019 | while the TUI remains stale. |
| 1020 | |
| 1021 | ### Changed |
| 1022 | |
| 1023 | - **`read_file` now extracts PDFs in pure Rust by default — no |
| 1024 | Poppler install required.** Before v0.8.32 the PDF path shelled |
| 1025 | out to `pdftotext` (Poppler), so first-time users on hosts without |
| 1026 | it saw `read_file` return a `binary_unavailable` sentinel and had |
| 1027 | to `brew install poppler` / `apt install poppler-utils` before |
| 1028 | the model could open a PDF. The bundled `pdf-extract` crate |
| 1029 | (which already powered URL-fetched PDFs in `web_run`) now drives |
| 1030 | the local `read_file` path too. The `pages` parameter still |
| 1031 | filters by 1-indexed inclusive page range; both the whole-file |
| 1032 | and per-page variants run with no system dependency. Users with |
| 1033 | column-heavy or complex-table PDFs (academic papers, financial |
| 1034 | filings) where `pdftotext -layout` still wins can opt into the |
| 1035 | external path with `prefer_external_pdftotext = true` in |
| 1036 | `~/.config/deepseek/settings.toml` — when set, the previous |
| 1037 | Poppler dispatch (and the `binary_unavailable` install hint when |
| 1038 | the binary is missing) returns. `deepseek doctor` now reports |
| 1039 | `pdftotext` as optional and explains how to opt in instead of |
| 1040 | framing it as a missing dependency. |
| 1041 | |
| 1042 | ### Known issues |
| 1043 | |
| 1044 | - **Terminal-native text selection can still be blocked while the |
| 1045 | agent is thinking or streaming a response.** v0.8.32 removed the |
| 1046 | noisy Shift-to-bypass-mouse-capture path that caused visible |
| 1047 | scroll/redraw thrash, but the replacement selection path is not |
| 1048 | complete yet. v0.8.33 is planned to ship the text-selection fix |
| 1049 | alongside the sub-agent and RLM renovation. |
| 1050 | |
| 1051 | ## [0.8.31] - 2026-05-12 |
| 1052 | |
| 1053 | A "tools that actually work" release. `code_execution` no longer |
| 1054 | fails on Windows hosts where `python3` isn't on `PATH` — we probe |
| 1055 | for the interpreter at catalog-build time and only advertise the |
| 1056 | tool when one resolves, so the model never sees a runtime it can't |
| 1057 | actually use. The new `deepseek doctor` "Tool Dependencies" and |
| 1058 | "Terminal Quirks" sections surface external-binary status and |
| 1059 | active env-driven overrides so flicker / motion / missing-tool |
| 1060 | puzzles answer themselves before a bug report gets filed. Ptyxis |
| 1061 | 50.x users on Ubuntu 26.04 get a manual `synchronized_output = off` |
| 1062 | knob plus auto-detection that opts them out of the DEC 2026 |
| 1063 | synchronized-output wrap their VTE 0.84 mishandles. The CNB Cool |
| 1064 | mirror workflow is rewritten with concurrency and scoped pushes so |
| 1065 | release tags reliably reach `cnb.cool/deepseek-tui.com/DeepSeek-TUI` |
| 1066 | for users behind GitHub-blocking networks. Plus a new auto-close |
| 1067 | workflow that closes contributor PRs whose code has been harvested |
| 1068 | into `main`, so credit lands at the same moment the fix does. |
| 1069 | |
| 1070 | ### Fixed |
| 1071 | |
| 1072 | - **Windows `exec_shell` preserves MSVC toolchain env** (harvested |
| 1073 | from PR #1487 by **@Jianfengwu2024**). When the parent shell has |
| 1074 | already loaded `VsDevCmd` / `vcvars` (Developer Command Prompt, |
| 1075 | the standard way to run Rust + MSVC on Windows), `exec_shell` was |
| 1076 | stripping `LIB` / `LIBPATH` / `INCLUDE` and the related VS / SDK / |
| 1077 | CRT root variables on its way to the child. That made |
| 1078 | model-driven `cargo build` calls fail to resolve `kernel32.lib` |
| 1079 | even though `link.exe` was reachable via `PATH`. The allowlist |
| 1080 | in `child_env.rs` now preserves the 13 MSVC env vars so the |
| 1081 | toolchain context survives the sanitisation pass. |
| 1082 | - **`code_execution` no longer fails with "program not found" on |
| 1083 | Windows** (and any other host without `python3` on `PATH`). Before |
| 1084 | v0.8.31 the tool hardcoded `python3` and was unconditionally |
| 1085 | advertised in Agent / YOLO modes — so the model would call it, |
| 1086 | spawn would fail, and the error surfaced as a generic tool failure |
| 1087 | with no upstream hint. The fix probes for a Python interpreter |
| 1088 | (`python3` → `python` → `py -3`) at catalog-build time, caches the |
| 1089 | resolved interpreter, and only advertises `code_execution` when one |
| 1090 | resolves. On hosts with no Python the tool is not registered at all |
| 1091 | — the model never sees a tool it can't actually run. Reported by a |
| 1092 | Windows contributor; resolver lives at |
| 1093 | `crates/tui/src/dependencies.rs` and is also surfaced by |
| 1094 | `deepseek doctor`. Folds in the contributor's "write code to a |
| 1095 | tempfile and run the file" suggestion at the same time, so multiline |
| 1096 | code with quote nesting no longer round-trips through `python3 -c`. |
| 1097 | - **Termius and every SSH session auto-enable low-motion** |
| 1098 | (#1433, harvested from PR #1479 by **@CrepuscularIRIS / autoghclaw**). |
| 1099 | Termius desktop sets `TERM_PROGRAM=Termius`; sshd exports |
| 1100 | `SSH_CLIENT` for every TCP session and `SSH_TTY` for interactive |
| 1101 | PTY logins. Any of those signals now flips `low_motion` and |
| 1102 | `fancy_animations` like the existing VS Code / Ghostty path, so |
| 1103 | the 120 FPS cursor-repositioning that races the SSH round-trip |
| 1104 | no longer flickers a remote TUI. Disk-loaded `fancy_animations = |
| 1105 | true` is unconditionally overridden under these signals, |
| 1106 | matching the existing env-precedence contract. |
| 1107 | - **DEC 2026 synchronized output is auto-disabled on Ptyxis** (the new |
| 1108 | default terminal on Ubuntu 26.04 and an increasingly common Linux |
| 1109 | TUI host). Ptyxis 50.x ships on VTE 0.84.x, which parses the |
| 1110 | `\x1b[?2026h` / `\x1b[?2026l` begin/end pair but still flashes the |
| 1111 | entire viewport on every wrapped frame instead of deferring |
| 1112 | rendering — so a TUI that uses DEC 2026 to avoid tearing |
| 1113 | experiences visible flicker on every redraw. gnome-terminal 3.58 |
| 1114 | on the same VTE renders cleanly, so the heuristic must stay narrow: |
| 1115 | we trigger only on `TERM_PROGRAM` matching `ptyxis` |
| 1116 | case-insensitively, or `PTYXIS_VERSION` set to any non-empty value. |
| 1117 | Either signal flips the new `synchronized_output` setting from |
| 1118 | `auto` to `off`; the renderer then skips the begin/end pair on |
| 1119 | every draw, in `reset_terminal_viewport`, and in `resume_terminal`. |
| 1120 | Users on Ptyxis who upgrade past the upstream fix (or who want to |
| 1121 | confirm a fix landed) can override with |
| 1122 | `/set synchronized_output on` or by adding |
| 1123 | `synchronized_output = "on"` to `~/.config/deepseek/settings.toml`. |
| 1124 | |
| 1125 | ### Added |
| 1126 | |
| 1127 | - **`deepseek doctor` now reports tool-dependency status.** A new |
| 1128 | "Tool Dependencies" section lists which external binaries the |
| 1129 | registered tools rely on, with ✓ when present and ✗ + an |
| 1130 | install hint when missing. Today this covers the Python |
| 1131 | interpreter (`code_execution`) and `pdftotext` (`read_file` PDF |
| 1132 | path). A separate "Terminal Quirks" section shows which env-driven |
| 1133 | auto-overrides (VS Code / Ghostty / Termius / SSH / Ptyxis) are |
| 1134 | currently active so users can see at a glance why a particular |
| 1135 | rendering compromise is in effect. Foundation for surfacing future |
| 1136 | tool dependencies as the toolset grows. |
| 1137 | - **New `synchronized_output` setting** controls whether the renderer |
| 1138 | wraps each frame in DEC mode 2026 synchronized output. Accepts |
| 1139 | `auto` (default; respect the Ptyxis env opt-out), `on` (always emit |
| 1140 | DEC 2026, override the heuristic), or `off` (never emit DEC 2026). |
| 1141 | The cost of `off` is brief tearing on terminals that handle DEC |
| 1142 | 2026 cleanly; it is purely a rendering-quality knob, not a |
| 1143 | correctness one. Set via `/set synchronized_output <auto|on|off>` |
| 1144 | or in `~/.config/deepseek/settings.toml`. |
| 1145 | - **`read_file` accepts `start_line` and `max_lines`** for chunked, |
| 1146 | bounded reads of large files (#1450, harvested from PR #1451 by |
| 1147 | **@Oliver-ZPLiu**). Default window is 200 lines / ~16 KB; the hard |
| 1148 | cap is 500 lines. Small files (≤ 200 lines AND ≤ 16 KB) still |
| 1149 | return their contents unchanged, so existing prompts that read |
| 1150 | config files / single source files see no behavior change. Large |
| 1151 | files now return a `<file …>`-wrapped, line-numbered window with |
| 1152 | `shown_lines`, `truncated`, and `next_start_line` attributes plus |
| 1153 | a `[TRUNCATED]` continuation hint — so the model can page through |
| 1154 | a 50 KB file in 16 KB slices instead of dragging the whole thing |
| 1155 | into the conversation context on every turn. PDFs continue to use |
| 1156 | `pages`; `start_line` / `max_lines` apply to text files only. |
| 1157 | - **`web/` dependency security updates.** Bumps: |
| 1158 | - `next` 15.5.16 → 15.5.18 (GHSA-26hh-7cqf-hhc6 — App Router |
| 1159 | middleware/proxy bypass via segment-prefetch routes; high |
| 1160 | severity). |
| 1161 | - `mermaid` 11.14.0 → 11.15.0 (GHSA family: Gantt-chart infinite- |
| 1162 | loop DoS, `classDef` HTML injection, `classDefs` / |
| 1163 | configuration CSS injection; all medium severity). |
| 1164 | - `eslint-config-next` 15.5.16 → 15.5.18 (matches Next.js). |
| 1165 | `npm run build` confirmed clean on the bumped lockfile. None of |
| 1166 | these affect the Rust TUI binary; the bumps are for the |
| 1167 | separately-deployed `deepseek-tui.com` site. |
| 1168 | - **MCP HTTP servers accept custom headers** for authentication |
| 1169 | (#1454, harvested from PR #1456 by **@Oliver-ZPLiu**). Mirrors the |
| 1170 | `headers` field that Claude Code, Codex, and OpenCode already |
| 1171 | accept in their MCP config — add e.g. |
| 1172 | `"headers": { "Authorization": "Bearer ${HF_TOKEN}" }` under any |
| 1173 | HTTP server entry in `~/.deepseek/mcp.json` and the headers are |
| 1174 | sent on every Streamable HTTP request. Headers are sent |
| 1175 | literally — env-var interpolation is a follow-up, so tokens |
| 1176 | pasted directly into mcp.json live there as plain text. The |
| 1177 | Streamable HTTP transport filters out empty keys, framing |
| 1178 | overrides (`Accept`, `Content-Type`), and CR/LF in values |
| 1179 | (response-splitting defense) so a single bad entry can't break |
| 1180 | protocol negotiation or smuggle a header through a misbehaving |
| 1181 | proxy. Stdio servers (`command`-based) and the legacy SSE |
| 1182 | transport ignore the field; SSE coverage is a follow-up. |
| 1183 | |
| 1184 | ## [0.8.30] - 2026-05-11 |
| 1185 | |
| 1186 | A "tighten what we shipped" release. Bare single-letter keystrokes |
| 1187 | (`g`, `G`, `[`, `]`, `?`, `l`, `v`) no longer get eaten as transcript- |
| 1188 | nav shortcuts when the composer is empty — every one of them is now |
| 1189 | freely usable as the first character of a message. The water-spout |
| 1190 | animation in the footer is decoupled from `low_motion` so typewriter |
| 1191 | mode no longer hides the wave, and the v0.3.5-era 🐳→🐋 cycling |
| 1192 | indicator is back next to the effort chip after a long detour through |
| 1193 | geometric dots. Plus a handful of provider, shell, and config fixes |
| 1194 | that surfaced during v0.8.29 testing. |
| 1195 | |
| 1196 | ### Added |
| 1197 | |
| 1198 | - **The whale is back.** Restored the `🐳 → 🐳. → 🐳.. → 🐳... → 🐋 → 🐋. |
| 1199 | → 🐋.. → 🐋... → 🐋.. → 🐋. → 🐳..` cycling status indicator that |
| 1200 | originally shipped in v0.3.5 and silently disappeared in commit |
| 1201 | `1a04659a9` (the "smoother TUI streaming" pass, which swapped the |
| 1202 | 12-frame whale sequence for a 6-frame geometric `◍ ◉ ◌` ring) and then |
| 1203 | was deleted outright in `f4dbf828c` (footer-polish commit). The chip |
| 1204 | renders in the header status cluster, immediately before the |
| 1205 | reasoning-effort chip — exactly where long-time users remember it. |
| 1206 | Idle frame is a steady 🐳; the cycle advances every 420 ms keyed off |
| 1207 | `App::turn_started_at`, so the breaching whale shows up halfway |
| 1208 | through any active turn. |
| 1209 | |
| 1210 | Configurable via the new `status_indicator` setting: |
| 1211 | - `whale` (default) — the historical cycling whale. |
| 1212 | - `dots` — the geometric `◍ ◉ ◌` frames from the dots era. |
| 1213 | - `off` — hide the chip entirely. |
| 1214 | |
| 1215 | Set via `/config status_indicator <whale|dots|off>` or in |
| 1216 | `settings.toml`. |
| 1217 | |
| 1218 | ### Changed |
| 1219 | |
| 1220 | - **Transcript-nav single-letter shortcuts now require `Alt`.** Before |
| 1221 | v0.8.30, pressing a bare `g`, `G`, `[`, `]`, `?`, `l`, or `v` with an |
| 1222 | empty composer hijacked the keystroke for transcript navigation — so |
| 1223 | typing "good morning" produced "ood morning" with no warning, and the |
| 1224 | v0.8.29 spot-fix at `c13ddb04d` (gg double-tap) only suppressed the |
| 1225 | scroll, not the lost character. The bindings are now uniformly |
| 1226 | `Alt+<key>`, mirroring `Alt+R` (history search) and `Alt+V` (tool |
| 1227 | details) which already followed this pattern: |
| 1228 | |
| 1229 | | Old (bare) | New (`Alt+…`) | What it does | |
| 1230 | |---|---|---| |
| 1231 | | `gg` (double-tap) | `Alt+G` | scroll transcript to top | |
| 1232 | | `G` or `Shift+G` | `Alt+Shift+G` | scroll transcript to bottom | |
| 1233 | | `[` | `Alt+[` | jump to previous tool output | |
| 1234 | | `]` | `Alt+]` | jump to next tool output | |
| 1235 | | `?` | `Alt+?` | open the searchable help overlay (F1 / `Ctrl+/` also bound) | |
| 1236 | | `l` | `Alt+L` | open pager for last message | |
| 1237 | | `v` / `V` | `Alt+V` | open tool-details pager | |
| 1238 | |
| 1239 | Plain letters are now always inserted into the composer as text. The |
| 1240 | `App::transcript_pending_g` field from the v0.8.29 half-fix is removed; |
| 1241 | the unified `alt_nav_modifiers` predicate replaces the per-key |
| 1242 | `is_empty()` checks. |
| 1243 | |
| 1244 | ### Fixed |
| 1245 | |
| 1246 | - **`low_motion = true` no longer hides the footer water-spout** when |
| 1247 | `fancy_animations = true`. The spout-strip animation in the footer was |
| 1248 | hard-gated on `!low_motion`, collapsing two unrelated concerns — |
| 1249 | streaming pacing and footer animation — onto one flag. The two are now |
| 1250 | orthogonal: `low_motion` governs only streaming pacing (typewriter vs. |
| 1251 | upstream cadence), and `fancy_animations` alone decides whether the |
| 1252 | water-spout strip renders. The wave itself is unchanged from prior |
| 1253 | releases (wall-clock-driven sine, same cadence as v0.8.29). |
| 1254 | - **Custom-base-URL providers preserve the user's model name** (#857 |
| 1255 | class). Only OpenRouter was previously whitelisted; Sglang, Novita, |
| 1256 | Fireworks, Vllm, Ollama, and NvidiaNim users hitting custom gateways |
| 1257 | with a bare model name were getting HTTP 400s because the dispatcher |
| 1258 | rewrote the model identifier. Now any provider with a user-set |
| 1259 | `base_url` is treated as a custom endpoint and passes the model name |
| 1260 | through unchanged. |
| 1261 | - **`exec_shell` no longer freezes the TUI when a background subprocess |
| 1262 | outlives its parent shell** (#828, cherry-picked from PR #1475 by |
| 1263 | **@CrepuscularIRIS / autoghclaw**). Orphaned children that kept the |
| 1264 | pipe write-end open made `handle.join()` in `collect_output` block |
| 1265 | indefinitely; every transcript-rendering tick that called |
| 1266 | `list_jobs()` then hung the UI. The collector now kills the process |
| 1267 | group before joining the reader threads, and the previously dead |
| 1268 | `cleanup()` is now wired to drop completed jobs older than an hour. |
| 1269 | |
| 1270 | ## [0.8.29] - 2026-05-11 |
| 1271 | |
| 1272 | A maintenance release anchored by a regression fix for the |
| 1273 | "scroll demon" (#1085 class, re-introduced by v0.8.27's flicker |
| 1274 | patch) and a wrong-project session-restore bug (#1395). Plus 25 |
| 1275 | community PRs covering MCP transport, prompt steering, auto-routing |
| 1276 | language coverage, web-search SERP filtering, and broad test |
| 1277 | coverage additions. |
| 1278 | |
| 1279 | ### Fixed |
| 1280 | |
| 1281 | - **Scroll demon — alt-screen no longer drifts under parallel |
| 1282 | sub-agent load** (#1085 regression). The v0.8.27 flicker fix |
| 1283 | dropped the `\x1b[2J\x1b[3J` deep-clear from the viewport-reset |
| 1284 | path, which had been silently masking three `eprintln!` sites |
| 1285 | inside the sub-agent and network-policy modules. Each leak |
| 1286 | scrolled the alt-screen up by one row while ratatui's diff |
| 1287 | renderer remained convinced its model matched reality. Three |
| 1288 | layers of defence now ship together: a `tracing-subscriber` |
| 1289 | writing to `~/.deepseek/logs/tui-YYYY-MM-DD.log`, an fd-level |
| 1290 | `dup2` stderr redirect for the alt-screen lifetime (Unix only; |
| 1291 | Windows follow-up tracked), and module-level |
| 1292 | `#![deny(clippy::print_stdout, clippy::print_stderr)]` on |
| 1293 | `tools/`, `core/`, `tui/`, `runtime_threads.rs`, and |
| 1294 | `network_policy.rs`. The three known leak sites |
| 1295 | (`subagent::persist_state_best_effort`, |
| 1296 | `subagent::new_shared_subagent_manager`, `network_policy::record`) |
| 1297 | now route through `tracing::warn!` with structured fields. |
| 1298 | - **`Ctrl+R` session-restore picker is workspace-scoped** (#1395, |
| 1299 | PR #1397 from **@linzhiqin2003**). `SessionPickerView::new` |
| 1300 | previously listed every saved session on disk sorted globally — |
| 1301 | so opening DeepSeek-TUI in Project B and pressing `Ctrl+R` could |
| 1302 | hand back Project A's last conversation. The picker now filters |
| 1303 | by current workspace, with a fallback hint when no in-workspace |
| 1304 | sessions exist. |
| 1305 | - **MCP discovery survives malformed items** (PR #1410 from |
| 1306 | **@Liu-Vince**). The `tools/list`, `resources/list`, |
| 1307 | `resources/templates/list`, and `prompts/list` walks previously |
| 1308 | did `serde_json::from_value::<Vec<…>>(…).unwrap_or_default()`, |
| 1309 | which silently discarded the entire page when any single entry |
| 1310 | was misshapen. Each list now iterates per-item, skipping |
| 1311 | malformed entries with a `tracing::debug!` instead of dropping |
| 1312 | the rest of the catalogue. Composes with the v0.8.x pagination |
| 1313 | loop landed for #1256. |
| 1314 | - **MCP SSE transport accepts CRLF-framed endpoint events** (#1309, |
| 1315 | PR #1358 from **@reidliu41**). FastMCP / uvicorn-style SSE |
| 1316 | streams using `\r\n\r\n` separators now discover the endpoint and |
| 1317 | send initialization requests instead of timing out while waiting |
| 1318 | for an LF-only event boundary. |
| 1319 | - **Composer ignores leaked SGR mouse-report bursts** (#1418, |
| 1320 | PR #1421 from **@reidliu41**). Some SSH / IDE terminal chains |
| 1321 | leak fragments like `[<35;44;18M` into stdin while mouse capture |
| 1322 | is enabled; the composer now filters those bursts at the insertion |
| 1323 | boundary without stripping ordinary coordinate-like typed text. |
| 1324 | - **Footer right-cluster chips can no longer crowd the left status |
| 1325 | line** (#1357, PR #1417 from **@Wenjunyun123**). The footer now |
| 1326 | reserves visible space for the left status before selecting cache / |
| 1327 | aux chips, dropping oversized right-side chips instead of pushing |
| 1328 | the row over the available terminal width. |
| 1329 | - **Web search drops spam-stuffed SERPs** (#964, PR #1396 from |
| 1330 | **@linzhiqin2003**). The Bing / DDG fallback paths now filter |
| 1331 | the SEO-farm domains that were poisoning quick lookups. |
| 1332 | - **Language directive: `reasoning_content` follows the user's |
| 1333 | message language** (#1118, PR #1398 from **@linzhiqin2003**) — |
| 1334 | previously the project context's inferred `lang` could override |
| 1335 | the latest user message, leading to English thinking for a |
| 1336 | Chinese turn. |
| 1337 | - **Deferred tools hydrate their schema before first execution** |
| 1338 | (#1419, PR #1429 from **@SamhandsomeLee**). When the model asks |
| 1339 | for a deferred tool such as `edit_file` before seeing its schema, |
| 1340 | the engine now loads the tool, returns a non-executed hydration |
| 1341 | result with the expected fields, and requires a retry instead of |
| 1342 | executing guessed argument names. Common `edit_file` aliases such |
| 1343 | as `old_string -> search` and `new_string -> replace` are called |
| 1344 | out in the retry hint. |
| 1345 | - **DeepSeek public aliases replay thinking-mode tool turns** |
| 1346 | (PR #1428 from **@Beltran12138**). `deepseek-chat` and |
| 1347 | `deepseek-reasoner` now classify as V4 reasoning models for |
| 1348 | `reasoning_content` replay, preventing second-turn HTTP 400s |
| 1349 | after tool calls when users keep the onboarding default model |
| 1350 | alias. |
| 1351 | - **`Ctrl+O` expands thinking blocks still in flight.** |
| 1352 | Two compounding bugs were making the "thinking collapsed; press |
| 1353 | Ctrl+O for full text" affordance a lie. (1) `open_thinking_pager` |
| 1354 | only searched `app.history`, but after `ThinkingComplete` the |
| 1355 | finalized thinking entry sits in `app.active_cell` with |
| 1356 | `streaming = false` until the active cell flushes at end-of-turn; |
| 1357 | during that window the handler surfaced "No thinking blocks to |
| 1358 | expand" while the affordance pointed at the live entry. Routed |
| 1359 | through the existing `cell_at_virtual_index` / `virtual_cell_count` |
| 1360 | resolver that `open_tool_details_pager` already uses, so |
| 1361 | selection-based and most-recent lookups both reach in-flight |
| 1362 | entries. (2) The keybinding guard required `key.modifiers == |
| 1363 | KeyModifiers::CONTROL` (exact match), so any extra modifier bit |
| 1364 | set by the terminal — Shift while a native-selection bypass was |
| 1365 | active, Caps Lock indicator on some keyboard layouts — silently |
| 1366 | fell through to the `$EDITOR` arm and did nothing visible on an |
| 1367 | empty composer. Relaxed to `contains(KeyModifiers::CONTROL)` to |
| 1368 | match the existing Ctrl+P / Ctrl+B pattern. Regression-guarded by |
| 1369 | `open_thinking_pager_finds_thinking_in_active_cell`. |
| 1370 | - **Skill completions no longer flood the top-level slash menu** |
| 1371 | (#1437, PR #1442 from **@reidliu41**). Installed skills now |
| 1372 | complete under `/skill <name>` while the root `/` menu stays |
| 1373 | focused on built-in commands. |
| 1374 | - **`edit_file` rejects no-op replacements** (PR #1460 from |
| 1375 | **@xiluoduyu**). Identical `search` / `replace` arguments now |
| 1376 | fail fast with a clear validation error instead of producing an |
| 1377 | empty diff that can trap the model in retry loops. |
| 1378 | - **Windows-terminal glyph widths are stable** (#1314, PR #1465 |
| 1379 | from **@CrepuscularIRIS**). SMP emoji in the header and file tree |
| 1380 | were replaced with BMP-width-safe symbols / text so cmd, |
| 1381 | PowerShell, WezTerm, and Alacritty do not mismeasure rows. |
| 1382 | - **Ghostty defaults to low-motion rendering** (#1445, PR #1468 |
| 1383 | from **@CrepuscularIRIS**). `TERM_PROGRAM=ghostty` now receives |
| 1384 | the same animation cap as VS Code terminals to avoid redraw |
| 1385 | flicker on affected setups. |
| 1386 | - **Docker buildx provenance permission failures get an actionable |
| 1387 | hint** (#1449, PR #1469 from **@CrepuscularIRIS**). macOS shell |
| 1388 | outputs matching the restricted provenance metadata failure now |
| 1389 | include guidance to disable provenance for that build. |
| 1390 | - **Windows CMD mouse-wheel fallback scrolls the transcript** |
| 1391 | (#1443, PR #1471 from **@CrepuscularIRIS**). When mouse capture is |
| 1392 | off, composer arrow-scroll defaults on so terminal wheel events |
| 1393 | mapped to Up / Down do not cycle composer history. |
| 1394 | |
| 1395 | ### Added |
| 1396 | |
| 1397 | - **MCP HTTP transport honors `HTTP(S)_PROXY` / `NO_PROXY`** (#1408 |
| 1398 | from **@hlx98007**). Reqwest 0.13 does not auto-detect proxy env |
| 1399 | vars by default, so MCP HTTP connections were bypassing the |
| 1400 | proxy that every other tool on the box (curl, npm, git, …) was |
| 1401 | using. Connections behind corporate egress proxies and |
| 1402 | China-mainland Clash / Shadowsocks tunnels now work transparently. |
| 1403 | Malformed `HTTPS_PROXY` values log a `tracing::warn!` and the |
| 1404 | connection proceeds without a proxy rather than failing the MCP |
| 1405 | attach. |
| 1406 | - **Note management slash commands** (PR #1407 from |
| 1407 | **@reidliu41**). `/note add`, `/note list`, and friends for |
| 1408 | persistent maintainer-style notes inside the TUI, backed by |
| 1409 | `~/.deepseek/notes/`. |
| 1410 | - **Header surfaces the runtime version chip.** A `v0.8.29` tag |
| 1411 | sits in the header's right cluster after the provider / effort / |
| 1412 | Live / context chips. Styled with `palette::TEXT_HINT` so it |
| 1413 | reads behind the streaming indicators. Drops first under tight |
| 1414 | terminal width. |
| 1415 | - **Global `~/.deepseek/AGENTS.md` now merges with project |
| 1416 | AGENTS.md** (#1157, PR #1399 from **@linzhiqin2003**) instead of |
| 1417 | being shadowed when a workspace ships its own. |
| 1418 | - **Auto-routing recognises CJK debug / search keywords** (PRs |
| 1419 | #1401 and #1402 from **@linzhiqin2003**) — `--model auto` and |
| 1420 | the reasoning-effort picker correctly route Chinese / Japanese |
| 1421 | technical queries that previously fell through to the generic |
| 1422 | baseline. |
| 1423 | |
| 1424 | ### Security |
| 1425 | |
| 1426 | - **`sync-cnb.yml` workflow hardened** (CodeQL finding from |
| 1427 | v0.8.28). Adds explicit `permissions: contents: read` |
| 1428 | (least-privilege), bumps `actions/checkout` v3 → v4, and |
| 1429 | narrows the trigger from `on: [push]` to `on: push.branches: |
| 1430 | [main]` + `tags: ['v*']`. Feature branches no longer mirror to |
| 1431 | CNB; only `main` and tagged releases do. |
| 1432 | - **Post-exit resume hint avoids session-id taint.** The TUI now |
| 1433 | checks whether a session exists separately from the constant |
| 1434 | resume-hint text it prints after leaving the alt-screen, resolving |
| 1435 | the `rust/cleartext-logging` CodeQL alert without reintroducing |
| 1436 | scroll-demon stdout writes. |
| 1437 | |
| 1438 | ### Internal |
| 1439 | |
| 1440 | - **+438 LOC of new test coverage** across four PRs from |
| 1441 | **@linzhiqin2003**: `error_taxonomy::classify_error_message` |
| 1442 | and Display impls (#1403), `parse_pages_arg` edge cases (#1404), |
| 1443 | `optional_search_max_results` precedence (#1405), and |
| 1444 | `sanitize_stream_chunk` control-byte filtering (#1406). |
| 1445 | - **`runtime_log` module** ships with a regression test pinning |
| 1446 | the `HOME` / `USERPROFILE` / `dirs::home_dir()` resolution |
| 1447 | order, holding the process-wide `test_support::lock_test_env()` |
| 1448 | lock for env-mutation safety. |
| 1449 | - **Header rendering** gains two regression tests |
| 1450 | (`header_renders_version_chip_when_width_allows` and |
| 1451 | `narrow_header_drops_version_chip_before_dropping_mode`) |
| 1452 | pinning the version chip's cascade priority. |
| 1453 | - **Workspace/session test isolation** tightened (PR #1431 from |
| 1454 | **@reidliu41**). Git-root detection ignores invalid parent `.git` |
| 1455 | markers, env-mutating tests share the crate-wide test lock, and |
| 1456 | the streamable HTTP MCP mock server stays alive for the full test. |
| 1457 | - **Config-mutating smoke tests now isolate `DEEPSEEK_CONFIG_PATH`.** |
| 1458 | The command registry and web-config commit tests no longer rewrite |
| 1459 | the developer's real `~/.deepseek/config.toml` while validating |
| 1460 | release candidates locally. |
| 1461 | |
| 1462 | ## [0.8.28] - 2026-05-10 |
| 1463 | |
| 1464 | A maintenance release bundling four streaming / approvals / cache |
| 1465 | bug-fix cherry-picks, six smaller community fixes, a Cmux |
| 1466 | notification probe, GPU-terminal flicker hardening via DEC 2026 |
| 1467 | synchronized output, VS Code low-motion auto-detection, a CNB |
| 1468 | mirror workflow, V4-steered tool descriptions, and test-suite |
| 1469 | stabilization for parallel-test environment races. |
| 1470 | |
| 1471 | ### Added |
| 1472 | |
| 1473 | - **CNB mirror workflow** (PR #1373 from **@Anyexyz**) — a |
| 1474 | GitHub Actions workflow (`sync-cnb.yml`) mirrors every push to |
| 1475 | the `cnb.cool/deepseek-tui.com/DeepSeek-TUI` repository, |
| 1476 | closing out the long-standing China-mirror request. Requires the |
| 1477 | `CNB_GIT_TOKEN` repo secret. |
| 1478 | - **Cmux desktop notification support via `LC_TERMINAL`** (#1281, |
| 1479 | PR #1340 from **@CrepuscularIRIS**) — Cmux sets |
| 1480 | `LC_TERMINAL=Cmux` rather than `TERM_PROGRAM`, so the previous |
| 1481 | notification probe fell back to `BEL` instead of using OSC 9. |
| 1482 | `resolve_method()` now checks `LC_TERMINAL` as a secondary probe |
| 1483 | and adds Cmux to the OSC 9 allowlist. Terminals that set |
| 1484 | neither env var can still force OSC 9 via |
| 1485 | `[notifications].method = "osc9"`. Two regression tests pin the |
| 1486 | Cmux and WezTerm `LC_TERMINAL` paths; the existing |
| 1487 | unknown-terminal-on-Unix test now clears `LC_TERMINAL` before |
| 1488 | asserting fallback so it doesn't flake on CI hosts that set it. |
| 1489 | - **DEC 2026 synchronized output around terminal repaints** (PR |
| 1490 | #1361 from **@xuezhaoyu**) — the viewport-reset path now wraps |
| 1491 | `terminal.clear()` in `\x1b[?2026h` / `\x1b[?2026l` so |
| 1492 | GPU-accelerated terminals (Ghostty, VSCode Terminal, Kitty, |
| 1493 | WezTerm) defer rendering until the whole frame is staged, |
| 1494 | eliminating mid-frame flicker on resize / focus / TurnComplete. |
| 1495 | The earlier "drop destructive 2J/3J" fix from v0.8.27 stays; |
| 1496 | this PR is complementary, batching the same lighter reset |
| 1497 | sequence into a single synchronized frame. Terminals without |
| 1498 | DEC 2026 support silently ignore the sequence. |
| 1499 | - **`low_motion` auto-enables under VS Code integrated terminal** |
| 1500 | (PR #1365 from **@CrepuscularIRIS**) — `apply_env_overrides()` |
| 1501 | now treats `TERM_PROGRAM=vscode` the same way it treats |
| 1502 | `NO_ANIMATIONS=1`: force `low_motion = true` and |
| 1503 | `fancy_animations = false`. The VS Code terminal compositor |
| 1504 | cannot keep up with 120 fps redraws and produces rapid flicker |
| 1505 | (#1356); the 30 fps low-motion cap is the right default there. |
| 1506 | Env overlays always win over the disk-loaded value, matching |
| 1507 | the existing precedence for `NO_ANIMATIONS`. |
| 1508 | |
| 1509 | ### Fixed |
| 1510 | |
| 1511 | - **Cache usage shows 0 when API omits cache data** (#1391, PR #1392 |
| 1512 | from **@Oliver-ZPLiu**) — `SessionUsage.cache_creation_input_tokens` / |
| 1513 | `cache_read_input_tokens` are now `Option<u64>` instead of `u64` |
| 1514 | defaulting to 0. When the upstream API doesn't report cache |
| 1515 | hit/miss, the model sees `null` instead of misleading zeros, and |
| 1516 | reasoning about cache utilization is accurate. |
| 1517 | - **Deny of one tool call no longer blocks all future calls of the |
| 1518 | same tool** (#1377, PR #1388 from **@Oliver-ZPLiu**) — denying a |
| 1519 | tool call now only caches the per-call `approval_key`, not the |
| 1520 | tool type. Subsequent invocations of the same tool prompt for |
| 1521 | approval again instead of being silently auto-denied. |
| 1522 | - **Streaming thinking blocks no longer drop their tail on |
| 1523 | MessageComplete** (#861 RC3, PR #1389 from **@linzhiqin2003**) — |
| 1524 | the active streaming entry is now drained into the finalized |
| 1525 | cell on `MessageComplete`, eliminating a data-loss path where |
| 1526 | the last chunk(s) of a streaming "thinking" reply could be |
| 1527 | discarded when `MessageComplete` arrived ahead of |
| 1528 | `ThinkingComplete` in a bursty event stream. Also closes a |
| 1529 | related HTTP 400 on the next turn (DeepSeek V4 requires |
| 1530 | `reasoning_content` replay for assistant messages that carry |
| 1531 | tool calls). |
| 1532 | - **Streaming thinking renders live in collapsed view** (#861 RC4, |
| 1533 | #1324, PR #1390 from **@linzhiqin2003**) — collapsed thinking |
| 1534 | cells now stream their content as it arrives instead of staying |
| 1535 | at a static "thinking..." placeholder until streaming ends. When |
| 1536 | the live body exceeds the collapsed budget, the truncation |
| 1537 | affordance ("thinking continues; press Ctrl+O for full text") |
| 1538 | now fires during streaming with head lines dropped so the |
| 1539 | visible window tracks the live cursor at the bottom. |
| 1540 | - **First-turn latency bounded on large workspaces** (#697, PR #1386 |
| 1541 | from **@linzhiqin2003**) — the working-set file walker now caps |
| 1542 | the number of entries it visits during initial indexing, so |
| 1543 | starting a session in a workspace with a deep `node_modules`, |
| 1544 | `target`, or `.venv` no longer stalls the first response on |
| 1545 | filesystem traversal. |
| 1546 | - **Duplicate error toast on transcript-rendered turn errors** (PR |
| 1547 | #1368 from **@douglarek**) — when a turn error is already in the |
| 1548 | transcript as a system/error cell, the status-line toast is |
| 1549 | suppressed so the user doesn't see the same failure twice. |
| 1550 | - **Clearer continue tip on idle prompts** (PR #1370 from |
| 1551 | **@nightfallsad**) — the "press Tab to continue" affordance now |
| 1552 | uses concrete language instead of a vague hint. |
| 1553 | - **Ctrl+Enter content lost when engine is idle** (#1331, PR #1347 |
| 1554 | from **@Oliver-ZPLiu**) — when no turn was active, `Ctrl+Enter` |
| 1555 | routed the message to `rx_steer` (only monitored inside |
| 1556 | `handle_deepseek_turn`), so the user saw their message in the |
| 1557 | transcript via the local mirror but the LLM never received it — |
| 1558 | the next regular Enter would drain it as a "stale steer". The |
| 1559 | idle path now sends through the standard `handle_send_message` |
| 1560 | flow so the submission reaches the engine. |
| 1561 | - **Explicit hidden / ignored `@`-mention completions work** |
| 1562 | (#1270 follow-up) — PR #1270 from **@SamhandsomeLee** landed |
| 1563 | the `add_local_reference_completions` helper and tests in |
| 1564 | v0.8.27 but never wired it into `Workspace::completions()` or |
| 1565 | `build_file_index`. The two regression tests were ignored with |
| 1566 | a "v0.8.28 follow-up" marker. This release wires the helper |
| 1567 | into both entry points so `@.deepseek/commands/start-task.md` |
| 1568 | and `@.generated/specs/device-layout.md` (and the basename |
| 1569 | fuzzy-resolve equivalent) now surface from gitignored |
| 1570 | user-folders while `.deepseekignore` entries stay blocked. |
| 1571 | Both tests un-ignored. |
| 1572 | |
| 1573 | ### Changed |
| 1574 | |
| 1575 | - **Prompt-side reliability guidance** (PR #1393 from |
| 1576 | **@Oliver-ZPLiu**) — `prompts/base.md` gains three Verification |
| 1577 | Principle bullets steering the model to verify before reporting |
| 1578 | complete, preserve only key facts from tool results, and |
| 1579 | inspect errors before retrying. Combined with the truthful- |
| 1580 | reporting addition from #1392, the model is less likely to claim |
| 1581 | unverified successes or repeat the identical failing tool call. |
| 1582 | - **V4-steered tool descriptions** (#711, PR #1379 from |
| 1583 | **@linzhiqin2003**) — every model-visible tool description |
| 1584 | (`read_file`, `write_file`, `edit_file`, `list_dir`, |
| 1585 | `grep_files`, `file_search`, `web_search`, `apply_patch`, |
| 1586 | `fetch_url`) now opens with a short *"use this instead of X |
| 1587 | in exec_shell"* steering line, the return shape, and the |
| 1588 | limits. Routes V4 toward our typed tools and away from |
| 1589 | shell footguns. All description strings stay under 1024 |
| 1590 | chars (max: 350) with no embedded newlines so the cached |
| 1591 | tool catalogue stays prefix-stable for V4's KV cache. |
| 1592 | Removes the unused legacy `normal.txt` / `plan.txt` / |
| 1593 | `yolo.txt` prompt templates (referenced only by their own |
| 1594 | self-tests). |
| 1595 | |
| 1596 | ### Internal |
| 1597 | |
| 1598 | - Test-suite parallelism stabilization (commit |
| 1599 | `test: stabilize parallel test execution`). Folds three local |
| 1600 | test-mutex implementations into the process-wide |
| 1601 | `test_support::lock_test_env`, eliminating a class of |
| 1602 | intermittent failures (`refresh_system_prompt_is_noop_when_unchanged`, |
| 1603 | `save_api_key_for_openrouter_writes_provider_table`, |
| 1604 | `list_archives_sorts_by_cycle_number`) observed during the |
| 1605 | v0.8.27 release cycle. |
| 1606 | - Windows `task_manager` timeout bumped 3s → 10s on four tests |
| 1607 | exercising durable task recovery, addressing an intermittent |
| 1608 | CI timeout on Windows under file-I/O load. |
| 1609 | - `provider_switch_clears_turn_cache_history` now isolates |
| 1610 | `HOME` / `USERPROFILE` to a tempdir for its lifetime. The test |
| 1611 | was silently writing `default_provider = "ollama"` to the |
| 1612 | developer's real `~/Library/Application Support/deepseek/settings.toml` |
| 1613 | on every run, which then contaminated parallel-running picker |
| 1614 | tests because Ollama is a pass-through provider that hides the |
| 1615 | DeepSeek model rows. |
| 1616 | - `settings::tests::no_animations_test_guard` and |
| 1617 | `term_program_test_guard` both now return |
| 1618 | `crate::test_support::lock_test_env()` instead of their own |
| 1619 | module-local mutexes — folding them into the same |
| 1620 | process-wide test env lock the v0.8.27 EnvGuard family was |
| 1621 | migrated to. Without this, a `NO_ANIMATIONS=1` write from one |
| 1622 | test family could race a `TERM_PROGRAM=iTerm.app` write from |
| 1623 | the other through the shared `apply_env_overrides` path and |
| 1624 | flip `low_motion` to `true` on the assertion side. |
| 1625 | |
| 1626 | ## [0.8.27] - 2026-05-10 |
| 1627 | |
| 1628 | A polish release bundling 17 community PRs plus a focused user-issue |
| 1629 | sweep over the 24–48 hours after v0.8.26 shipped. Headline fixes: |
| 1630 | cross-terminal flicker on Ghostty / VSCode / Win10 conhost (most- |
| 1631 | reported v0.8.26 regression), long-text right-edge overflow, an |
| 1632 | in-app pager copy-out, context-sensitive Ctrl+C, an MCP pool that |
| 1633 | auto-reloads on config changes, and a model-callable `notify` tool. |
| 1634 | Big thanks to every contributor below. |
| 1635 | |
| 1636 | ### Added |
| 1637 | |
| 1638 | - **Unified `/mode` command** (#1247) — `/mode [agent|plan|yolo|1|2|3]` |
| 1639 | replaces the separate `/agent`, `/plan`, and `/yolo` commands. Running |
| 1640 | `/mode` without arguments opens a picker modal. The legacy aliases |
| 1641 | (`/yolo`, `/agent`, `/plan`) are kept as compatibility shorthands. |
| 1642 | Thanks **@reidliu41**. |
| 1643 | - **`/status` runtime diagnostics** (#1223) — shows version, provider, |
| 1644 | model, workspace, mode, permissions, context-window usage, cache |
| 1645 | hit/miss, and session cost. Previously `/status` was an alias for |
| 1646 | `/statusline` (footer config); that alias is now `/statusline` only. |
| 1647 | Thanks **@reidliu41**. |
| 1648 | - **`/feedback` command** (#1185) — opens the matching GitHub issue |
| 1649 | template (bug report, feature request) in the browser. Security |
| 1650 | vulnerability reports route through the project's security policy |
| 1651 | page first. Thanks **@reidliu41**. |
| 1652 | - **Session artifact metadata** (#1220) — large tool outputs spilled to |
| 1653 | the session artifacts directory are now tracked in a durable metadata |
| 1654 | index, so saved sessions retain references across save/restore cycles. |
| 1655 | Thanks **@THINKER-ONLY**. |
| 1656 | - **Subagent results are self-reports** (#1140) — the compacted result |
| 1657 | summary now notes that child-agent outputs are unverified self-reports. |
| 1658 | The parent model should verify side effects with tools like `read_file` |
| 1659 | or `list_dir` before claiming success. Thanks **@THINKER-ONLY**. |
| 1660 | - **Global AGENTS.md fallback** (#1197) — when the workspace and its |
| 1661 | parents don't provide project instructions, the TUI now loads |
| 1662 | `~/.deepseek/AGENTS.md` before falling back to auto-generated |
| 1663 | instructions. Repo-local context still takes priority. |
| 1664 | Thanks **@manaskarra**. |
| 1665 | - **`--yolo` forwarded from CLI to TUI** (#1233) — the `deepseek --yolo` |
| 1666 | flag now propagates through the dispatcher to the TUI binary via |
| 1667 | `DEEPSEEK_YOLO=true`. Previously the flag set `yolo` in the CLI |
| 1668 | process but the TUI session started in its default mode. |
| 1669 | Thanks **@fuleinist**. |
| 1670 | - **`composer_arrows_scroll` config** (#1211) — a new |
| 1671 | `tui.composer_arrows_scroll` option (default `false`) makes plain |
| 1672 | Up/Down arrow keys scroll the transcript when the composer is empty, |
| 1673 | instead of navigating input history. Helpful for terminals that map |
| 1674 | trackpad gestures to arrow keys. Thanks **@lbcheng888**. |
| 1675 | - **Session cost persistence** (#1192) — accumulated costs (session + |
| 1676 | sub-agents, both USD and CNY) and the displayed-cost high-water mark |
| 1677 | now survive session save/restore, so the monotonic cost guarantee |
| 1678 | (#244) holds across restarts. Thanks **@lbcheng888**. |
| 1679 | - **Provider-aware model picker and provider persistence** (#1320) — |
| 1680 | switching providers now persists the choice to |
| 1681 | `~/.deepseek/settings.toml` so it survives restarts. The model |
| 1682 | picker hides DeepSeek-specific models when a non-DeepSeek provider |
| 1683 | is active. `OPENAI_MODEL` env var now overrides the per-provider |
| 1684 | model rather than the global `default_text_model`. Bailian / ZhiPu |
| 1685 | Coding Plan endpoints are now supported. |
| 1686 | Thanks **@imkingjh999**. |
| 1687 | - **HTTP User-Agent header** (#1320) — all outbound API requests now |
| 1688 | carry `deepseek-tui/{version}` in the User-Agent, matching the format |
| 1689 | `fetch_url` already uses. Thanks **@imkingjh999**. |
| 1690 | |
| 1691 | ### Fixed |
| 1692 | |
| 1693 | - **Cross-terminal flicker on TurnComplete / focus / resize** (#1119, |
| 1694 | #1260, #1295, #1352, #1356, #1363, #1366) — the viewport-reset |
| 1695 | sequence emitted before each forced repaint no longer includes |
| 1696 | `\x1b[2J\x1b[3J`. Combined with the immediately-following ratatui |
| 1697 | `terminal.clear()`, the destructive pair produced a double-clear that |
| 1698 | Ghostty, the VSCode integrated terminal, and Win10 conhost rendered |
| 1699 | as a visible blank-then-repaint flicker. The lighter sequence |
| 1700 | (`\x1b[r\x1b[?6l\x1b[H`) plus the alt-screen buffer's double-buffering |
| 1701 | handles viewport correctness without flicker. macOS Terminal.app / |
| 1702 | iTerm2 / alacritty users were already unaffected and remain so. |
| 1703 | - **`/skills --remote` and `/skills sync` diagnostics** (#1329) — the |
| 1704 | underlying anyhow chain has always been formatted with `{err:#}`, but |
| 1705 | the chain alone is often opaque (e.g. "error sending request"). The |
| 1706 | error message now appends a one-line hint when the chain matches a |
| 1707 | common failure pattern: DNS / connection refused / TLS / 4xx / 429 / |
| 1708 | timeout. Each hint points at the most likely cause and a concrete |
| 1709 | next step. |
| 1710 | |
| 1711 | ### Added |
| 1712 | |
| 1713 | - **Pager copy-out** (#1354) — full-screen pagers (`Alt+V` tool details, |
| 1714 | `Ctrl+O` thinking content, shell-job / task / MCP-manager pagers, and |
| 1715 | the selection pager) now accept `c` or `y` to copy the entire body to |
| 1716 | the system clipboard. The pager intercepts mouse capture so terminal- |
| 1717 | native selection isn't available inside it; this restores the |
| 1718 | copy-out path that users on macOS / Windows / WSL expect. The footer |
| 1719 | hint now reads `… / search c copy q/Esc close`. A status toast |
| 1720 | confirms success ("Pager content copied"), empty-body, or failure. |
| 1721 | - **`notify` tool** (#1322) — model-callable desktop notification. |
| 1722 | Always-loaded (no ToolSearch round-trip). Routes through the existing |
| 1723 | `tui::notifications` infrastructure: OSC 9 on iTerm2 / Ghostty / |
| 1724 | WezTerm, BEL fallback on macOS / Linux, `MessageBeep` on Windows when |
| 1725 | explicitly opted in. Honours the user's `[notifications].method` |
| 1726 | config — when set to `off`, the tool is a silent no-op. Title and |
| 1727 | body are length-capped (80 / 200 chars) on character (not byte) |
| 1728 | boundaries to keep the OSC 9 escape clean and avoid mid-grapheme |
| 1729 | truncation. The tool description steers the model away from chatter: |
| 1730 | use only when a long-running task completes or genuinely needs the |
| 1731 | user's attention. |
| 1732 | |
| 1733 | ### Fixed (cont.) |
| 1734 | |
| 1735 | - **Long output text overflowed the right edge** (#1344, #1351) — |
| 1736 | paragraph rendering (`render_line_with_links`) and code-block |
| 1737 | wrapping (`wrap_text` for `Block::Code`) were word-based: a single |
| 1738 | word wider than the available column was placed alone on a line and |
| 1739 | silently overflowed. Long URLs, paths, hashes, and no-whitespace CJK |
| 1740 | runs all hit this. Both paths now hard-break overlong words at the |
| 1741 | character level, matching the v0.8.25 fix for table cells. The |
| 1742 | rendered width is capped at the budget for every line; full content |
| 1743 | is preserved across wrapped segments. Snapshot-style tests pin the |
| 1744 | invariant at widths 40, 60, 80, and 120. |
| 1745 | |
| 1746 | ### Changed |
| 1747 | |
| 1748 | - **`Ctrl+C` now copies an active transcript selection** (#1337) — on |
| 1749 | Windows, plain `Ctrl+C` is the OS-wide copy chord, and treating it |
| 1750 | as "exit" stole work whenever a user copy-pasted from the |
| 1751 | transcript. `Ctrl+C` is now a four-stage decision: 1) selection |
| 1752 | active → copy + clear (matches the OS convention); 2) turn in |
| 1753 | flight → cancel (unchanged); 3) quit-armed within 2s → exit cleanly |
| 1754 | (unchanged); 4) idle, no selection → arm the 2-second |
| 1755 | "press Ctrl+C again to quit" prompt (unchanged). The decision is |
| 1756 | factored into a `CtrlCDisposition` helper with a unit-tested |
| 1757 | priority table. `Cmd+C` (macOS) and `Ctrl+Shift+C` continue to copy |
| 1758 | unchanged. |
| 1759 | - **Cancel-key discoverability hint on turn start** (#1367) — when a |
| 1760 | turn begins, the status-message slot now surfaces "Press Esc or |
| 1761 | Ctrl+C to cancel" if the slot is otherwise empty. Real transient |
| 1762 | status messages still take precedence; the hint clears as soon as |
| 1763 | any other update fires. Closes the loop on users who didn't know |
| 1764 | how to interrupt a long-running turn. |
| 1765 | - **Lazy auto-reload of MCP pool on config-file change** (#1267 part 2) — |
| 1766 | v0.8.26 surfaced the underlying spawn errors; v0.8.27 closes the |
| 1767 | loop on the second half of the report (manual `/mcp reload` after |
| 1768 | `~/.deepseek/mcp.json` edits). `McpPool::get_or_connect` now does a |
| 1769 | cheap `stat` + content-hash check before each connection lookup. If |
| 1770 | the on-disk file's mtime moved AND its content hash changed since |
| 1771 | the pool was loaded, all live connections are dropped so the next |
| 1772 | `get_or_connect` reattaches under the new config. Pool-construction |
| 1773 | via `McpPool::new` (tests, ad-hoc snapshots) is unaffected — only |
| 1774 | pools built with `from_config_path` watch the source file. No file |
| 1775 | watcher; no long-lived task. mtime-only churn (touched but |
| 1776 | byte-unchanged content) does not trigger a reload, so networked |
| 1777 | filesystems with coarse mtime granularity won't churn the pool. |
| 1778 | - **Paste consolidation now happens at paste time, not submit time** — |
| 1779 | large bracketed pastes that exceed the 16 000-char safety cap are |
| 1780 | now folded into a workspace `.deepseek/pastes/paste-…md` file and |
| 1781 | swapped for an `@`-mention immediately on paste, instead of waiting |
| 1782 | until the user presses Enter. The user sees the `@`-mention in the |
| 1783 | composer (and the "consolidated → @mention" toast) before deciding |
| 1784 | whether to send, eliminating the "I pressed Enter and an `@`-mention |
| 1785 | appeared in the chat I didn't authorise" surprise. The submit-time |
| 1786 | consolidation remains as a safety net for any other code path that |
| 1787 | fills the buffer above the cap, so the cap is still enforced exactly |
| 1788 | once. |
| 1789 | - **Auto-disable paste-burst once bracketed paste verified** — the |
| 1790 | rapid-keystroke paste-burst heuristic (default-on for terminals |
| 1791 | without bracketed paste) used to keep running on every session. |
| 1792 | Once a real `Event::Paste` arrives in a session, paste-burst now |
| 1793 | short-circuits — bracketed paste is verified working, and running |
| 1794 | the heuristic alongside it just creates false positives on fast |
| 1795 | typing / IME commits / autocomplete bursts. Terminals that never |
| 1796 | deliver bracketed paste (the original target audience) are |
| 1797 | unaffected; the heuristic still fires there. |
| 1798 | - **Short CJK multi-line paste no longer auto-submits first line** |
| 1799 | (#1302) — pasting `请联网搜索：\nSTM32 …` (short non-ASCII first line |
| 1800 | followed by a newline) used to fail the paste-burst detection |
| 1801 | heuristic because the first line had no whitespace and was under |
| 1802 | the 16-char threshold; the trailing pasted newline then fell |
| 1803 | through as a real Enter and submitted the first line on its own. |
| 1804 | The heuristic now treats any non-ASCII run as paste-like, so the |
| 1805 | Enter is absorbed into the burst buffer. Thanks **@reidliu41** |
| 1806 | (PR #1342). |
| 1807 | - **Onboarding screens render in the selected language** — when a |
| 1808 | user picked 简体中文 / 日本語 / Português (Brasil) at the language |
| 1809 | step, every subsequent screen (API key entry, workspace trust |
| 1810 | prompt, final tips) used to remain in English. The |
| 1811 | `set_locale_from_onboarding` path now drives the title, body |
| 1812 | copy, hints, and footer of each onboarding screen through the |
| 1813 | localization table, so once you pick your language the rest of |
| 1814 | the flow is in that language. Particularly nice for users on |
| 1815 | CJK input methods who want to avoid IME juggling during setup. |
| 1816 | - **`/skills <prefix>` filters the local skills list** (#1318) — on |
| 1817 | top of the v0.8.26 inter-row spacing (#1328 from @reidliu41), the |
| 1818 | list now narrows to skills whose names start with the typed |
| 1819 | prefix. Case-insensitive. The header reflects matched count vs |
| 1820 | registry total; an empty match set says so explicitly and points |
| 1821 | back at unfiltered `/skills`. `--remote` and `sync` stay |
| 1822 | reserved as subcommands; any `--`-prefixed argument is rejected |
| 1823 | rather than being silently treated as a no-match prefix. |
| 1824 | - **HTTP 400 quota errors retried** (#1203) — some OpenAI-compatible |
| 1825 | gateways return quota/rate-limit errors as HTTP 400 instead of 429. |
| 1826 | These are now classified as retryable `RateLimited` errors. |
| 1827 | Thanks **@dst1213**. |
| 1828 | - **Explicit hidden/ignored file completions** (#1270) — when the user |
| 1829 | types an explicit path starting with `.` (e.g., `.deepseek/commands/`), |
| 1830 | the file-completion system now surfaces hidden and gitignored entries |
| 1831 | while still respecting `.deepseekignore`. Thanks **@SamhandsomeLee**. |
| 1832 | |
| 1833 | ### Changed |
| 1834 | |
| 1835 | - **Windows mouse capture docs** (#1181) — the `--mouse-capture` help |
| 1836 | text and the configuration docs now mention scrollbar dragging and |
| 1837 | note that raw terminal selection on Windows may cross the sidebar. |
| 1838 | Thanks **@Oliver-ZPLiu**. |
| 1839 | - **README zh-CN sync** (#1235) — the Chinese README's quickstart section |
| 1840 | now shows `deepseek run pr <N>` instead of the outdated |
| 1841 | `deepseek pr <N>`. Thanks **@whtis**. |
| 1842 | - **Tool output render perf** (#1098) — tool output summaries and the |
| 1843 | "is this a diff?" check are now pre-computed once at cell creation |
| 1844 | instead of re-parsed every frame. Tool output cells also got a visual |
| 1845 | card-rail (`╭ │ ╰`) for clearer grouping. Thanks **@lbcheng888**. |
| 1846 | |
| 1847 | ### Internal |
| 1848 | |
| 1849 | - Test coverage for approval decision branches (@tuohai666, #1316) |
| 1850 | - Test coverage for hook event dispatch paths (@tuohai666, #1317) |
| 1851 | |
| 1852 | ## [0.8.26] - 2026-05-09 |
| 1853 | |
| 1854 | A security + polish release. Two responsibly-disclosed issues were |
| 1855 | patched, plus a small batch of internal release-pipeline fixes. Big |
| 1856 | thanks to **@JafarAkhondali** and **@47Cid** for the disclosures. |
| 1857 | |
| 1858 | ### Security |
| 1859 | |
| 1860 | - Hardened the `fetch_url` tool's network-target validation |
| 1861 | (GHSA-88gh-2526-gfrr). Thanks to **@JafarAkhondali**. |
| 1862 | - Tightened the default privileges of sub-agents created through |
| 1863 | `task_create` (GHSA-72w5-pf8h-xfp4). Thanks to **@47Cid**. |
| 1864 | |
| 1865 | Both items will have full advisory text once the GHSA entries are |
| 1866 | published. |
| 1867 | |
| 1868 | ### Fixed |
| 1869 | |
| 1870 | - **Hint when root `base_url` is set with a non-DeepSeek provider |
| 1871 | (#1308)** — config load now logs a warning telling the user to |
| 1872 | move the URL under the matching `[providers.<name>]` table or use |
| 1873 | the `*_BASE_URL` env var. Closes the silent-ignore footgun for |
| 1874 | Ollama / vLLM / OpenAI-compatible setups. |
| 1875 | - **Insecure base-URL error message is more discoverable (#1303)** — |
| 1876 | the rejection now spells out which env var to set (with underscores |
| 1877 | visible), notes that loopback hosts are auto-allowed, and shows a |
| 1878 | one-line `DEEPSEEK_ALLOW_INSECURE_HTTP=1 deepseek` example. |
| 1879 | - **Workspace skills survive prompt truncation** — when the skill |
| 1880 | catalog needs trimming to fit the prompt budget, workspace-local |
| 1881 | skills now keep precedence over global ones rather than being |
| 1882 | truncated indiscriminately. Thanks **@hhhaiai**. |
| 1883 | - **`/skills` listing has visual spacing** between entries so long |
| 1884 | skill descriptions don't run together. Thanks **@reidliu41**. |
| 1885 | - **Provider base-URL overrides reach the active provider** — the |
| 1886 | per-provider `*_BASE_URL` env vars (e.g. `OPENAI_BASE_URL`, |
| 1887 | `OPENROUTER_BASE_URL`) now propagate into the active provider's |
| 1888 | config entry consistently. Closes a gap where the override was |
| 1889 | parsed but never applied. Thanks **@reidliu41**. |
| 1890 | - **WSL2 turn-start timeout** — `TurnStarted` is now emitted before the |
| 1891 | snapshot step so a slow snapshot on WSL2's `/mnt/*` volumes doesn't |
| 1892 | push past the runtime watchdog and surface a spurious "engine may |
| 1893 | have stopped" error. Thanks **@michaeltse321**. |
| 1894 | - **`/init` auto-adds `.deepseek/` to `.gitignore` (#1326)** when the |
| 1895 | workspace is a git repo, so workspace-local snapshots, instructions, |
| 1896 | and pastes don't get accidentally committed. Idempotent on repeated |
| 1897 | runs. Thanks **@Giggitycountless**. |
| 1898 | - **MCP tool ordering is deterministic** — discovered tools and the |
| 1899 | resulting API tool block are now sorted by name so the prompt |
| 1900 | prefix the model sees is stable across runs, regardless of |
| 1901 | server-side pagination order. Improves prompt-cache hit rates with |
| 1902 | multi-server MCP setups. Thanks **@hxy91819**. |
| 1903 | - **Error cells render as plain text** so env-var names (`API_KEY_FOO`) |
| 1904 | in error messages keep their underscores instead of being parsed as |
| 1905 | markdown emphasis. Thanks **@douglarek**. |
| 1906 | - **`/clear` resets the Todos sidebar (#1258)** — previously `/clear` |
| 1907 | only reset the Plan panel; the Todos checklist persisted across |
| 1908 | clears. Thanks **@Giggitycountless**. |
| 1909 | - **Drag-select past the viewport edge auto-scrolls (#1163, #1255, |
| 1910 | #1292, #1298)** — when the mouse drag reaches the top or bottom of |
| 1911 | the transcript area the viewport now scrolls to follow the |
| 1912 | selection, the way text editors do. **Copy strips every visual-only |
| 1913 | decoration glyph** — tool-card rails (`╭│╰`), transcript rails |
| 1914 | (`▏`), reasoning rails (`╎`), tool-status symbols (`·•◦`), and |
| 1915 | tool-family glyphs no longer leak into clipboard output. Thanks |
| 1916 | **@Oliver-ZPLiu**. |
| 1917 | - MCP stdio servers no longer discard stderr. The spawn site now pipes |
| 1918 | stderr through a bounded ring buffer; when a server crashes |
| 1919 | mid-session, the transport-closed error includes the captured stderr |
| 1920 | tail instead of disappearing into `Stdio::null`. Useful for debugging |
| 1921 | Node/Python MCP servers that fail well after `initialize`. |
| 1922 | - Mouse capture now defaults on inside Windows Terminal (#1169, #1298, |
| 1923 | #1331). When `WT_SESSION` is set, in-app text selection is enabled |
| 1924 | by default and the wheel scrolls the transcript again (rather than |
| 1925 | the terminal interpreting wheel events as input-history keys). |
| 1926 | Legacy conhost stays opt-in via `--mouse-capture` or `[tui] |
| 1927 | mouse_capture = true` to preserve the protections from #878 / #898. |
| 1928 | Selection now clamps to the transcript region instead of the |
| 1929 | terminal painting native selection across the sidebar. |
| 1930 | - The build script now invalidates its cache on `.git/HEAD` changes, so |
| 1931 | the embedded short-SHA in `deepseek --version` stays current after |
| 1932 | commits and branch switches without needing `cargo clean`. Both |
| 1933 | regular checkouts and `git worktree` layouts are handled. |
| 1934 | - The release-time `changelog_entry_exists_for_current_package_version` |
| 1935 | gate walks up from the crate manifest to find `CHANGELOG.md` instead |
| 1936 | of assuming a fixed `../../CHANGELOG.md` layout. The workspace path |
| 1937 | still resolves; running the suite from a packaged crate skips the |
| 1938 | gate quietly instead of panicking. |
| 1939 | |
| 1940 | ## [0.8.25] - 2026-05-09 |
| 1941 | |
| 1942 | A stabilization + drift-fixes release. Headline work hardens the |
| 1943 | self-update path (no more `curl` shellout, real SHA-256 verification), |
| 1944 | fixes long-cell truncation in markdown tables, centralizes the MCP |
| 1945 | JSON-RPC framing, and unifies terminal-mode recovery on focus events. |
| 1946 | Big thanks to **Reid Liu (@reidliu41)** (Streamable HTTP MCP transport, |
| 1947 | `/config` column alignment), **Duducoco (@Duducoco)** (cache-stable |
| 1948 | `reasoning_content` replay), **jinpengxuan (@jinpengxuan)** (provider |
| 1949 | credentials during onboarding), **heloanc (@heloanc)** (Home/End cursor |
| 1950 | keys), **Wenjunyun123 (@Wenjunyun123)** (docs anchor scroll), and |
| 1951 | **Liu-Vince (@Liu-Vince)** (zh-Hans approval-dialog wording) for the |
| 1952 | contributions below. |
| 1953 | |
| 1954 | ### Added |
| 1955 | |
| 1956 | - **Streamable HTTP MCP endpoints with SSE fallback (#1300)** — adds |
| 1957 | the third MCP transport alongside stdio and SSE. The new transport |
| 1958 | posts JSON-RPC over plain HTTP with optional SSE upgrade for servers |
| 1959 | that prefer streaming responses. Thanks **Reid Liu (@reidliu41)**. |
| 1960 | - **`recall_archive` exposed in the parent agent registry** — the |
| 1961 | read-only BM25 archive search tool was previously only available to |
| 1962 | sub-agents; it is now callable from Plan, Agent, and YOLO parent |
| 1963 | registries. Plan mode's read-only contract is preserved (the existing |
| 1964 | registry test was updated to assert membership while still rejecting |
| 1965 | write/exec tools). |
| 1966 | |
| 1967 | ### Changed |
| 1968 | |
| 1969 | - **Markdown tables wrap long cells instead of truncating (#1163-adjacent)** |
| 1970 | — long cell content is word-wrapped within the column instead of |
| 1971 | collapsing to `…`. Column separators are preserved on every wrapped |
| 1972 | line so the table grid stays readable. |
| 1973 | - **MCP JSON-RPC framing centralized** — request/response correlation, |
| 1974 | timeout handling, and message framing now live above the byte-level |
| 1975 | transports. Stdio, SSE, and the new Streamable HTTP transport share a |
| 1976 | single protocol layer instead of each maintaining its own copy of the |
| 1977 | framing code. |
| 1978 | - **Self-update is curl-free and verifies SHA-256** — `deepseek update` |
| 1979 | no longer shells out to system `curl` (and no longer needs the |
| 1980 | Schannel `--ssl-no-revoke` Windows hack from v0.8.23). Downloads now |
| 1981 | use `reqwest::blocking` with rustls, and the aggregated |
| 1982 | `deepseek-artifacts-sha256.txt` manifest is parsed and checked |
| 1983 | against each downloaded asset before it is installed. Verification |
| 1984 | status is surfaced in the update output. |
| 1985 | - **Terminal-mode recovery unified in `recover_terminal_modes()`** — |
| 1986 | startup, `FocusGained`, and `resume_terminal` all route through one |
| 1987 | idempotent helper that re-establishes keyboard enhancement flags, |
| 1988 | mouse capture, bracketed paste, and focus events. Adding a new mode |
| 1989 | flag now only has to happen in one place. |
| 1990 | |
| 1991 | ### Fixed |
| 1992 | |
| 1993 | - **`reasoning_content` replay stable for prompt cache (#1297)** — |
| 1994 | reasoning text replayed from saved sessions now hashes consistently |
| 1995 | across turns so the cache-aware prompt builder's static-prefix |
| 1996 | stability isn't broken by replays. Thanks **Duducoco (@Duducoco)**. |
| 1997 | - **Active provider credentials respected during onboarding (#1265)** |
| 1998 | — the onboarding flow now reads credentials from the active provider |
| 1999 | instead of falling back to the default DeepSeek path when another |
| 2000 | provider is selected. Thanks **jinpengxuan (@jinpengxuan)**. |
| 2001 | - **Home/End keys move the input cursor (#1246)** — Home and End now |
| 2002 | jump the composer cursor to line start/end instead of being |
| 2003 | swallowed. Thanks **heloanc (@heloanc)**. |
| 2004 | - **Docs anchor scroll-margin overrideable (#1282)** — the |
| 2005 | scroll-margin offset on docs anchors is now overrideable so embedded |
| 2006 | contexts can adjust it without forking the stylesheet. Thanks |
| 2007 | **Wenjunyun123 (@Wenjunyun123)**. |
| 2008 | - **`/config` view columns aligned (#1290)** — the `/config` table now |
| 2009 | sizes the key column from the actual data instead of a fixed width, |
| 2010 | so long keys no longer overflow into the value column. Thanks |
| 2011 | **Reid Liu (@reidliu41)**. |
| 2012 | - **zh-Hans approval dialog wording (#1274)** — uses 终止 (terminate) |
| 2013 | instead of 中止 (abort) in the Chinese approval dialog, matching the |
| 2014 | English semantics. Thanks **Liu-Vince (@Liu-Vince)**. |
| 2015 | |
| 2016 | ### Removed |
| 2017 | |
| 2018 | - **Unwired `[context.per_model]` config field** — the field had no |
| 2019 | runtime consumer and was only present in the config schema. Removed |
| 2020 | to keep the schema honest. Existing configs that still contain a |
| 2021 | `[context.per_model.*]` table continue to load (serde ignores |
| 2022 | unknown keys; covered by a regression test). |
| 2023 | - **Stale aspirational `[cycle.per_model]` comments** — reference to a |
| 2024 | config table that was never wired. No behavior change. |
| 2025 | |
| 2026 | ### Documentation |
| 2027 | |
| 2028 | - **`.claude/CODEMAP_v0.8.25_dead_code.md`** — committed the |
| 2029 | cycle/seam/coherence/capacity codemap with a softened |
| 2030 | `cycle_manager` classification: live by code trace, design |
| 2031 | load-bearing, practical load-bearing unproven. Use this to decide |
| 2032 | the v0.8.26+ product direction for the cycle/seam/capacity |
| 2033 | subsystems. |
| 2034 | |
| 2035 | ### Known issues |
| 2036 | |
| 2037 | - **Windows 10 conhost flicker regression (#1260, #1251)** — |
| 2038 | v0.8.22-and-later content flickering on Windows 10 is still present. |
| 2039 | The viewport-reset escape sequence added in v0.8.22 needs a Windows |
| 2040 | guard. Deferred to v0.8.26. |
| 2041 | - **Snapshot system still snapshots every turn** — the v0.8.24 500 MB |
| 2042 | hard cap protects against blowups, but the underlying design still |
| 2043 | snapshots on every turn regardless of whether the workspace changed. |
| 2044 | A write-aware skip is planned for v0.8.26. |
| 2045 | - **`▏` glyph leak in code blocks (#1212)**, **mouse selection |
| 2046 | crossing the sidebar (#1169)**, **drag-select edge auto-scroll |
| 2047 | (#1163)**, **mid-run MCP server stderr capture** — all deferred to |
| 2048 | v0.8.26. |
| 2049 | |
| 2050 | ## [0.8.24] - 2026-05-09 |
| 2051 | |
| 2052 | A bugfix + refactor release picking up the backlog after the v0.8.23 security |
| 2053 | release. Big thanks to **wplll** (cache-aware prompt + `/cache inspect`), |
| 2054 | **Liu-Vince** (MCP pagination diagnosis), **@Giggitycountless** (snapshot cap |
| 2055 | proposal), and to issue reporters **@SamhandsomeLee**, |
| 2056 | **@barjatiyasaurabh**, **@tyculw**, **@hongyuatcufe**, and **@ljlbit** for |
| 2057 | the bugs fixed below. |
| 2058 | |
| 2059 | ### Fixed |
| 2060 | |
| 2061 | - **Mouse-wheel scroll survives focus toggles** — on macOS, switching away |
| 2062 | (Cmd+Tab, opening the screenshot tool, etc.) and back can drop the |
| 2063 | terminal's mouse-tracking mode, leaving wheel scroll dead until restart. |
| 2064 | The TUI now re-arms `EnableMouseCapture` on `FocusGained` alongside the |
| 2065 | existing keyboard-mode recapture, so wheel events keep flowing after a |
| 2066 | focus round-trip. |
| 2067 | - **Workspace-local slash commands are now loaded (#1259)** — user command |
| 2068 | files placed in `<workspace>/.deepseek/commands/`, |
| 2069 | `<workspace>/.claude/commands/`, and `<workspace>/.cursor/commands/` are |
| 2070 | now discovered alongside the existing global `~/.deepseek/commands/`. |
| 2071 | Workspace-local commands shadow global by name, matching the precedence |
| 2072 | model already used for skills. Reported by **@SamhandsomeLee**. |
| 2073 | - **`@`-mention completion finds AI-tool dot-directories** — files inside |
| 2074 | `.deepseek/`, `.cursor/`, `.claude/`, and `.agents/` are now discoverable |
| 2075 | in `@`-mention Tab-completion even when those directories are excluded by |
| 2076 | `.gitignore`. The fix also applies to the Ctrl+P file picker and fuzzy |
| 2077 | file resolution. |
| 2078 | - **MCP paginated discovery (#1250, #1256)** — tools, resources, resource |
| 2079 | templates, and prompts from MCP servers that paginate their responses |
| 2080 | (e.g., gbrain at 5 items per page) are now fully discovered by following |
| 2081 | the MCP spec's `nextCursor` across all pages. Reported by |
| 2082 | **@hongyuatcufe**; thanks to **Liu-Vince** for the diagnosis and PR |
| 2083 | #1256 with the same fix shape. |
| 2084 | - **Snapshot storage has a disk-space cap (#1112)** — the snapshot side repo |
| 2085 | now enforces a 500 MB hard limit. When the limit is exceeded at snapshot |
| 2086 | time, the oldest snapshots are pruned aggressively to stay under a 400 MB |
| 2087 | target. Guards against the reported 1.2 TB snapshot blowup during |
| 2088 | high-churn sessions. Reported by **@tyculw**; thanks to |
| 2089 | **@Giggitycountless** for the PR #1131 proposal that informed the |
| 2090 | hard-cap approach. |
| 2091 | - **`/clear` now resets the Todos sidebar (#1258)** — previously `/clear` |
| 2092 | only reset the Plan panel; the Todos checklist persisted across clears |
| 2093 | until app restart. The fix ensures `clear_todos()` clears the |
| 2094 | `SharedTodoList` inner state. Reported by **@barjatiyasaurabh**. |
| 2095 | |
| 2096 | ### Added |
| 2097 | |
| 2098 | - **Cache-aware prompt diagnostics + payload optimization (#1196)** — adds |
| 2099 | a `PromptBuilder` that classifies the system prompt into `static` / |
| 2100 | `history` / `dynamic` layers for cache-prefix stability, plus: |
| 2101 | - `/cache inspect` — shows SHA-256 hashes per layer, base static prefix |
| 2102 | hash vs full request prefix hash, static-prefix stability across |
| 2103 | turns, and first-divergence tracking. Does not print prompt text. |
| 2104 | - `/cache warmup` — prefetches the stable prefix to seed the DeepSeek |
| 2105 | context cache. |
| 2106 | - **Project Context Pack injected into the stable prefix by default** |
| 2107 | — a structured workspace summary (directory listing up to 4 levels / |
| 2108 | 400 entries, README excerpt up to 4 KB, config + key source file |
| 2109 | lists). Adds **~1–10 KB to every prompt depending on repo size**, in |
| 2110 | exchange for a much more cacheable prefix. **Default ON**; disable |
| 2111 | with `[context] project_pack = false` in `~/.deepseek/config.toml` |
| 2112 | if you'd rather keep prompts minimal. |
| 2113 | - Wire-payload optimization: large tool outputs are budgeted, repeated |
| 2114 | identical tool outputs and `<turn_meta>` blocks are deduplicated |
| 2115 | with stable refs (wire-only — local session messages stay intact). |
| 2116 | - Footer cache-hit % chip from `prompt_cache_hit_tokens` / |
| 2117 | `prompt_cache_miss_tokens` in the API response. |
| 2118 | |
| 2119 | Thanks **wplll** for the design and implementation. |
| 2120 | |
| 2121 | ### Changed |
| 2122 | |
| 2123 | - **Language directive strengthened against project-context bias (#1118)** |
| 2124 | — the system prompt now explicitly instructs the model that project |
| 2125 | context (AGENTS.md, auto-generated instructions, file trees) is NOT a |
| 2126 | language signal. Chinese filenames in a repo no longer bias the model |
| 2127 | toward Chinese replies when the user writes in English. Reported by |
| 2128 | **@ljlbit**. |
| 2129 | |
| 2130 | ### Known issues |
| 2131 | |
| 2132 | - **Windows flicker/shake regression (#1260, #1251)** — v0.8.22 and v0.8.23 |
| 2133 | exhibit content flickering on Windows 10 (v0.8.20 works correctly). The |
| 2134 | issue is likely caused by the viewport-reset escape sequence |
| 2135 | (`\x1b[r\x1b[?6l\x1b[H\x1b[2J\x1b[3J`) added in v0.8.22 to fix viewport |
| 2136 | drift. On Windows conhost, this sequence may trigger a full screen clear |
| 2137 | on every repaint. A platform guard or less aggressive sequence is needed. |
| 2138 | |
| 2139 | ## [0.8.23] - 2026-05-08 |
| 2140 | |
| 2141 | A security-focused follow-up to v0.8.22. The bulk of the diff is hardening of |
| 2142 | the child-process surface — shells, MCP stdio servers, and other spawned |
| 2143 | subprocesses — plus a related set of MCP, secret-store, and tool-policy |
| 2144 | fixes uncovered during follow-up review. |
| 2145 | |
| 2146 | ### Security |
| 2147 | |
| 2148 | - **Sanitized child-process environments** - shells, MCP stdio servers, hooks, |
| 2149 | and other child processes spawned from the TUI now start from an explicit |
| 2150 | allowlist of parent environment variables rather than inheriting every |
| 2151 | parent var. The base allowlist covers `PATH`, `HOME`, `USER`, `LANG`/`LC_*`, |
| 2152 | `TERM`/`COLORTERM`, `SHELL`, `TMPDIR`/`TMP`/`TEMP`, and the corresponding |
| 2153 | Windows variables. Stops casual exfiltration of `*_API_KEY`, `AWS_*`, |
| 2154 | `GITHUB_TOKEN`, and similar through a spawned subprocess. |
| 2155 | - **Tighter shell safety classification** - the `exec_shell` deny-list was |
| 2156 | reviewed and broadened to cover additional dangerous command patterns. |
| 2157 | - **Plan mode tool surface narrowed** - planning sub-agents see a smaller, |
| 2158 | read-only tool surface so a plan-mode call can no longer mutate workspace |
| 2159 | state. |
| 2160 | - **Sub-agent approval boundaries preserved** - sub-agents inherit the |
| 2161 | parent's approval policy and cannot escalate beyond it. |
| 2162 | - **Symlinked workspace walks no longer followed** - workspace-relative |
| 2163 | walkers (file-search, project context) now refuse to traverse symlinks |
| 2164 | pointing outside the workspace root. |
| 2165 | - **Path and output handling tightened** - several tools that build paths |
| 2166 | from model output now reject `..` segments and absolute paths outside the |
| 2167 | workspace. |
| 2168 | - **Runtime API requires authentication by default** - `deepseek serve --http` |
| 2169 | no longer accepts unauthenticated requests in its default configuration. |
| 2170 | - **Security-sensitive dependencies bumped** - routine bump pass for crates |
| 2171 | with recent advisories. |
| 2172 | - **MCP config paths reject traversal** - `load_config`/`save_config` now |
| 2173 | refuse paths containing `..` components. |
| 2174 | - **Hardened `run_tests` approval policy.** Thanks to **@47Cid** for the |
| 2175 | responsible disclosure. |
| 2176 | |
| 2177 | ### Fixed |
| 2178 | |
| 2179 | - **macOS Keychain prompt at startup** - the file-backed secret store is now |
| 2180 | the default. The OS keyring is opt-in via |
| 2181 | `DEEPSEEK_SECRET_BACKEND=system|keyring`, and the auth status surface |
| 2182 | refers to "secret store" rather than "keyring" where appropriate. |
| 2183 | - **MCP stdio spawn errors are now visible (#1244)** - when spawning a stdio |
| 2184 | MCP server fails (e.g., `npx` not on `PATH`), the underlying OS error is |
| 2185 | now shown ("No such file or directory (os error 2)") instead of the opaque |
| 2186 | wrapper "MCP stdio spawn failed (...)". The fix applies to the snapshot, |
| 2187 | the `mcp connect` / `mcp validate` CLI commands, and the in-TUI status |
| 2188 | events. |
| 2189 | - **MCP servers no longer break under env scrub (#1244)** - MCP stdio launches |
| 2190 | now inherit a wider env allowlist than arbitrary shell tools, so common |
| 2191 | `npx ...`, `uvx ...`, `python -m mcp_server_*`, and proxy-bound corporate |
| 2192 | setups keep working under the new child-env scrub. Pass-through includes |
| 2193 | `NVM_DIR`, `NODE_OPTIONS`, `NODE_PATH`, `NODE_EXTRA_CA_CERTS`, |
| 2194 | `NPM_CONFIG_*`, `VOLTA_HOME`, `COREPACK_HOME`, `PYTHONPATH`, `PYTHONHOME`, |
| 2195 | `VIRTUAL_ENV`, `PIPX_*`, `POETRY_HOME`, `UV_*`, `GEM_*`, `BUNDLE_*`, |
| 2196 | `JAVA_HOME`, `HTTP_PROXY` / `HTTPS_PROXY` / `NO_PROXY` / `ALL_PROXY` / |
| 2197 | `FTP_PROXY` (case-insensitive), `SSL_CERT_FILE`, `SSL_CERT_DIR`, |
| 2198 | `REQUESTS_CA_BUNDLE`, `CURL_CA_BUNDLE`. Secret-bearing parent env stays |
| 2199 | scrubbed. |
| 2200 | |
| 2201 | ### Changed |
| 2202 | |
| 2203 | - **Live thinking is compact by default** - the streaming "thinking" panel |
| 2204 | collapses by default; expand via the existing details toggle. |
| 2205 | |
| 2206 | ### Added |
| 2207 | |
| 2208 | - **`docs/RELEASE_CHECKLIST.md`** - explicit pre-tag checklist (CHANGELOG, |
| 2209 | versions, preflight, npm wrapper smoke) so the v0.8.21/v0.8.22 CHANGELOG |
| 2210 | gap does not recur. |
| 2211 | |
| 2212 | ### Known issues |
| 2213 | |
| 2214 | - **Mid-run MCP server stderr is still suppressed** - if a stdio MCP server |
| 2215 | spawns successfully but exits later (e.g., crashes during `initialize`), |
| 2216 | its stderr is not yet captured. Spawn-time OS errors (the most common |
| 2217 | case from #1244) are visible. Full mid-run stderr capture is planned for |
| 2218 | v0.8.24. |
| 2219 | |
| 2220 | ## [0.8.22] - 2026-05-08 |
| 2221 | |
| 2222 | A focused security release. |
| 2223 | |
| 2224 | ### Security |
| 2225 | |
| 2226 | - **Hardened `fetch_url` redirect handling.** Thanks to **@47Cid** for the |
| 2227 | responsible disclosure. |
| 2228 | |
| 2229 | ## [0.8.21] - 2026-05-08 |
| 2230 | |
| 2231 | A community-heavy release rolling up two weeks of contributor PRs across the |
| 2232 | TUI, runtime, and docs. Big thanks to **Reid (@reidliu41)**, |
| 2233 | **jiaren wang (@JiarenWang)**, **Friende (@pengyou200902)**, |
| 2234 | **ZzzPL (@Oliver-ZPLiu)**, **Sun**, **Liu-Vince**, **kitty**, and |
| 2235 | **Aqil Aziz** for the contributions below. |
| 2236 | |
| 2237 | ### Added |
| 2238 | |
| 2239 | - **Distinct user-message body color** (#1168) - user turns now render in a |
| 2240 | green body color so the conversation flow is easier to scan at a glance. |
| 2241 | |
| 2242 | ### Fixed |
| 2243 | |
| 2244 | - **Plan mode enforces read-only tool boundaries** (#1114) - planning calls |
| 2245 | can no longer reach into write-side tools. Thanks **jiaren wang**. |
| 2246 | - **Composer arrow keys navigate input history** (#1117) - up/down in the |
| 2247 | composer cycles through prior prompts when the cursor is on the first/last |
| 2248 | line. Thanks **Reid**. |
| 2249 | - **RLM preserves prompt cache usage** (#1127) - the RLM batch path no longer |
| 2250 | resets prompt-cache hits between calls. Thanks **Sun**. |
| 2251 | - **`fetch_url` proxy DNS opt-in** (#1103) - the proxy DNS path is now opt-in |
| 2252 | rather than always forced, fixing breakage in environments where the proxy |
| 2253 | cannot resolve the target host. Thanks **Sun**. |
| 2254 | - **Undo syncs session context after snapshot restore** (#1150, fixes #1139) - |
| 2255 | rolling back a turn now correctly resyncs the in-memory session so a |
| 2256 | follow-up turn doesn't see stale context. Thanks **jiaren wang**. |
| 2257 | - **Stale busy-state watchdog** (#1170) - the TUI now recovers if the busy |
| 2258 | indicator gets stuck after an aborted turn. Thanks **ZzzPL**. |
| 2259 | - **`gh` discovered across common install paths** - the `gh` tool is found |
| 2260 | whether installed via Homebrew, apt, the Windows MSI, or the GitHub CLI |
| 2261 | installer. Thanks **kitty**. |
| 2262 | - **Code block indentation preserved in transcript** - leading whitespace |
| 2263 | inside fenced code blocks is no longer collapsed during rendering. |
| 2264 | Thanks **Liu-Vince**. |
| 2265 | - **Stream pacing preserves upstream cadence** - long streaming responses |
| 2266 | no longer chunk together when the upstream is bursty. |
| 2267 | Thanks **Sun**. |
| 2268 | - **Task list output gets headers** - the long-form `/tasks` output now has |
| 2269 | group headers so it scans cleanly. Thanks **Reid**. |
| 2270 | - **macOS option-V details shortcut** - the details toggle now works correctly |
| 2271 | on US Mac keyboards where Option+V produces `√`. |
| 2272 | - **Uppercase approval shortcuts accepted** - `[A]/[D]/[V]` work in either |
| 2273 | case in the approval dialog. |
| 2274 | - **Transcript scrollbar inert** - the transcript scrollbar no longer captures |
| 2275 | clicks intended for content below it. |
| 2276 | - **Hide transcript rail before code blocks** - the rail glyph no longer |
| 2277 | bleeds onto the line just above a fenced code block. |
| 2278 | - **Pager exit hint prominent** - the "press q to exit" hint is now visible |
| 2279 | on the pager footer. |
| 2280 | - **Empty tool call names fall back to a placeholder** - a model that returns |
| 2281 | an empty `function.name` in a tool call no longer hangs the turn. |
| 2282 | - **MCP SSE waits for endpoint before connect returns** (#1225) - the SSE |
| 2283 | transport no longer reports "connected" before the endpoint event has been |
| 2284 | received, fixing a race where the first request was lost. |
| 2285 | - **Git branch status item renders** (#1226, fixes #1217) - the |
| 2286 | `StatusItem::GitBranch` toggle now produces a footer entry instead of a |
| 2287 | blank slot. |
| 2288 | - **Beta endpoint routes non-beta paths to v1** (#1174) - paths that aren't |
| 2289 | available on the DeepSeek beta host are transparently redirected to the v1 |
| 2290 | host instead of failing. |
| 2291 | - **Skill packs accept workflow-pack archive layouts** (#1164) - skill |
| 2292 | archives produced by the workflow pack tool now install correctly. |
| 2293 | - **Interactive sessions stay in alternate screen** (#1158) - returning from |
| 2294 | a sub-process no longer kicks the TUI back to the primary screen mid-turn. |
| 2295 | - **Slash-menu arrow navigation wraps** (#1152) - up at the top / down at the |
| 2296 | bottom of the slash menu wraps to the other end. |
| 2297 | - **CLI preserves split prompt words from Windows shims** (#1160) - prompt |
| 2298 | arguments forwarded by the npm wrapper on Windows are no longer joined into |
| 2299 | one giant token. |
| 2300 | - **`libc` extended to all Unix targets** (#1173) - improves FreeBSD build |
| 2301 | compatibility. |
| 2302 | - **Memory truncation marker reports omitted bytes** - the `[…N bytes |
| 2303 | omitted]` marker now shows an accurate count. Thanks **Friende**. |
| 2304 | |
| 2305 | ### Docs |
| 2306 | |
| 2307 | - **Memory skill link** (#1096) - corrected. Thanks **Aqil Aziz**. |
| 2308 | - **Help keybinding reference** (#1095) - corrected. Thanks **Friende**. |
| 2309 | - **Additional environment variables** documented in the config reference. |
| 2310 | Thanks **Liu-Vince**. |
| 2311 | - **Docker volume guidance** - the install snippet now uses a writable named |
| 2312 | data volume rather than a bind mount that may be read-only on some hosts. |
| 2313 | - **Competitive analysis reflects LSP diagnostics** (#1171) - the doc now |
| 2314 | matches the shipping LSP diagnostics implementation. |
| 2315 | - **Dispatcher path for `/run-pr`** (#1227) - the README now points at the |
| 2316 | dispatcher binary. |
| 2317 | |
| 2318 | ## [0.8.20] - 2026-05-08 |
| 2319 | |
| 2320 | ### Added |
| 2321 | - **Global AGENTS.md fallback** - when a workspace and its parents do not |
| 2322 | provide project instructions, DeepSeek TUI now loads `~/.deepseek/AGENTS.md` |
| 2323 | before falling back to auto-generated `.deepseek/instructions.md`, keeping |
| 2324 | repo-local instructions higher priority while supporting shared defaults. |
| 2325 | |
| 2326 | ### Fixed |
| 2327 | - **Chinese reasoning stays Chinese** - restore the #588 language contract after |
| 2328 | the deterministic environment prompt regressed it. The latest user message now |
| 2329 | chooses the natural language for both `reasoning_content` and the final reply; |
| 2330 | the resolved `lang` field is only a fallback when the user turn is ambiguous. |
| 2331 | |
| 2332 | ## [0.8.19] - 2026-05-08 |
| 2333 | |
| 2334 | ### Fixed |
| 2335 | - **DeepSeek beta endpoint stays default for Chinese locales** - the legacy |
| 2336 | `deepseek-cn` runtime path no longer routes users to the non-beta |
| 2337 | `https://api.deepseek.com` base URL. It is now a backwards-compatible alias |
| 2338 | for the normal `deepseek` provider default, `https://api.deepseek.com/beta`, |
| 2339 | so strict tool mode and other beta-gated features stay available worldwide. |
| 2340 | - **Provider docs stop advertising `deepseek-cn` as a separate provider** - |
| 2341 | runtime docs now describe it only as a legacy config alias. DeepSeek uses the |
| 2342 | same official host worldwide; users with private mirrors should set |
| 2343 | `base_url` explicitly. |
| 2344 | |
| 2345 | ## [0.8.18] - 2026-05-07 |
| 2346 | |
| 2347 | This is the v0.8.17 follow-up release: a tighter TUI/runtime/install pass with |
| 2348 | safer session startup semantics, Docker images promoted to a supported install |
| 2349 | path, and several community PRs harvested into the release branch. VS Code and |
| 2350 | Feishu/Lark/mobile companion work remain out of scope for this release. |
| 2351 | |
| 2352 | ### Added |
| 2353 | - **Prebuilt Docker images on GHCR** - release builds now publish |
| 2354 | `ghcr.io/hmbown/deepseek-tui` with `latest`, semver, and `vX.Y.Z` tags, and |
| 2355 | the GitHub release notes include a Docker install snippet. Docker publishing |
| 2356 | is now a release gate rather than a best-effort check. |
| 2357 | - **Draggable transcript scrollbar** (#1075, #1076) - when mouse capture is |
| 2358 | enabled, drag the transcript scrollbar thumb to move through long sessions. |
| 2359 | The implementation also clears stale drag state on resize and new left-clicks. |
| 2360 | Thanks @Oliver-ZPLiu. |
| 2361 | - **PTY regression for viewport drift** (#1085) - the QA harness now covers the |
| 2362 | blank-top-rows failure after a failed/long turn so future layout changes catch |
| 2363 | terminal viewport drift. |
| 2364 | |
| 2365 | ### Changed |
| 2366 | - **Plain `deepseek` starts a fresh session** - opening a second `deepseek` in |
| 2367 | the same folder no longer silently attaches to the same in-flight checkpoint. |
| 2368 | Crash/interrupted checkpoints are preserved as saved sessions and recovered |
| 2369 | explicitly through `deepseek --continue`. |
| 2370 | - **npm postinstall is recoverable for transient download failures** (#1059) - |
| 2371 | install-time GitHub download/extract failures are non-blocking and documented, |
| 2372 | while unsupported platforms, checksum mismatches, glibc preflight failures, |
| 2373 | and runtime wrapper failures remain fatal. Thanks @Fire-dtx. |
| 2374 | - **Docker Buildx cargo caches are platform-isolated and locked** - registry, |
| 2375 | git, and target caches now use platform-specific cache IDs plus locked |
| 2376 | sharing to avoid the `.cargo-ok File exists` unpack race in release checks. |
| 2377 | - **Long-session palette is easier to read** (#1070, #936 partial) - default |
| 2378 | body text is slightly softer, reasoning/thinking text uses a warmer accent, |
| 2379 | and `/theme` now updates the terminal color adapter so light mode keeps those |
| 2380 | contrasts coherent after an in-session toggle. Thanks @bevis-wong and |
| 2381 | @oooyuy92 for the readability reports. |
| 2382 | - **Install docs add a second rustup mirror fallback** (#1011) - `rsproxy.cn` |
| 2383 | is documented as an alternate rustup mirror, and old Debian/Ubuntu Cargo |
| 2384 | `edition2024` failures now point users to rustup stable. Thanks @wuwuzhijing. |
| 2385 | |
| 2386 | ### Fixed |
| 2387 | - **Chinese destructive approval dialogs keep explicit risk wording** (#1087, |
| 2388 | #1091) - zh-Hans destructive approval copy now localizes the operation label, |
| 2389 | title, prompt, and destructive-risk warning without changing English default |
| 2390 | behavior. Thanks @qinxianyuzou and @axobase001. |
| 2391 | - **Terminal viewport is reset before repaint** (#1085) - the TUI now clears |
| 2392 | scroll margins/origin mode before key repaints after resume, resize, and turn |
| 2393 | completion, preventing alt-screen content from drifting downward and leaving |
| 2394 | blank rows at the top. |
| 2395 | - **Interactive subprocesses wait for terminal release** (#1085) - shell/editor |
| 2396 | handoff now waits until the UI has actually left alt-screen/raw mode before |
| 2397 | launching the child process, preventing the TUI from repainting into host |
| 2398 | scrollback after interactive tool use. |
| 2399 | - **Light theme reasoning blocks stay light** (#1070, #936 partial) - |
| 2400 | thinking/reasoning background tints now map to the light reasoning surface |
| 2401 | instead of keeping the dark-mode tint after `/theme light`. |
| 2402 | - **FreeBSD can compile the secrets crate** (#1089) - platforms without a native |
| 2403 | `keyring` dependency now fail the OS-keyring probe cleanly and fall back to |
| 2404 | the file-backed secret store instead of referencing a missing crate. Thanks |
| 2405 | @avysk for the FreeBSD report. |
| 2406 | - **Windows sandbox docs no longer overstate guarantees** (#1015, #1058) - the |
| 2407 | docs and code comments now describe the future Windows helper as |
| 2408 | process-tree containment only until filesystem, network, registry, or |
| 2409 | AppContainer isolation is actually implemented. Thanks @axobase001. |
| 2410 | |
| 2411 | ## [0.8.17] - 2026-05-07 |
| 2412 | |
| 2413 | A focused reliability release built almost entirely from community contributions. |
| 2414 | Fixes Plan-mode safety, paste-Enter auto-submit, slash-menu skills coverage, the |
| 2415 | `deepseek-cn` endpoint preset, and a handful of platform / streaming / |
| 2416 | gateway-compatibility issues. Also lands a small PTY-driven QA harness so the |
| 2417 | next round of TUI fixes can be verified against real terminal behaviour. |
| 2418 | |
| 2419 | ### Added |
| 2420 | - **`/theme` command** (#1057) — toggle between dark and light themes inline, |
| 2421 | without round-tripping through `/config`. Thanks @MengZ-super. |
| 2422 | - **PTY/frame-capture TUI QA harness** — new |
| 2423 | `crates/tui/tests/support/qa_harness/` lets integration tests spawn |
| 2424 | `deepseek-tui` in a real pseudo-terminal, send scripted keys / paste / |
| 2425 | resize, and assert on the parsed terminal frame plus the workspace |
| 2426 | filesystem. Initial scenarios cover boot smoke and the #1073 paste regression. |
| 2427 | Adding-a-scenario walkthrough lives in `crates/tui/tests/support/qa_harness/README.md`. |
| 2428 | - **Whalescale desktop runtime bridge** — the local runtime API now exposes |
| 2429 | `POST /v1/approvals/{id}`, `GET /v1/runtime/info`, `enabled` flags on |
| 2430 | `GET /v1/skills`, and `POST /v1/skills/{name}` toggles. Runtime thread |
| 2431 | events also carry `agent_reasoning` items so desktop clients can render |
| 2432 | thinking separately from assistant text. |
| 2433 | |
| 2434 | ### Changed |
| 2435 | - **`deepseek-cn` provider preset now defaults to the official |
| 2436 | `https://api.deepseek.com` host** (#1079, #1084) — matches |
| 2437 | [api-docs.deepseek.com](https://api-docs.deepseek.com/). The legacy typo |
| 2438 | host `api.deepseeki.com` is still recognized in URL heuristics and chat-client |
| 2439 | normalization so existing user configs keep working. Thanks @Jefsky. |
| 2440 | - **Plan mode runs shell commands in a read-only sandbox** (#1077) — was |
| 2441 | `WorkspaceWrite` with the workspace as a writable root, which let |
| 2442 | `python -c "open('f','w').write('x')"` mutate files inside the workspace. |
| 2443 | Now `SandboxPolicy::ReadOnly`: no writes anywhere on the filesystem, no |
| 2444 | network. Read-only inspection commands (`ls`, `git log`, `grep`, |
| 2445 | `cargo metadata`, …) keep working through the per-platform sandbox; for |
| 2446 | anything that creates or modifies files, switch to Agent mode (`/agent`). |
| 2447 | Thanks @DI-HUO-MING-YI. |
| 2448 | |
| 2449 | ### Fixed |
| 2450 | - **Pasting multi-line text with a trailing newline no longer auto-submits** |
| 2451 | (#1073) — the composer's Enter handler now consults the paste-burst |
| 2452 | suppression state and either appends `\n` to the in-flight burst buffer or |
| 2453 | inserts it into the composer text directly, instead of falling through to |
| 2454 | `submit_input()`. Reproduced from the original Windows / PowerShell |
| 2455 | symptom; fix covers both the bracketed-paste and rapid-keystroke detection |
| 2456 | paths. Thanks @bevis-wong for the precise reproducer. |
| 2457 | - **Slash menu, `/skills`, and `/skill <name>` show project-local AND global |
| 2458 | skills** (#1068, #1083) — switched the cache to `discover_in_workspace`, so |
| 2459 | the UI surfaces stay in sync with the system-prompt skills block. Bonus |
| 2460 | fix: `SKILL.md` frontmatter values are now stripped of surrounding YAML |
| 2461 | quotes, so `name: "hud"` registers as `hud` and matches prefix lookup. |
| 2462 | Thanks @AlphaGogoo / @Duducoco. |
| 2463 | - **Windows shell output is decoded as UTF-8 even on non-UTF-8 system code |
| 2464 | pages** (#982, #1018) — Windows shell commands are now wrapped with |
| 2465 | `chcp 65001 >NUL & ` so subprocesses output UTF-8 instead of GBK / other |
| 2466 | ANSI code pages. `display_command` strips the prefix so transcripts and |
| 2467 | approval prompts stay clean. Thanks @chnjames. |
| 2468 | - **Stale snapshot `tmp_pack_*` files are cleaned up on startup** (#975, |
| 2469 | #1055) — interrupted side-repo git pack operations no longer leak orphaned |
| 2470 | temp files; `prune_unreachable_objects` runs during the regular prune |
| 2471 | cycle to drop loose objects from rolled-back snapshots. Closes the |
| 2472 | ~30 GB+ disk-usage report. Thanks @axobase001. |
| 2473 | - **Window-resize artifacts on macOS Terminal.app and Windows ConHost are |
| 2474 | gone** (#993) — forces the resize-event size during the post-resize draw |
| 2475 | so ratatui's internal `autoresize()` cannot shrink the viewport back to a |
| 2476 | stale dimension and leave the newly-expanded area filled with stale |
| 2477 | content. Same class as #582 for additional emulator families. Thanks |
| 2478 | @ArronAI007. |
| 2479 | - **Streaming thinking blocks finalize cleanly on stream errors and |
| 2480 | restarts** (#861 partial, #1078) — the engine-error handler now drains |
| 2481 | the in-flight thinking block into the transcript instead of leaving the |
| 2482 | partial reasoning orphaned in `StreamingState`. Refactor extracts the |
| 2483 | thinking lifecycle into named helpers (`start_streaming_thinking_block`, |
| 2484 | `finalize_current_streaming_thinking`, `stash_reasoning_buffer_into_last_reasoning`). |
| 2485 | Thanks @reidliu41. |
| 2486 | - **OpenRouter and other custom-endpoint providers preserve explicit model |
| 2487 | IDs** (#1066) — when a provider has an explicit model AND a custom |
| 2488 | `base_url` (different from the provider default), the model name is no |
| 2489 | longer rewritten by provider-specific normalization. Lets OpenAI-compatible |
| 2490 | gateways accept bare IDs like `deepseek/deepseek-v4-pro`, |
| 2491 | `accounts/fireworks/models/...`, or `glm-5`. Thanks @THINKER-ONLY. |
| 2492 | - **Auto-generated `.deepseek/instructions.md` stabilizes the KV prefix |
| 2493 | cache** (#1080) — replaces the per-turn filesystem-scan fallback in |
| 2494 | `prompts.rs` with a real on-disk artifact when no context file exists, so |
| 2495 | the system prompt's prefix stays byte-stable across turns and prefix-cache |
| 2496 | hit-rate improves. The auto-generated file is plainly labelled and the |
| 2497 | user can edit or delete it freely. Thanks @lloydzhou. |
| 2498 | - **SSE responses behind compressing gateways decode correctly** (#1061) — |
| 2499 | enables reqwest's `gzip` and `brotli` features so streams through proxies |
| 2500 | that compress the response come through clean instead of as protocol |
| 2501 | corruption. Quiets one of the failure modes behind some "stuck working" |
| 2502 | reports. Thanks @MengZ-super. |
| 2503 | - **NVIDIA NIM provider configs use their own API key even when a legacy |
| 2504 | root DeepSeek key is present** (#1081) — `[providers.nvidia_nim] api_key` |
| 2505 | now wins for NIM requests, avoiding 401s caused by accidentally sending the |
| 2506 | top-level DeepSeek credential to NVIDIA. Thanks @wlon for the focused |
| 2507 | diagnosis. |
| 2508 | - **npm installs explain the release-mirror escape hatch when GitHub Releases |
| 2509 | are blocked** (#1051, #1056) — network/DNS failures now point at the |
| 2510 | existing `DEEPSEEK_TUI_RELEASE_BASE_URL` override and the required checksum |
| 2511 | manifest / binary layout instead of stopping at a raw `ENOTFOUND github.com`. |
| 2512 | Thanks @axobase001. |
| 2513 | |
| 2514 | ### Notes for contributors |
| 2515 | |
| 2516 | This release shifts the project's PR-handling philosophy: every contribution |
| 2517 | has value somewhere; the maintainer's job is to find it, use it, and credit |
| 2518 | the contributor — never to close a PR with nothing taken. If a PR is too |
| 2519 | large or scope-mixed to merge whole, useful commits / files / ideas are |
| 2520 | harvested directly rather than asking the contributor to split it. Trust |
| 2521 | boundary on credentials, sandbox, providers, publishing, telemetry, |
| 2522 | sponsorship, branding, and global prompts still requires explicit |
| 2523 | maintainer sign-off, but the burden of getting there is on us. See |
| 2524 | `AGENTS.md` for the full text. |
| 2525 | |
| 2526 | ## [0.8.16] - 2026-05-07 |
| 2527 | |
| 2528 | A focused hotfix for v0.8.15 regressions in RLM, sub-agent visibility, and |
| 2529 | terminal ownership. This release keeps the v0.8.15 feature set intact while |
| 2530 | making long-running delegated work easier to inspect and safer to run. |
| 2531 | |
| 2532 | ### Changed |
| 2533 | - **RLM has no fixed 180s wall-clock timeout** (#955) — RLM turns can continue |
| 2534 | past the old hard limit when the long-input REPL is still making progress. |
| 2535 | - **RLM output is easier to audit** (#955) — final reports now include compact |
| 2536 | execution metadata: input size, iteration count, elapsed time, sub-LLM RPC |
| 2537 | count, and termination state. |
| 2538 | - **RLM chunking guidance is stricter for exact work** (#955) — prompts now |
| 2539 | tell the sub-agent to use deterministic Python over the full `context` for |
| 2540 | counts/aggregation and to report chunk coverage when splitting a whole input. |
| 2541 | - **Tool guidance is less defensive** (#955) — the system prompt now explains |
| 2542 | when to use tools instead of discouraging the model from using capabilities |
| 2543 | that are actually available. |
| 2544 | |
| 2545 | ### Fixed |
| 2546 | - **Active RLM work stays visible** (#955) — foreground RLM calls surface in the |
| 2547 | active task/right-rail state instead of leaving the Tasks panel saying |
| 2548 | `No active tasks`. |
| 2549 | - **`/subagents` no longer reports false emptiness** (#955) — the sub-agent |
| 2550 | overlay now includes live progress-only agents and transcript fanout workers |
| 2551 | when the manager cache has not refreshed yet. |
| 2552 | - **Sub-agent cards are quieter and more useful** (#955) — low-signal scheduler |
| 2553 | lines such as `step 1/100: requesting model response` are hidden, while |
| 2554 | compact tool activity remains visible. |
| 2555 | - **Sub-agent completion protocol stays internal** (#955) — completion |
| 2556 | sentinels are routed as internal runtime events instead of user messages, so |
| 2557 | the parent agent does not explain raw protocol XML back to the user. |
| 2558 | - **Sub-agents cannot take over the parent terminal** (#955) — background |
| 2559 | agents reject `exec_shell` with `interactive=true`; they can still use |
| 2560 | non-interactive shell, background shell, `tty=true`, and task-shell tools. |
| 2561 | - **Terminal scrollback ownership is restored** (#955) — the TUI re-enters |
| 2562 | alternate-screen mode after foreground/sub-agent work drains, preventing the |
| 2563 | host terminal scrollbar from taking over the live interface. |
| 2564 | |
| 2565 | ## [0.8.15] - 2026-05-06 |
| 2566 | |
| 2567 | An auth, Windows, editor-integration, and setup stabilization release. This |
| 2568 | release keeps the existing DeepSeek V4 architecture intact while landing small |
| 2569 | community fixes that make first-run setup, terminal behavior, skills, cost |
| 2570 | display, and recovery paths easier to trust. |
| 2571 | |
| 2572 | ### Added |
| 2573 | - **ACP stdio adapter for Zed/custom agents** (#782) — `deepseek serve --acp` |
| 2574 | starts a local Agent Client Protocol server over stdio. The first slice |
| 2575 | supports new sessions and prompt responses through the user's existing |
| 2576 | DeepSeek config/API key; tool-backed editing and checkpoint replay remain |
| 2577 | outside the ACP surface for now. |
| 2578 | - **Yuan/CNY cost display** (#806) — `cost_currency = "cny"` (also accepts |
| 2579 | `yuan` / `rmb`) switches footer, context panel, `/cost`, `/tokens`, and |
| 2580 | long-turn notification summaries from USD to CNY. |
| 2581 | - **Slash autocomplete for skills** (#808) — installed skills are visible in |
| 2582 | the slash-command autocomplete menu. |
| 2583 | - **`/rename` session titles** (#836) — sessions can be renamed without |
| 2584 | editing save files manually. |
| 2585 | |
| 2586 | ### Changed |
| 2587 | - **Current local date in turn metadata** (#893, closes #865) — real user turns |
| 2588 | now include the current local date in `<turn_meta>`, without changing the |
| 2589 | stable system prompt/cache prefix. |
| 2590 | - **Doctor endpoint diagnostics** (#823) — `deepseek doctor` shows the resolved |
| 2591 | provider/API endpoint to make proxy, China endpoint, and inherited-env |
| 2592 | debugging more concrete. |
| 2593 | - **More conservative request sizing** (#826) — API requests cap `max_tokens` |
| 2594 | against the active model/context budget before dispatch. |
| 2595 | - **Safer config and secret file writes** (#833, #837) — generated config files |
| 2596 | use restrictive permissions and improved secret redaction. |
| 2597 | |
| 2598 | ### Fixed |
| 2599 | - **Env-only API key failure recovery** (#892) — runtime auth failures now say |
| 2600 | when the rejected key came from inherited `DEEPSEEK_API_KEY` and no saved |
| 2601 | config key is present, matching the clearer `deepseek doctor` guidance. |
| 2602 | - **Windows Unicode output** (#887, closes #872) — TUI startup now best-effort |
| 2603 | switches the Windows console input/output codepages to UTF-8, improving |
| 2604 | Chinese and other non-ASCII rendering. |
| 2605 | - **Windows resume picker** (#886, closes #866) — the dispatcher keeps the |
| 2606 | resume picker path on Windows instead of bypassing it. |
| 2607 | - **Windows clipboard fallback** (#850) — copy operations have a fallback path |
| 2608 | when the primary clipboard backend is unavailable. |
| 2609 | - **Workspace trust persistence** (#870) — approval/trust choices persist in |
| 2610 | global config instead of surprising users on the next launch. |
| 2611 | - **Ctrl+E composer behavior** (#883, closes #876) — plain Ctrl+E moves to the |
| 2612 | end of the composer again; file-tree toggling moved to the shifted shortcut. |
| 2613 | - **Plain Markdown skills** (#869) — `SKILL.md` files without frontmatter now |
| 2614 | fall back to the first `# Heading` instead of being ignored. |
| 2615 | - **Workspace-scoped latest resume** (#830, closes #779) — `resume --last`, |
| 2616 | `--continue`, and fork/resume helpers choose the latest session for the |
| 2617 | current workspace/repo rather than the newest saved session globally. |
| 2618 | - **Npm wrapper version fallback** (#885) — `deepseek --version` / `-v` can |
| 2619 | report the package version when the native binary has not been downloaded |
| 2620 | yet. |
| 2621 | - **TUI exit resume hint** (#863, closes #682) — exiting the TUI now points |
| 2622 | users toward the relevant resume command. |
| 2623 | - **Startup and terminal reliability** — includes bounded stream-open waits |
| 2624 | (#847), cursor-lag reduction for `@` mentions (#849), OSC52 clipboard fallback |
| 2625 | for SSH (#845), legacy Ctrl+V paste recognition (#786), Windows mouse capture |
| 2626 | defaulting off (#785), and UTF-8-preserving ANSI stripping (#784). |
| 2627 | - **Install and policy reliability** — avoids unstable Rust file-locking APIs |
| 2628 | (#821), enforces network policy in `web_run` (#800), fixes repeated setup |
| 2629 | language prompts after API-key setup (#844), and explains dispatcher TUI spawn |
| 2630 | failures (#853). |
| 2631 | - **Workspace safety** — refuses dangerous snapshots for `$HOME` or unsafe |
| 2632 | workspaces (#798, #804), fixes path-escape false positives for double-dots in |
| 2633 | names (#824), scopes snapshot built-in excludes (#854), and replaces provider |
| 2634 | `unreachable!()` paths with proper errors (#835). |
| 2635 | - **Skills discovery** — recursively reads the skills directory (#811), ignores |
| 2636 | symlinks outside the selected install root (#814), discovers global Agents |
| 2637 | skills (#848), and includes `.cursor/skills` (#817). |
| 2638 | - **Provider/model compatibility** — restores auto model routing (#772), |
| 2639 | completes vLLM provider integration (#737), accepts provider-prefixed DeepSeek |
| 2640 | model IDs (#794), preserves requested model ID casing (#733), and pins RLM |
| 2641 | child calls to Flash (#832). |
| 2642 | |
| 2643 | ### Thanks |
| 2644 | - Thanks to [@reidliu41](https://github.com/reidliu41) for the resume hint and |
| 2645 | workspace trust fixes (#863, #870). |
| 2646 | - Thanks to [@Oliver-ZPLiu](https://github.com/Oliver-ZPLiu) for the Windows |
| 2647 | clipboard fallback (#850). |
| 2648 | - Thanks to [@xieshutao](https://github.com/xieshutao) for the plain Markdown |
| 2649 | skill fallback (#869). |
| 2650 | - Thanks to [@GK012](https://github.com/GK012) for the npm wrapper version |
| 2651 | fallback (#885). |
| 2652 | - Thanks to everyone filing Windows, Chinese-language setup, auth, and |
| 2653 | first-run reports. Those concrete reproductions shaped the release. |
| 2654 | |
| 2655 | ## [0.8.13] - 2026-05-05 |
| 2656 | |
| 2657 | A stabilization release for DeepSeek V4 runtime and TUI reliability. The |
| 2658 | v0.8.13 milestone was narrowed to direct runtime/TUI fixes; prompt hygiene, |
| 2659 | trajectory logging, Anthropic-wire support, and larger UI cleanup were moved |
| 2660 | out of this release. |
| 2661 | |
| 2662 | ### Added |
| 2663 | - **No-LLM tool-result prune before compaction** (#710) — old verbose tool |
| 2664 | results are mechanically summarized before the paid summary pass. Duplicate |
| 2665 | reads keep the freshest full body and replace older copies with one-line |
| 2666 | summaries; if that gets the session back under the compaction threshold, the |
| 2667 | LLM summary call is skipped entirely. |
| 2668 | - **Repeated-tool anti-loop guard** (#714) — the engine now tracks |
| 2669 | `(tool_name, args)` pairs per user turn. On the third identical call it |
| 2670 | inserts a synthetic corrective tool result instead of running the same tool |
| 2671 | again unchanged; per-tool failures warn at three and halt at eight. |
| 2672 | - **V4 cache-hit telemetry fallback** (#721) — usage parsing now recognizes |
| 2673 | `usage.prompt_tokens_details.cached_tokens`, so the existing footer cache-hit |
| 2674 | chip works with DeepSeek V4's automatic prefix-cache telemetry as well as the |
| 2675 | older explicit hit/miss fields. |
| 2676 | |
| 2677 | ### Fixed |
| 2678 | - **Invalid tool-call JSON repair** (#712) — malformed streamed tool arguments |
| 2679 | now pass through a deterministic repair ladder before dispatch. |
| 2680 | - **Hallucinated tool-name recovery** (#713) — common non-canonical tool names |
| 2681 | are resolved through the registry before the engine reports a missing tool. |
| 2682 | - **Tool-schema sanitation** (#715) — schemas are normalized before API |
| 2683 | emission so provider-strict JSON Schema handling does not reject valid tools. |
| 2684 | - **Case-sensitive model IDs** (#717, #729) — valid configured model IDs keep |
| 2685 | caller-provided case while compact DeepSeek aliases still canonicalize. |
| 2686 | - **Stale `working...` state after failed dispatch** (#738) — if the UI fails |
| 2687 | to send a message to the engine before a turn starts, the composer loading |
| 2688 | state is cleared instead of trapping later input in pending state. |
| 2689 | - **Prompt-free doctor key checks** — `deepseek doctor` no longer reads the OS |
| 2690 | keyring, avoiding macOS Keychain prompts during diagnostics. |
| 2691 | - **macOS Terminal color compatibility** — `xterm-256color` sessions now |
| 2692 | receive 256-color palette indexes instead of truecolor SGR, preventing |
| 2693 | Apple Terminal from misrendering whale blues as green/cyan blocks. |
| 2694 | - **Chat client repair after Responses cleanup** — restored the chat client |
| 2695 | body and regression coverage after removing the dead experimental Responses |
| 2696 | fallback path. |
| 2697 | - **Up/Down arrow transcript scroll when composer is empty** — bare Up/Down |
| 2698 | arrows now scroll the transcript when the composer input is empty (or |
| 2699 | whitespace-only); with text present they still navigate composer history. |
| 2700 | Previously the gate was hardcoded to false, leaving users in virtual |
| 2701 | terminals (Ghostty, Codex, Kitty-protocol) unable to scroll without |
| 2702 | modifier shortcuts. |
| 2703 | |
| 2704 | ## [0.8.11] - 2026-05-04 |
| 2705 | |
| 2706 | ### Changed |
| 2707 | - **Cache-maxing prompt path for DeepSeek V4** — the engine now skips |
| 2708 | system-prompt reassignment when the assembled stable prompt is unchanged, |
| 2709 | keeps the volatile repo working-set summary out of the system prompt, and |
| 2710 | injects it as per-turn metadata on the latest user message instead. |
| 2711 | - **Tool catalog cache anchor** — the model-visible tool array now marks |
| 2712 | the final native tool with `cache_control: ephemeral` so DeepSeek can |
| 2713 | anchor the stable tool prefix explicitly. |
| 2714 | - **V4-scale automatic compaction defaults** — automatic compaction keeps a |
| 2715 | 500K-token hard floor and the fallback compaction threshold now reflects |
| 2716 | the V4-scale late-trigger policy instead of the old 50K-era default. |
| 2717 | - **Token-only compaction trigger** — the message-count compaction trigger |
| 2718 | was a 128K-era heuristic that fired on long sessions of small messages |
| 2719 | — exactly the case where rewriting V4's prefix cache is most wasteful. |
| 2720 | Removed `CompactionConfig::message_threshold` and the message-count |
| 2721 | branch in `should_compact`; token budget is now the sole automatic |
| 2722 | trigger (gated by the 500K floor). Manual `/compact` is unchanged. |
| 2723 | |
| 2724 | ### Fixed |
| 2725 | - **Legacy 128K context naming** — the 128K fallback is now named and |
| 2726 | documented as legacy DeepSeek-only behavior, reducing ambiguity with the |
| 2727 | 1M-token DeepSeek V4 defaults. |
| 2728 | - **`npm install` resilience for slow / firewalled networks** — the |
| 2729 | postinstall binary fetch from GitHub Releases now retries on transient |
| 2730 | errors (5 attempts, 1-16 s exponential backoff with jitter), enforces a |
| 2731 | per-attempt timeout (default 5 min, configurable via |
| 2732 | `DEEPSEEK_TUI_DOWNLOAD_TIMEOUT_MS`) plus a 30 s stall detector, honors |
| 2733 | `HTTPS_PROXY` / `HTTP_PROXY` / `NO_PROXY` env vars (pure-Node CONNECT |
| 2734 | tunneling, no new dependencies), and prints a download-progress line |
| 2735 | to stderr so users know it isn't hung. Suppressible with |
| 2736 | `DEEPSEEK_TUI_QUIET_INSTALL=1`. Reported by a community user from China |
| 2737 | whose install through a CN npm mirror took 18 minutes — the bottleneck |
| 2738 | was the GitHub fetch, which CN npm mirrors do not proxy. |
| 2739 | - **YOLO sandbox dropped to DangerFullAccess** — YOLO mode was still |
| 2740 | routing shell commands through the WorkspaceWrite sandbox, which |
| 2741 | intercepted legitimate outside-workspace writes (package installs, |
| 2742 | sub-agent workspaces, `~/.cache`, brew, `npm install -g`, pipx) and |
| 2743 | forced approval round-trips — contradicting the "no guardrails" |
| 2744 | contract. YOLO already auto-approves all tools and enables trust mode; |
| 2745 | the sandbox was the last residual restriction. Now uses |
| 2746 | DangerFullAccess (no sandbox), consistent with the full YOLO posture. |
| 2747 | - **Scroll position lock preserved across render resolve** — user |
| 2748 | scroll-up during live streaming was being yanked back to the live tail |
| 2749 | on the next chunk. The `user_scrolled_during_stream` lock was cleared |
| 2750 | prematurely when content briefly fit in one screen, or when the |
| 2751 | transcript shrank between renders (e.g. sub-agent card collapsed). |
| 2752 | Fixed by snapshotting the prior tail state before `resolve_top` and |
| 2753 | only clearing the lock when the user was deliberately at the bottom. |
| 2754 | - **Capacity controller disabled by default** — the capacity controller |
| 2755 | was silently clearing the transcript (`messages.clear()`) based on |
| 2756 | slack-based `p_fail` calculations, independent of token utilization or |
| 2757 | the `auto_compact` setting. This contradicted the v0.8.11 default of |
| 2758 | `auto_compact = false` — the user opted into trusting the model with |
| 2759 | the full 1M-token V4 window, and the controller was auto-managing the |
| 2760 | prefix on their behalf. The controller now defaults to `enabled = false`; |
| 2761 | power users can opt in via `capacity.enabled = true`. |
| 2762 | |
| 2763 | ### Docs |
| 2764 | - **README clarity pass** (#685) — title-cased section headings, an explicit |
| 2765 | Node + npm prerequisites block before the `npm install -g` snippet, a |
| 2766 | China-friendly `--registry=https://registry.npmmirror.com` install |
| 2767 | variant, a DeepWiki badge for AI-assisted repo browsing, and a 🐳 mark |
| 2768 | on the title. *Thanks to [@Agent-Skill-007](https://github.com/Agent-Skill-007) |
| 2769 | for this PR.* |
| 2770 | |
| 2771 | ## [0.8.12] - 2026-05-05 |
| 2772 | |
| 2773 | A feature release built on the v0.8.11 cache-maxing foundation: 20 community |
| 2774 | PRs merged, covering reasoning-effort automation, V4 FIM edits, bash-arity |
| 2775 | execpolicy, skill-registry sync, vim composer mode, large-tool-output routing, |
| 2776 | pluggable sandbox backends, layered permission rulesets, and cache-aware |
| 2777 | resident sub-agents. No breaking changes. |
| 2778 | |
| 2779 | ### Added |
| 2780 | - **Reasoning-effort auto mode** (#669) — `reasoning_effort = "auto"` inspects |
| 2781 | the last user message for keywords (debug/error → Max, search/lookup → Low, |
| 2782 | default → High) and resolves the tier before each API request. Sub-agents |
| 2783 | always get Low. |
| 2784 | - **FIM edit tool for V4 /beta** (#668) — `fim_edit` tool sends |
| 2785 | fill-in-the-middle requests to DeepSeek's `/beta` endpoint for surgical code |
| 2786 | edits. |
| 2787 | - **Bash arity dictionary** (#655) — `auto_allow = ["git status"]` now matches |
| 2788 | `git status -s` but NOT `git push`. The arity dictionary knows command |
| 2789 | structure for git, cargo, npm, yarn, pnpm, docker, kubectl, aws, make, and |
| 2790 | others. Legacy flat prefix matching still works for unlisted commands. |
| 2791 | - **Unified slash-command namespace** (#661) — user-defined commands in |
| 2792 | `~/.deepseek/commands/` support `$1`, `$2`, `$ARGUMENTS` template |
| 2793 | substitution. User commands override built-in commands. |
| 2794 | - **Skill registry sync** (#654) — `/skills sync` fetches the community skill |
| 2795 | registry and installs/updates all listed skills. Network-gated by the |
| 2796 | existing `[network]` policy. |
| 2797 | - **Vim modal editing in composer** (#659) — `vim.insert_mode` / `vim.normal_mode` |
| 2798 | settings enable modal editing in the message composer with standard Vim |
| 2799 | keybindings. |
| 2800 | - **Separate tui.toml** (#657) — theme colors and keybind overrides can live in |
| 2801 | `~/.deepseek/tui.toml` alongside the main `config.toml`. *Note: file format |
| 2802 | is defined but not yet loaded at startup — wiring deferred to v0.8.13.* |
| 2803 | - **Large-tool-output routing** (#658) — tool results exceeding a configurable |
| 2804 | token threshold are routed through a workshop with truncated previews, |
| 2805 | protecting the parent context window. Synthesis is currently truncation-only; |
| 2806 | V4-Flash sub-agent synthesis deferred to follow-up. |
| 2807 | - **Pluggable sandbox backends** (#645) — a `SandboxBackend` trait and |
| 2808 | Alibaba OpenSandbox HTTP adapter let `exec_shell` route commands to a remote |
| 2809 | sandbox instead of spawning locally. Config keys: `sandbox_backend`, |
| 2810 | `sandbox_url`, `sandbox_api_key`. |
| 2811 | - **Layered permission rulesets** (#653) — `ExecPolicyEngine` supports |
| 2812 | builtin, agent, and user-priority layers for allow/deny prefix rules. |
| 2813 | Deny-always-wins semantics. |
| 2814 | - **Cache-aware resident sub-agents** (#660) — sub-agents spawned with |
| 2815 | `resident_file` prepend the file contents to their system prefix for V4 |
| 2816 | prefix-cache locality. A global lease table prevents two agents from holding |
| 2817 | a resident lease on the same file simultaneously. Leases are released on |
| 2818 | agent completion. |
| 2819 | - **Context-limit handoff** (#667) — engine-level support for replacing |
| 2820 | routine compaction with a `.deepseek/handoff.md` file write when context |
| 2821 | pressure triggers. *Note: config knob removed pending implementation.* |
| 2822 | - **LSP auto-attach diagnostics** (#656) — edit results now include post-edit |
| 2823 | diagnostics via the engine-level LSP hooks path. |
| 2824 | |
| 2825 | ### Docs |
| 2826 | - **README install section rewritten** (#672) — the previous lede claimed |
| 2827 | "no Node.js or Python runtime" but the very next paragraph told readers to |
| 2828 | install Node before continuing. Replaced with a three-path Install block |
| 2829 | (npm / cargo / direct download) that makes the npm wrapper's role explicit: |
| 2830 | it downloads the prebuilt binary, but `deepseek` itself does not depend on |
| 2831 | Node at runtime. zh-CN README mirrored. |
| 2832 | - **Windows Scoop install instructions** (#696) — README and zh-CN README now |
| 2833 | document `scoop install deepseek-tui` for Windows users. *Thanks to |
| 2834 | [@woyxiang](https://github.com/woyxiang) for this PR.* |
| 2835 | - **DeepSeek Pro discount window extended** (#692) — pricing footnote updated |
| 2836 | from 5 May 2026 to 31 May 2026 to match the platform-side promotion. *Thanks |
| 2837 | to [@wangfeng](mailto:wangfengcsu@qq.com) for this PR.* |
| 2838 | - **`deepseek resume <SESSION_ID>` surfaced in Usage** — the command exists |
| 2839 | since v0.7 but was undocumented. Reported via #682. |
| 2840 | - **SECURITY.md** (#648) — vulnerability reporting policy and supported |
| 2841 | versions. |
| 2842 | - **CODE_OF_CONDUCT.md** (#686) — Contributor Covenant v2.1. *Thanks to |
| 2843 | [@zichen0116](https://github.com/zichen0116) for this PR.* |
| 2844 | - **zh-Hans locale activation docs** (#652) — README.zh-CN.md and |
| 2845 | config.example.toml now document `locale = "zh-Hans"`. |
| 2846 | |
| 2847 | ### Fixed |
| 2848 | - **Cross-workspace session bleed (security)** — launching `deepseek` from |
| 2849 | any directory silently auto-recovered the most recent interrupted session, |
| 2850 | even if that session originated in a completely different workspace. Tools |
| 2851 | then operated on the prior workspace's file paths while the status bar |
| 2852 | displayed the *current* workspace name — a confusing trust-boundary |
| 2853 | violation that could leak `api_messages`, `working_set` entries, and any |
| 2854 | secrets the prior session had accumulated into a new terminal that was |
| 2855 | never meant to see them. `try_recover_checkpoint()` now compares the saved |
| 2856 | session's workspace to `std::env::current_dir()` (canonicalised, with a |
| 2857 | strict-equality fallback when canonicalisation fails) and only auto-recovers |
| 2858 | on a match. On a mismatch the checkpoint is persisted as a regular session |
| 2859 | (so the user can find it via `deepseek sessions` / `deepseek resume <id>`) |
| 2860 | and cleared, and the new launch starts fresh — no data is lost. Hotfixed |
| 2861 | to `main` ahead of the v0.8.12 tag. |
| 2862 | - **`cargo install` on stable Rust** — the language-picker match guard at |
| 2863 | `crates/tui/src/tui/ui.rs:1603` used `&& let Some(...) = ...` inside an |
| 2864 | `if`-guard, which requires the nightly-only `if_let_guard` feature on Rust |
| 2865 | before 1.94. Reported by an external user whose `cargo install |
| 2866 | deepseek-tui` failed with E0658. Rewrote as a plain match guard with a |
| 2867 | nested `if let` inside the arm body. The workspace also now declares |
| 2868 | `rust-version = "1.88"` (the actual minimum for `let_chains` in |
| 2869 | `if`/`while`) so users on too-old toolchains see a clear cargo error |
| 2870 | instead of a confusing rustc one. AGENTS.md gains a "stable Rust only" |
| 2871 | section so this doesn't regress. |
| 2872 | - **Resident-file lease never released after spawn** (#660) — the lease was |
| 2873 | stamped as `"pending"` at spawn time because the agent id is only assigned |
| 2874 | by the manager after the spawn call returns. The release-on-terminal-state |
| 2875 | path (added in the original #660 commit) matched leases by agent id, so |
| 2876 | it could never find these placeholder entries. Now the placeholder is |
| 2877 | replaced with the real agent id immediately after spawn so existing |
| 2878 | release wiring fires. Resolves the v0.8.12 caveat documented at RC time. |
| 2879 | - **Color::Reset across all UI widgets** (#651, #671) — replaced hardcoded |
| 2880 | `Color::Black` and `Color::Rgb(18, 29, 39)` backgrounds with `Color::Reset` |
| 2881 | so the TUI respects the terminal's actual background color on light-themed |
| 2882 | and non-standard terminals. |
| 2883 | - **Windows MessageBeep** (#646) — `notify_done_to` now calls `MessageBeep` on |
| 2884 | Windows when BEL method is selected. |
| 2885 | - **truncate_id optimization** (#649) — replaced manual string slicing with a |
| 2886 | shared `truncate_id` helper across session, picker, and UI call sites. |
| 2887 | |
| 2888 | ### Maintenance |
| 2889 | - Workspace `cargo fmt` sweep across community PRs that landed unformatted. |
| 2890 | - Issue-triage GitHub Actions added (#688): keyword-driven auto-labeller, |
| 2891 | stale-bot for `needs-info` issues (14 d → stale → 7 d → close), and a |
| 2892 | spam lockdown that auto-closes promotional issues from accounts <30 d |
| 2893 | old. All pure GitHub Actions — no third-party services. |
| 2894 | - Annotated `TuiPrefs` (#657) and `handoff::THRESHOLDS` (#667) with |
| 2895 | `#[allow(dead_code)]` so the deferred APIs don't trip CI's `-D warnings` |
| 2896 | flag while their call sites are staged for v0.8.13. |
| 2897 | - Removed dead `prefer_handoff` field from `CompactionConfig` — config knob |
| 2898 | existed but zero code paths consulted it (#667). |
| 2899 | - Removed dead `use_terminal_colors` field from `TuiConfig` — no rendering |
| 2900 | code read the value (#671). |
| 2901 | - Fixed `expect()` panic risk in `OpenSandboxBackend::new()` — now returns |
| 2902 | `Result` (#645). |
| 2903 | - Fixed broken `section_bg` test assertion after Color::Reset migration (#651). |
| 2904 | - Fixed `resolve_prefixes` docstring to accurately describe deny-always-wins |
| 2905 | behavior (#653). |
| 2906 | - Wired `create_backend()` into `Engine::build_tool_context` — sandbox backend |
| 2907 | was defined but never activated (#645). |
| 2908 | - Wired resident lease release on agent completion/cancellation/failure (#660). |
| 2909 | |
| 2910 | ### Contributors |
| 2911 | |
| 2912 | First-time contributor to this release: **@zichen0116** (#686). Welcome — and |
| 2913 | thank you. |
| 2914 | |
| 2915 | Bulk community contributions by [@merchloubna70-dot](https://github.com/merchloubna70-dot) |
| 2916 | (#645–#681, 28 PRs spanning features, fixes, and VS Code extension scaffolding). |
| 2917 | *Thank you for the remarkable volume and quality of work.* |
| 2918 | |
| 2919 | ## [0.8.10] - 2026-05-04 |
| 2920 | |
| 2921 | A patch release: hotfixes, small UX polish, and four whalescale-unblocking |
| 2922 | runtime API additions. No breaking changes. |
| 2923 | |
| 2924 | ### Added |
| 2925 | - **OPENCODE shell.env hook** (#456) — lifecycle hooks can now inject |
| 2926 | shell environment into spawned commands without hard-coding env in |
| 2927 | prompts or wrapper scripts. |
| 2928 | - **Stacked toast overlay** (#439) — status toasts can queue and render |
| 2929 | together instead of overwriting each other. |
| 2930 | - **File @-mention frecency** (#441) — file mention suggestions learn |
| 2931 | from recent selections via `~/.deepseek/file-frecency.jsonl`. |
| 2932 | - **Durable keybinding catalog** (#559) — `docs/KEYBINDINGS.md` is now |
| 2933 | the source-of-truth audit for current shortcuts and the future |
| 2934 | configurable-keymap registry. |
| 2935 | - **Runtime API quartet for whalescale-desktop integration** (#561, #562, #563, |
| 2936 | #564, #567) — addresses whalescale#255/256/260/261: |
| 2937 | - `[runtime_api] cors_origins` config / `--cors-origin URL` flag (repeatable) / |
| 2938 | `DEEPSEEK_CORS_ORIGINS` env var, all stacking on top of the built-in |
| 2939 | dev-origin defaults (#561 / whalescale#255). |
| 2940 | - `PATCH /v1/threads/{id}` extended from `archived`-only to the full |
| 2941 | editable field set: `allow_shell`, `trust_mode`, `auto_approve`, `model`, |
| 2942 | `mode`, `title`, `system_prompt`. Empty string clears `title` / |
| 2943 | `system_prompt`. New `title` field on `ThreadRecord` is additive — no |
| 2944 | schema_version bump (#562 / whalescale#256). |
| 2945 | - `archived_only=true` query param on `GET /v1/threads` and |
| 2946 | `/v1/threads/summary`, backed by a new `ThreadListFilter` enum |
| 2947 | (#563 / whalescale#260). |
| 2948 | - `GET /v1/usage?since=&until=&group_by=<day|model|provider|thread>` |
| 2949 | aggregates token totals + cost (via `pricing.rs`) across all |
| 2950 | threads/turns. Empty time ranges yield empty `buckets` (never 404) |
| 2951 | (#564 / whalescale#261). |
| 2952 | - **Language picker in first-run onboarding** (#566) — new step between |
| 2953 | Welcome and ApiKey lists every shipped locale (`auto` / `en` / `ja` / |
| 2954 | `zh-Hans` / `pt-BR`) with the native name (日本語, 简体中文, …) plus an |
| 2955 | English label so the target language is reachable without already |
| 2956 | speaking it. Hotkeys 1-5 select; persists immediately to |
| 2957 | `~/.deepseek/settings.toml`. |
| 2958 | - **Windows + China install documentation** (#578) — expanded |
| 2959 | `docs/INSTALL.md` with Windows source-build setup, Visual Studio Build |
| 2960 | Tools / MSVC environment notes, rustup and Cargo mirror guidance, and |
| 2961 | antivirus troubleshooting. *Thanks to |
| 2962 | [@loongmiaow-pixel](https://github.com/loongmiaow-pixel) for this PR.* |
| 2963 | |
| 2964 | ### Changed |
| 2965 | - **Agent prompt now explicitly describes DeepSeek cache-aware behavior** |
| 2966 | — long-session guidance explains why stable prompt prefixes, sub-agents, |
| 2967 | RLM, and late compaction matter for V4 cache economics. |
| 2968 | - **Whale sub-agent nicknames now interleave Simplified Chinese with |
| 2969 | English** (`Blue` / `蓝鲸` / `Humpback` / `座头鲸` / …). Pure cosmetic; |
| 2970 | doubles the labeling pool size and gives a roughly even mix on each |
| 2971 | new spawn. |
| 2972 | - **User memory docs + help polish** (#497, #569) — `/memory` is now |
| 2973 | listed in slash-command help, supports `/memory help`, and the README |
| 2974 | / configuration docs now point at the full `docs/MEMORY.md` guide and |
| 2975 | document both `[memory].enabled` and `DEEPSEEK_MEMORY`. *Thanks to |
| 2976 | [@20bytes](https://github.com/20bytes) for this PR.* |
| 2977 | |
| 2978 | ### Fixed |
| 2979 | - **Compaction summaries are cache-aligned for DeepSeek V4** (#575, #580) |
| 2980 | — when the summarized message prefix fits the large V4 context budget, |
| 2981 | the summary request now reuses the original messages and appends the |
| 2982 | summary instruction as a normal user message instead of rebuilding a |
| 2983 | fresh `SUMMARY_PROMPT + dropped messages` input. This lets the summary |
| 2984 | call benefit from DeepSeek prefix caching. *Thanks to |
| 2985 | [@lloydzhou](https://github.com/lloydzhou) and |
| 2986 | [@jeoor](https://github.com/jeoor) for the cost reports and concrete |
| 2987 | strategy.* |
| 2988 | - **Windows Terminal API-key paste during onboarding** (#577) — the |
| 2989 | setup wizard now handles Ctrl/Cmd+V before generic character input and |
| 2990 | filters control/meta-modified keys out of the API-key text path. |
| 2991 | *Thanks to [@toi500](https://github.com/toi500) for the report and |
| 2992 | workaround details.* |
| 2993 | - **Terminal startup repaint** (#581) — the TUI clears the terminal |
| 2994 | immediately after initialization so normal-screen startup no longer |
| 2995 | leaves stale default-background rows above the first frame. *Thanks to |
| 2996 | [@xsstomy](https://github.com/xsstomy) for the screenshot.* |
| 2997 | - **Markdown rendering for tables, bold/italic, and horizontal rules** |
| 2998 | (#579) — transcript markdown now handles table rows, strips separator |
| 2999 | rows, renders horizontal rules, applies inline bold/italic styles, and |
| 3000 | avoids an infinite-loop edge case on unclosed markers. *Thanks to |
| 3001 | [@WyxBUPT-22](https://github.com/WyxBUPT-22) for the PR, screenshots, |
| 3002 | and tests.* |
| 3003 | - **Slash-prefix Enter activation** (#573) — typing a short prefix such |
| 3004 | as `/mo` and pressing Enter now activates the first slash-command |
| 3005 | match. *Thanks to [@melody0709](https://github.com/melody0709) for |
| 3006 | the report.* |
| 3007 | - **macOS seatbelt blocked `~/.cargo/registry`** (#558) — `cargo publish` |
| 3008 | / `cargo build` from inside the TUI's shell tool was getting |
| 3009 | sandbox-denied. The seatbelt now allows read on `(param "CARGO_HOME")` |
| 3010 | and write on the `registry/` and `git/` subpaths whenever the policy |
| 3011 | isn't read-only. Honors `CARGO_HOME` env with a `$HOME/.cargo` |
| 3012 | fallback. |
| 3013 | - **Stdio MCP servers now receive SIGTERM on shutdown** (#420) — instead |
| 3014 | of SIGKILL via `kill_on_drop`. New `async fn shutdown` on |
| 3015 | `McpTransport` overrides on `StdioTransport` to send SIGTERM and wait |
| 3016 | up to 2s for graceful exit before drop fires SIGKILL as the backstop. |
| 3017 | Wired into the engine's `Op::Shutdown` path so graceful exit is the |
| 3018 | default. A Drop fallback still SIGTERMs on abnormal exit paths. |
| 3019 | - **Shell-spawned children get `PR_SET_PDEATHSIG(SIGTERM)` on Linux** |
| 3020 | (#421) — the kernel sends SIGTERM the moment the parent (TUI) exits, |
| 3021 | even on SIGKILL of the parent. Closes the leak window the cooperative |
| 3022 | cancellation path can't cover. macOS / Windows watchdog tracked as a |
| 3023 | follow-up; the existing `kill_on_drop` + process_group SIGKILL on |
| 3024 | cancellation still cover normal shutdown there. |
| 3025 | - **npm install on older glibc now fails fast** (#555, #560, #556, #565) |
| 3026 | — the prebuilt Linux x64 / arm64 binaries are now built via |
| 3027 | `cargo zigbuild` targeting `x86_64-unknown-linux-gnu.2.28` / |
| 3028 | `aarch64-unknown-linux-gnu.2.28`, lowering the requirement from glibc |
| 3029 | ≥ 2.39 to ≥ 2.28. The npm postinstall also runs a Linux-only glibc |
| 3030 | preflight that fails fast with a clear "build from source" message |
| 3031 | when the host is incompatible (or musl). *Thanks to |
| 3032 | [@staryxchen](https://github.com/staryxchen) (#556) and |
| 3033 | [@Vishnu1837](https://github.com/Vishnu1837) (#565) for these PRs.* |
| 3034 | - **Shell tool `cwd` parameter now validated against the workspace |
| 3035 | boundary** (#524) — the model could previously pass `cwd` paths |
| 3036 | outside the workspace; now `exec_shell` runs `ToolContext::resolve_path` |
| 3037 | on `cwd` like every other path-taking file tool, returning |
| 3038 | `PathEscape` on violations. `trust_mode = true` still bypasses, |
| 3039 | consistent with the file-tool pattern. *Thanks to |
| 3040 | [@shentoumengxin](https://github.com/shentoumengxin) for this PR.* |
| 3041 | |
| 3042 | ### Contributors |
| 3043 | |
| 3044 | First-time contributors to this release: **@staryxchen** (#556), |
| 3045 | **@shentoumengxin** (#524), **@Vishnu1837** (#565), **@20bytes** |
| 3046 | (#569), **@loongmiaow-pixel** (#578), and **@WyxBUPT-22** (#579). |
| 3047 | Welcome — and thank you. |
| 3048 | |
| 3049 | ## [0.8.8] - 2026-05-03 |
| 3050 | |
| 3051 | ### Added |
| 3052 | - **User memory MVP** (#489–#493) — opt-in persistent note file |
| 3053 | injected into the system prompt as a `<user_memory>` block. |
| 3054 | - `# foo` typed in the composer appends a timestamped bullet |
| 3055 | without firing a turn (#492). |
| 3056 | - `/memory [show|path|clear|edit]` slash command for inline |
| 3057 | inspection / editing hints (#491). |
| 3058 | - `remember` model-callable tool so the agent can capture |
| 3059 | durable preferences itself; auto-approved because writes are |
| 3060 | scoped to the user's own file (#489). |
| 3061 | - Hierarchy loader pulls `~/.deepseek/memory.md` (path |
| 3062 | configurable via `memory_path` / `DEEPSEEK_MEMORY_PATH`) and |
| 3063 | injects above the volatile-content boundary in the prompt |
| 3064 | (#490). |
| 3065 | - Default off; enable with `[memory] enabled = true` or |
| 3066 | `DEEPSEEK_MEMORY=on` (#493). |
| 3067 | - Full feature documentation in `docs/MEMORY.md`. |
| 3068 | - **Inline diff rendering for `edit_file` / `write_file`** (#505) — |
| 3069 | tool results now emit a unified diff at the head of the body, |
| 3070 | picked up by the existing diff-aware renderer with line numbers |
| 3071 | and coloured `+`/`-` gutters. New `similar` crate dep. |
| 3072 | - **OSC 8 hyperlinks** (#498) — URLs in the transcript become |
| 3073 | Cmd+click-openable in supporting terminals (iTerm2, Terminal.app |
| 3074 | 13+, Ghostty, Kitty, WezTerm, Alacritty). Clipboard path strips |
| 3075 | the escapes so yanked text stays clean. Off-switch: |
| 3076 | `[tui] osc8_links = false`. |
| 3077 | - **Retry/backoff visual countdown** (#499) — `⟳ retry N in Ms — reason` |
| 3078 | banner ticks down during HTTP backoff. On exhaustion the row turns |
| 3079 | red `× failed: <reason>` until the next turn starts. |
| 3080 | - **MCP server health chip** (#502) — colour-coded `MCP M/N` in the |
| 3081 | footer's right-cluster: success / warning / error / muted by |
| 3082 | reachability. Hidden when zero MCP servers are configured. |
| 3083 | - **Per-project config overlay** (#485) — `<workspace>/.deepseek/config.toml` |
| 3084 | overlays a curated set of fields on top of the user-global config: |
| 3085 | `model`, `reasoning_effort`, `approval_policy`, `sandbox_mode`, |
| 3086 | `notes_path`, `max_subagents`, `allow_shell`, plus the |
| 3087 | `instructions = [...]` array (#454). Pass `--no-project-config` |
| 3088 | to bypass for one launch. |
| 3089 | - **Project-scope deny-list for credentials/redirects** (#417) — |
| 3090 | `api_key`, `base_url`, `provider`, and `mcp_config_path` are |
| 3091 | refused at project scope. A malicious |
| 3092 | `<workspace>/.deepseek/config.toml` would otherwise be able to |
| 3093 | exfiltrate prompts to an attacker-controlled endpoint by |
| 3094 | swapping the user's credentials and target host with |
| 3095 | project-controlled values, or redirect the MCP loader at a |
| 3096 | config that spawns arbitrary stdio servers under the user's |
| 3097 | identity. The denied key emits a stderr warning so a user who |
| 3098 | expected the override sees the deny instead of a silent drop. |
| 3099 | - **Project-scope value-deny for the loosest postures** (#417 |
| 3100 | follow-up) — `approval_policy = "auto"` and |
| 3101 | `sandbox_mode = "danger-full-access"` are pure escalation |
| 3102 | values, denied unconditionally at project scope regardless |
| 3103 | of the user's prior value. Sub-tightening comparisons |
| 3104 | (e.g. user `"never"` → project `"on-request"` is allowed |
| 3105 | even though it loosens) stay v0.8.9 follow-up because they |
| 3106 | need a richer ordering check. |
| 3107 | - **`SSL_CERT_FILE` honored in the HTTPS client** (#418) — corporate |
| 3108 | proxy / TLS-inspecting MITM users can now point at their custom |
| 3109 | CA bundle and have it added alongside the platform's system |
| 3110 | trust store. Tries PEM-bundle parsing first (covers single-cert |
| 3111 | files too), falls back to DER. Failures log a warning and |
| 3112 | continue — the existing system roots still apply, so a |
| 3113 | malformed env var won't bring down the launch. Documented in |
| 3114 | `docs/CONFIGURATION.md`. |
| 3115 | - **Execpolicy heredoc handling** (#419) — `normalize_command` now |
| 3116 | strips heredoc bodies before shlex tokenization so a user's |
| 3117 | `auto_allow = ["cat > file.txt"]` pattern matches the heredoc |
| 3118 | form `cat <<EOF > file.txt\nbody\nEOF` cleanly. Recognises the |
| 3119 | common forms (`<<DELIM`, `<<-DELIM`, `<<'DELIM'`, `<<"DELIM"`) |
| 3120 | while leaving the here-string operator (`<<<`) untouched. |
| 3121 | Without this fix, heredoc-form file writes would skip the |
| 3122 | user's auto-approve list and route through the approval modal |
| 3123 | even for explicitly-blessed commands. |
| 3124 | - **Sub-agent role taxonomy expansion** (#404) — adds `Implementer` |
| 3125 | ("land this change with the minimum surrounding edit") and |
| 3126 | `Verifier` ("run the test suite, report pass/fail with evidence") |
| 3127 | to the existing `general` / `explore` / `plan` / `review` / |
| 3128 | `custom` set. Each role has a distinct system prompt posture. |
| 3129 | Documented in `docs/SUBAGENTS.md`. |
| 3130 | - **`docs/SUBAGENTS.md`** — full sub-agent reference: role taxonomy, |
| 3131 | alias map, concurrency cap, lifecycle, session-boundary |
| 3132 | classification, output contract. |
| 3133 | - **`docs/MEMORY.md`** — user-facing memory feature documentation. |
| 3134 | - **Competitive analysis doc** — `docs/COMPETITIVE_ANALYSIS.md` |
| 3135 | catalogues capability matrix vs OpenCode and Codex CLI. |
| 3136 | - **Session prune helper + `/sessions prune <days>`** (#406 phase-1) — |
| 3137 | drops persisted sessions older than N days from |
| 3138 | `~/.deepseek/sessions/`. Skips the checkpoint subdirectory and |
| 3139 | compares against metadata `updated_at` (not fs mtime, which can |
| 3140 | lie after an rsync). 10 total tests cover the helper's contract |
| 3141 | and the slash-command dispatch surface. Phase 2 (boot-prune + |
| 3142 | retention policy) stays v0.8.9 work. |
| 3143 | - **`deepseek doctor --json`** now surfaces a `memory` block |
| 3144 | (`enabled` / `path` / `file_present`) so operators can verify |
| 3145 | memory configuration without booting the TUI. |
| 3146 | - **Tool-output spillover** (#422 + #423 + #500) — tool outputs over |
| 3147 | 100 KiB now spill to `~/.deepseek/tool_outputs/<id>.txt` from the |
| 3148 | engine's tool-execution path. The model receives a 32 KiB head plus |
| 3149 | a footer pointing at the spillover file (`Use read_file path=…`), |
| 3150 | the tool cell renders an inline `full output: <path>` annotation in |
| 3151 | live mode, and a 7-day boot prune keeps the directory bounded. |
| 3152 | Spillover is skipped on error results so the model still sees the |
| 3153 | failure message verbatim. The existing tool-details pager surfaces |
| 3154 | the truncated head so the user can verify what the model saw. |
| 3155 | |
| 3156 | ### Changed |
| 3157 | - **Sub-agent concurrency cap raised to 10 by default** (#509) — |
| 3158 | was 5; configurable via `[subagents].max_concurrent` (hard |
| 3159 | ceiling 20). Running-count now ignores non-running, no-handle, |
| 3160 | and finished handles so completed agents stop occupying slots. |
| 3161 | - **`SharedSubAgentManager` is `Arc<RwLock<...>>`** (#510) — read |
| 3162 | paths take read locks, eliminating the multi-agent fan-out UI |
| 3163 | freeze. |
| 3164 | - **Sub-agent output summarized before parent context** (#511) — |
| 3165 | `compact_tool_result_for_context` now compresses |
| 3166 | `agent_result` / `agent_wait` payloads instead of dumping the |
| 3167 | full snapshot back into the parent's context window. |
| 3168 | - **`agent_list` defaults to current-session view** (#405) — each |
| 3169 | manager mints a `session_boot_id` and stamps every spawn; agents |
| 3170 | loaded from prior sessions are filtered unless |
| 3171 | `include_archived=true` is passed. Each result carries a |
| 3172 | `from_prior_session` flag. |
| 3173 | - **Concise todo / checklist update rendering** (#403) — repeat |
| 3174 | `todo_update` / `checklist_update` calls render a one-line |
| 3175 | `Todo #N: <title> → STATUS` card with full list still |
| 3176 | reachable via Alt+V instead of dumping the entire item array on |
| 3177 | every call. |
| 3178 | - **Compact `agent_spawn` rendering** (#409) — the generic tool |
| 3179 | block for `agent_spawn` collapses to one header line in live |
| 3180 | mode (`◐ delegate · agent-abc12 [running]`) since the |
| 3181 | `DelegateCard` already owns live action progress. Transcript |
| 3182 | replay keeps the full block. |
| 3183 | - **Plan panel role clarified** (#408) — drops the "No active |
| 3184 | plan" placeholder when the panel is otherwise empty; documents |
| 3185 | the panel's narrow role (`update_plan` tool output + `/goal` + |
| 3186 | cycle counter, distinct from todos). |
| 3187 | - **Sub-agent description copy** — `agent_spawn` tool description |
| 3188 | and `prompts/base.md` updated to reflect the new default cap of |
| 3189 | 10 (was stale "Max 5 in flight"). |
| 3190 | - **`agent_spawn` / `agent_assign` schema descriptions** (#404 |
| 3191 | follow-up) — type/agent_name property descriptions now list |
| 3192 | `implementer` and `verifier` so the model surfaces those roles |
| 3193 | without having to discover them from `docs/SUBAGENTS.md`. Adds |
| 3194 | the long-form aliases (`builder` / `validator` / `tester`) on |
| 3195 | `agent_assign` for parity with the alias map. |
| 3196 | - **Multi-day duration formatting** (#447) — `humanize_duration` |
| 3197 | now caps at two units and promotes through h/d/w boundaries. |
| 3198 | Long-running sessions render as `2d 3h` instead of `188415s`, |
| 3199 | and the previous "192m 30s" cycle output becomes `3h 12m`. The |
| 3200 | `/goal` status line picks up the same formatter so multi-day |
| 3201 | goal-elapsed times stay readable. |
| 3202 | - **Accessibility flag** (#450) — `NO_ANIMATIONS=1` env var now |
| 3203 | forces `low_motion = true` and `fancy_animations = false` at |
| 3204 | startup, regardless of the saved `settings.toml`. Recognises |
| 3205 | the standard truthy spellings (`1`, `true`, `yes`, `on`). |
| 3206 | Documented end-to-end in the new `docs/ACCESSIBILITY.md`, |
| 3207 | including the existing `low_motion` / `calm_mode` / |
| 3208 | `show_thinking` / `show_tool_details` toggles for |
| 3209 | screen-reader users. |
| 3210 | - **Cumulative session-elapsed footer chip** (#448) — a |
| 3211 | low-priority `worked 3h 12m` chip in the footer's right |
| 3212 | cluster shows session age once it crosses 60s. Hidden during |
| 3213 | the first minute of a launch so a fresh start doesn't flash a |
| 3214 | ticker. Drops first under narrow widths so the existing chips |
| 3215 | (coherence / agents / replay / cache / mcp) keep their slots. |
| 3216 | Sampled at props-build time (matches the `retry` capture |
| 3217 | pattern) so render stays pure for tests. |
| 3218 | - **`instructions = [...]` config array** (#454) — declare |
| 3219 | additional instruction files (`./AGENTS.md`, |
| 3220 | `~/.deepseek/global.md`, …) and they're concatenated into the |
| 3221 | system prompt in declared order, above the skills block. Each |
| 3222 | file is capped at 100 KiB; missing files log a warning and are |
| 3223 | skipped instead of failing the launch. Project config replaces |
| 3224 | the user-level array wholesale (the typical "merge" pattern is |
| 3225 | for users who want both — they list `~/global.md` inside the |
| 3226 | project array). Documented in `config.example.toml`. |
| 3227 | - **Keyboard-enhancement flags pop on suspend paths too** (#443 |
| 3228 | follow-up) — `pause_terminal` (Ctrl+Z / shell-suspend) and |
| 3229 | `external_editor::spawn_editor_for_input` (composer `$EDITOR` |
| 3230 | launch) now pop the flags before handing the terminal to the |
| 3231 | child process, matching the existing shutdown and panic-hook |
| 3232 | paths. Defense-in-depth: if a future code path enables the |
| 3233 | flags explicitly, the suspend handlers won't leak them to a |
| 3234 | Vim / less / shell child that hasn't asked for them. |
| 3235 | - **`load_skill` tool** (#434) — model-callable tool that takes a |
| 3236 | skill id and returns the SKILL.md body plus the sibling |
| 3237 | companion-file list in one call. Faster than the existing |
| 3238 | `read_file` + `list_dir` dance; surfaces the skill's |
| 3239 | description as a quote block at the head so a single tool |
| 3240 | result is self-contained. Resolves the skills directory with |
| 3241 | the same hierarchy `App::new` uses (`.agents/skills` → |
| 3242 | `skills` → `~/.deepseek/skills`). Available in Plan and |
| 3243 | Agent/Yolo modes. |
| 3244 | - **Kitty keyboard protocol opt-in** (#442) — pushes |
| 3245 | `DISAMBIGUATE_ESCAPE_CODES` at startup so terminals that |
| 3246 | support the protocol (Kitty, Ghostty, Alacritty 0.13+, |
| 3247 | WezTerm, recent Konsole / xterm) report unambiguous events |
| 3248 | for Option/Alt-modified keys, plain Esc, and multi-byte |
| 3249 | sequences. Legacy terminals silently discard the escape and |
| 3250 | see no change. Only the disambiguation tier is pushed — |
| 3251 | release-event reporting was deliberately skipped because the |
| 3252 | existing handlers would mis-route releases as duplicate |
| 3253 | presses. The flags are popped on shutdown / panic / suspend |
| 3254 | paths (#443). |
| 3255 | - **Multi-directory skill discovery** (#432) — the system |
| 3256 | prompt's `## Skills` listing and the `load_skill` tool now |
| 3257 | walk every candidate directory in the workspace plus the |
| 3258 | global default: `<workspace>/.agents/skills` → |
| 3259 | `<workspace>/skills` → `<workspace>/.opencode/skills` → |
| 3260 | `<workspace>/.claude/skills` → `~/.deepseek/skills`. Skills |
| 3261 | installed for any AI-tool convention show up in the same |
| 3262 | catalogue. Name conflicts resolve first-match-wins per the |
| 3263 | precedence order so workspace-local skills shadow user/global |
| 3264 | ones. New `skills_directories()` and |
| 3265 | `discover_in_workspace()` helpers in |
| 3266 | `crates/tui/src/skills/mod.rs`. |
| 3267 | - **`tool.spillover` audit event** (#500 polish) — emit a |
| 3268 | discrete audit-log entry whenever `apply_spillover` writes a |
| 3269 | spillover file, so operators tailing |
| 3270 | `~/.deepseek/audit.log` can correlate large-output episodes |
| 3271 | with disk-usage growth in `~/.deepseek/tool_outputs/`. Fires |
| 3272 | in both the sequential and parallel tool paths. |
| 3273 | - **Prompt stash** (#440) — Ctrl+S in the composer parks the |
| 3274 | current draft to a JSONL-backed stash at |
| 3275 | `~/.deepseek/composer_stash.jsonl` (no-op on empty composer). |
| 3276 | `/stash list` shows parked drafts (oldest first, with one-line |
| 3277 | previews and timestamps); `/stash pop` restores the most |
| 3278 | recently parked draft into the composer (LIFO). Self-healing |
| 3279 | parser drops malformed lines instead of poisoning the stash. |
| 3280 | Capped at 200 entries; multiline drafts round-trip intact via |
| 3281 | JSON's newline escaping. |
| 3282 | - **`deepseek pr <N>` subcommand** (#451) — fetches PR |
| 3283 | title/body/diff via `gh` and launches the interactive TUI |
| 3284 | with a review prompt pre-populated in the composer. The |
| 3285 | diff is capped at 200 KiB (codepoint-safe truncation) so a |
| 3286 | massive PR doesn't blow the context window before the user |
| 3287 | hits Enter. Optional `--repo <owner/name>` and `--checkout` |
| 3288 | flags; falls back gracefully with an actionable error |
| 3289 | message if `gh` isn't on PATH. Adds a new |
| 3290 | `TuiOptions::initial_input` plumb that any future caller can |
| 3291 | reuse to drop the model into a session with text already |
| 3292 | typed. |
| 3293 | - **`/stash clear` subcommand** (#440 polish) — wipes the |
| 3294 | entire stash file and reports how many parked drafts were |
| 3295 | dropped. Pairs with `/stash list` and `/stash pop` so the |
| 3296 | user can fully manage the stash from inside the TUI without |
| 3297 | reaching for `rm`. |
| 3298 | - **`/hooks` read-only listing** (#460 MVP) — slash command |
| 3299 | enumerates configured lifecycle hooks grouped by event, |
| 3300 | showing each hook's name, command preview, timeout, and |
| 3301 | condition. Notes the global `[hooks].enabled` flag's state. |
| 3302 | No more `cat ~/.deepseek/config.toml` to debug "did my hook |
| 3303 | actually load". The picker / persisted enable-disable |
| 3304 | surface from #460 stays as v0.8.9 follow-up. Available via |
| 3305 | `/hooks` or `/hooks list`; aliased to `/hook`. Localized in |
| 3306 | en/ja/zh-Hans/pt-BR. |
| 3307 | - **`deepseek doctor` reports cross-tool skill dirs** (#432 |
| 3308 | follow-up) — both the human-readable and JSON outputs now |
| 3309 | surface `.opencode/skills/` and `.claude/skills/` presence / |
| 3310 | count, so operators can confirm at a glance whether any |
| 3311 | cross-tool skill folder is contributing to the merged |
| 3312 | catalogue. Empty dirs are omitted from the human-readable |
| 3313 | output to keep the report scannable; JSON always emits all |
| 3314 | five slots (`global`, `agents`, `local`, `opencode`, |
| 3315 | `claude`) for stable machine consumption. |
| 3316 | - **`deepseek doctor` reports storage surfaces** (#422 / #440 / |
| 3317 | #500 follow-up) — new `Storage:` section surfaces the |
| 3318 | tool-output spillover dir |
| 3319 | (`~/.deepseek/tool_outputs/`) with file count and the |
| 3320 | composer stash file |
| 3321 | (`~/.deepseek/composer_stash.jsonl`) with parked-draft |
| 3322 | count. Mirrored under `storage.{spillover,stash}` in the |
| 3323 | JSON output so `deepseek doctor --json` keeps a stable |
| 3324 | schema. |
| 3325 | - **`/hooks events` subcommand** (#460 polish) — lists every |
| 3326 | supported `HookEvent` value with a short blurb so users can |
| 3327 | discover which events to target in `[[hooks.hooks]]` entries |
| 3328 | without reading source. Ordered lifecycle → per-tool → |
| 3329 | situational, stable across releases. |
| 3330 | - **Structured-Markdown compaction template** (#429) — |
| 3331 | `prompts/compact.md` switches from the legacy |
| 3332 | Active-task/Files-touched/Key-decisions/Open-blockers |
| 3333 | framing to the spec'd structure: Goal / Constraints / |
| 3334 | Progress (Done / In Progress / Blocked) / Key Decisions / |
| 3335 | Next step. The richer Progress sub-bullets help long |
| 3336 | resumed sessions distinguish "what's verified done" from |
| 3337 | "what's mid-flight" — useful when the model writes |
| 3338 | `.deepseek/handoff.md` before a long break. Backwards- |
| 3339 | compat: existing handoff.md files continue to render fine |
| 3340 | because the loader injects them as plain markdown (the |
| 3341 | template only guides what NEW handoffs look like). The |
| 3342 | pinned-tool-output configurability part of #429's spec |
| 3343 | stays a v0.8.9 follow-up — that requires changes to |
| 3344 | `cycle_manager.rs` compaction logic itself. |
| 3345 | - **`tool_call_before` / `tool_call_after` / `message_submit` / |
| 3346 | `on_error` hooks all fire now** (#455 observer-only slice) — |
| 3347 | these events were defined in the `HookEvent` enum but never |
| 3348 | fired from production code. Wired through: |
| 3349 | `tool_call_before` and `tool_call_after` fire from |
| 3350 | `tool_routing.rs`; `message_submit` fires from |
| 3351 | `dispatch_user_message` before engine dispatch; `on_error` |
| 3352 | fires from `apply_engine_error_to_app` before the error cell |
| 3353 | reaches the transcript. Hook contexts populate the relevant |
| 3354 | fields (`tool_name` + `tool_args` / `tool_result`, |
| 3355 | `message`, `error`). Hooks remain read-only in this slice; |
| 3356 | argument / result / message mutation is a v0.8.9 follow-up |
| 3357 | because it needs a synchronous-gate contract that doesn't |
| 3358 | exist today. Combined with the existing `session_start` / |
| 3359 | `session_end` / `mode_change` events, every variant in the |
| 3360 | `HookEvent` enum now has a live producer. Each fire is |
| 3361 | fast-path-gated by |
| 3362 | `HookExecutor::has_hooks_for_event(event)` so per-tool |
| 3363 | dispatch never pays for `HookContext` allocation when the |
| 3364 | user has no hooks configured (the common case). |
| 3365 | - **RLM tool family** (#512) — `rlm` tool cards map to |
| 3366 | `ToolFamily::Rlm` and render `rlm`, not `swarm`. Stale "swarm" |
| 3367 | wording cleaned out of docs / comments / tests. |
| 3368 | - **Foreground RLM visible in Agents sidebar** (#513 — stopgap) |
| 3369 | — projection now shows foreground RLM work; full async |
| 3370 | lifecycle remains v0.8.9. |
| 3371 | |
| 3372 | ### Fixed |
| 3373 | - **`Don't auto-approve git -C ...`** (#416, shipped 2026-05-03) — |
| 3374 | v0.8.8 release runtime fix; foundation for the rest of the |
| 3375 | stabilization batch. |
| 3376 | - **Self-update arch mapping** (#503) — `update.rs` uses release |
| 3377 | asset naming (`arm64`/`x64`) instead of raw Rust constants |
| 3378 | (`aarch64`/`x86_64`); rejects `.sha256` siblings as primary |
| 3379 | binaries. |
| 3380 | - **Composer Option+Backspace deletes by word** (#488) — was |
| 3381 | deleting by character. |
| 3382 | - **Offline composer queue is session-scoped** (#487) — legacy |
| 3383 | unscoped queues fail closed instead of leaking content into |
| 3384 | unrelated chats. |
| 3385 | - **`display_path` test race + Windows separator** (#506) — |
| 3386 | tests no longer mutate `$HOME`; `display_path_with_home` walks |
| 3387 | components and joins with `MAIN_SEPARATOR_STR` so Windows shows |
| 3388 | `~\projects\foo` not `~\projects/foo`. |
| 3389 | - **Footer reads statusline colours from `app.ui_theme`** (#449) — |
| 3390 | was using a bespoke palette. |
| 3391 | - **Keyboard-enhancement flags pop on panic exit too** (#443/#444) — |
| 3392 | raw-mode startup probe is now bounded by a configurable |
| 3393 | timeout. |
| 3394 | - **CI workflow cleanup** (#507) — pruned three duplicated/dead |
| 3395 | workflows (`crates-publish.yml`, `parity.yml`, `publish-npm.yml`); |
| 3396 | `release.yml` `build` job now allows `parity` to be skipped on |
| 3397 | manual `workflow_dispatch`; release-runbook reconciled. |
| 3398 | - **Slash-menu layout jitter on Windows** — typing through a |
| 3399 | `/foo` autocomplete used to shrink the matched-entry count, |
| 3400 | which shrank the composer height every keystroke, which forced |
| 3401 | the chat area above to repaint. On Windows 10 PowerShell + WSL |
| 3402 | the per-cell write cost made the jitter visible. Composer now |
| 3403 | reserves its panel-max envelope for the whole slash/mention |
| 3404 | session so the chat-area Rect stays stable; the menu still |
| 3405 | renders only the entries that actually match. |
| 3406 | |
| 3407 | - **Linux ARM64 prebuilt binaries** — the release workflow now publishes |
| 3408 | `deepseek-linux-arm64` and `deepseek-tui-linux-arm64` (built natively on |
| 3409 | GitHub's `ubuntu-24.04-arm` runner). The npm wrapper picks them up |
| 3410 | automatically on `arm64` Linux hosts, so HarmonyOS thin-and-light, |
| 3411 | openEuler/Kylin, Asahi Linux, Raspberry Pi, AWS Graviton, etc. now work |
| 3412 | with a plain `npm i -g deepseek-tui`. |
| 3413 | - **Interactive TUI hangs on `working.` at 100% CPU (#549)** — the event |
| 3414 | loop's blocking terminal poll starved the tokio runtime, preventing the |
| 3415 | engine task from dispatching the API request. Fixed by yielding to the |
| 3416 | scheduler before each poll cycle and clamping the event-poll timeout to |
| 3417 | a minimum of 1ms so a zero-timeout hot-loop can't monopolize the thread. |
| 3418 | - **Backspace key inserts "h" instead of deleting (#550)** — terminals |
| 3419 | that send `^H` (Ctrl+H) for Backspace were not recognized. Added |
| 3420 | `is_ctrl_h_backspace()` guard in both the composer and API-key input |
| 3421 | handlers so Ctrl+H is treated as a delete, matching the existing |
| 3422 | `KeyCode::Backspace` behavior. |
| 3423 | |
| 3424 | ### Changed |
| 3425 | - **npm `postinstall` failure messages** — when no prebuilt is available for |
| 3426 | the host's `os.platform() / os.arch()` combo, the wrapper now prints the |
| 3427 | full `cargo install` fallback recipe and a link to |
| 3428 | [`docs/INSTALL.md`](INSTALL.md) instead of just the bare error. |
| 3429 | - **`DEEPSEEK_TUI_OPTIONAL_INSTALL=1`** — new env knob that downgrades a |
| 3430 | postinstall failure to a warning + `exit 0`, so CI matrices that include |
| 3431 | unsupported platforms don't fail the whole `npm install`. |
| 3432 | |
| 3433 | ### Docs |
| 3434 | - New [`docs/INSTALL.md`](docs/INSTALL.md) — every supported platform, |
| 3435 | prebuilt vs. `cargo install` vs. manual download, cross-compiling x64 → ARM64 |
| 3436 | Linux with `cross` or `gcc-aarch64-linux-gnu`, and a troubleshooting section |
| 3437 | covering the common `Unsupported architecture`, `MISSING_COMPANION_BINARY`, |
| 3438 | and self-update mismatch errors. |
| 3439 | - README and `README.zh-CN.md` now have an explicit **Linux ARM64** quickstart |
| 3440 | pointing ARM64 users at `cargo install deepseek-tui-cli deepseek-tui --locked` |
| 3441 | for v0.8.7 and at `npm i -g deepseek-tui` for v0.8.8+. |
| 3442 | |
| 3443 | ### Releases |
| 3444 | - npm wrapper publish remains manual (npm 2FA OTP requirement). |
| 3445 | - GitHub release automation depends on `RELEASE_TAG_PAT` secret — |
| 3446 | without it `auto-tag.yml` creates the tag but `release.yml` |
| 3447 | doesn't fire. |
| 3448 | |
| 3449 | ## [0.8.7] - 2026-05-03 |
| 3450 | |
| 3451 | ### Fixed |
| 3452 | - **Selection across transcript cell types** — the selection-tightening from |
| 3453 | v0.8.6 (#383) restricted copy/select to user and assistant message bodies |
| 3454 | only, so text in system notes, thinking blocks, and tool output could not be |
| 3455 | copied. v0.8.7 removes the body-start gate; the rendered transcript block is |
| 3456 | fully selectable again. |
| 3457 | |
| 3458 | ## [0.8.6] - 2026-05-03 |
| 3459 | |
| 3460 | ### Added |
| 3461 | - **Long-session survivability by default** (#402) — capacity control and |
| 3462 | compaction defaults are enabled, transcript history is bounded, persisted |
| 3463 | sessions are capped, and oversized history folds into archived context |
| 3464 | placeholders instead of freezing the TUI. |
| 3465 | - **v0.8.6 feature batch** (#373-#402) — adds goal tracking, cache-hit chips, |
| 3466 | cycle-boundary visualization, file-tree pane, `/share`, `/model auto`, |
| 3467 | user-defined slash commands, `/profile`, LSP diagnostic wiring, |
| 3468 | crash-recovery, self-update, `/init`, `/diff`, patch-aware `/undo`, |
| 3469 | `/edit`, inline diff highlighting, smart clipboard, native-copy escape, |
| 3470 | right-click context menus, clickable file:line styling, and MCP Phase A. |
| 3471 | |
| 3472 | ### Fixed |
| 3473 | - **Lag and rendering regressions** (#399, #400) — moves git/file-tree work |
| 3474 | off the UI thread where possible, bounds render history, and tightens redraw |
| 3475 | behavior to avoid sidebar/chat text bleed-through. |
| 3476 | - **Release-hardening follow-ups** — `/share` now writes via secure temp files, |
| 3477 | self-update uses secure same-directory temps with Windows-safe replacement, |
| 3478 | and docs/rustfmt release gates are clean. |
| 3479 | |
| 3480 | ## [0.8.4] - 2026-05-02 |
| 3481 | |
| 3482 | ### Added |
| 3483 | - **Localization expansion (Phase 1, #285)** — every slash command's help |
| 3484 | description, the full `/tokens` / `/cost` / `/cache` debug output, the |
| 3485 | footer state and chip text, and the help-overlay section headings are |
| 3486 | now translated for all four shipped locales (`en`, `ja`, `zh-Hans`, |
| 3487 | `pt-BR`). Set the language with `/config locale zh-Hans` (or |
| 3488 | `LANG=zh_CN.UTF-8` / `LC_ALL=zh_CN.UTF-8` from the shell). Non-Latin |
| 3489 | scripts render via the same `unicode_width` plumbing the existing 27 |
| 3490 | chrome strings already use; the `shipped_first_pack_has_no_missing_core_messages` |
| 3491 | test enforces full coverage across all four locales for every new |
| 3492 | `MessageId`. Tool descriptions sent to the model and the base system |
| 3493 | prompt intentionally remain English (training-data alignment, prefix |
| 3494 | cache stability). |
| 3495 | - Phase 1a (#294): 44 new IDs covering slash commands. |
| 3496 | - Phase 1b (#295): 13 new IDs covering `/tokens` / `/cost` / `/cache` |
| 3497 | debug output. Templates use `{placeholder}` substitution so a |
| 3498 | translator can re-order args freely. |
| 3499 | - Phase 1c (#296): 11 new IDs covering footer state, sub-agent chip, |
| 3500 | quit-confirmation toast, and help-overlay section labels. |
| 3501 | - **Stable cache prefix** (#263) — five companion fixes to keep the |
| 3502 | DeepSeek prefix cache stable across turns: drop volatile fields from |
| 3503 | the working-set summary block (#280, #287), place handoff and |
| 3504 | working-set after the static prompt blocks (#288 → #292), memoise the |
| 3505 | tool catalog so descriptions stay byte-stable (#289), sort |
| 3506 | `project_tree` and `summarize_project` output (#290), and use a unique |
| 3507 | fallback id for parallel streaming tool calls so downstream tool-result |
| 3508 | routing doesn't match the first call twice (#291). The combined effect |
| 3509 | is a meaningful jump in cache hit rate after the third turn. |
| 3510 | |
| 3511 | ### Fixed |
| 3512 | - **Agent-mode shell exec could not reach the network** (#272) — the seatbelt |
| 3513 | default policy denies all outbound network including DNS, so any |
| 3514 | `exec_shell` command needing the network (`curl`, `yt-dlp`, package |
| 3515 | managers, …) failed in Agent mode unless the user dropped to Yolo. The |
| 3516 | engine now elevates the sandbox policy to `WorkspaceWrite { network_access: |
| 3517 | true, … }` for both Agent and Yolo. Plan mode is unchanged (read-only |
| 3518 | investigation never registers the shell tool). The application-level |
| 3519 | `NetworkPolicy` (`crates/tui/src/network_policy.rs`) remains the only |
| 3520 | outbound-traffic boundary. |
| 3521 | - **`/skill install <github-repo-url>` failed with `invalid gzip header`** (#269) |
| 3522 | — `https://github.com/<owner>/<repo>` parsed as a raw direct URL, so the |
| 3523 | installer downloaded the HTML repo page and tried to gzip-decode HTML. |
| 3524 | Bare GitHub repo URLs (with or without `.git`, with or without `www.`, |
| 3525 | with or without a trailing slash) now route to the `GitHubRepo` source the |
| 3526 | same as `github:<owner>/<repo>`. URLs that already point at a specific |
| 3527 | archive / blob / tree path still go through `DirectUrl`. |
| 3528 | - **V4 Pro discount expiry extended** (#267) — DeepSeek extended the V4 Pro 75% |
| 3529 | promotional discount from 2026-05-05 15:59 UTC to 2026-05-31 15:59 UTC. Without |
| 3530 | this update the TUI would have started showing 4× the actual billed cost on |
| 3531 | May 6 onwards. Verified at https://api-docs.deepseek.com/quick_start/pricing. |
| 3532 | |
| 3533 | ## [0.8.3] - 2026-05-01 |
| 3534 | |
| 3535 | ### Fixed |
| 3536 | - **Skills prompt referenced fabricated paths** — `render_available_skills_context` |
| 3537 | rendered each skill's file as `<skills_dir>/<frontmatter-name>/SKILL.md`, |
| 3538 | which did not exist when the directory name differed from the frontmatter |
| 3539 | `name` (community installs, manually-placed skills). `Skill` now carries the |
| 3540 | real path captured at discovery and renders that. |
| 3541 | - **Missing-companion error was hostile to direct GitHub Release downloaders** |
| 3542 | (#258) — replaced "Build workspace default members to install it" wall of |
| 3543 | text with a concrete three-path checklist: `npm install -g deepseek-tui`, |
| 3544 | `cargo install deepseek-tui-cli deepseek-tui --locked`, or downloading both |
| 3545 | `deepseek-<platform>` AND `deepseek-tui-<platform>` from the same Release |
| 3546 | page. `DEEPSEEK_TUI_BIN` stays as a power-user fallback. |
| 3547 | |
| 3548 | ### Added |
| 3549 | - **Privacy: `$HOME` contracts to `~` in viewer-visible paths** — the TUI, |
| 3550 | `deepseek doctor`, `deepseek setup`, and onboarding now contract the home |
| 3551 | directory to `~` in every path shown on screen, so screenshots, screencasts, |
| 3552 | and pasted help output do not leak the OS account name. Persisted state, |
| 3553 | audit log, session checkpoints, and LLM-bound system prompts intentionally |
| 3554 | keep absolute paths for full fidelity. |
| 3555 | - **`crates.io` badge** alongside the CI and npm badges in both English and |
| 3556 | Simplified Chinese READMEs. |
| 3557 | - **Engine decomposition** (#227) — `core/engine.rs` is split into focused |
| 3558 | submodules (`engine/{streaming,turn_loop,dispatch,tool_setup,tool_execution,tool_catalog,context,approval,capacity_flow,lsp_hooks,tests}.rs`). |
| 3559 | No behavior change; preparation for the future agent-loop work. |
| 3560 | |
| 3561 | ### Tests |
| 3562 | - RLM bridge: `batch_guard` extracted and tested for the empty-batch and |
| 3563 | oversize-batch invariants; depth-guard fallback covered (partial #231). |
| 3564 | - Persistence: schema-version rejection covered for `load_session`, |
| 3565 | `load_offline_queue_state`, `runtime_threads::load_turn`, |
| 3566 | `runtime_threads::load_item` (partial #233). |
| 3567 | - Command palette: `[disabled]` server description tag (closes the |
| 3568 | remaining #197 acceptance gap). |
| 3569 | - Protocol-recovery contract tests now scan the engine submodules in |
| 3570 | addition to `engine.rs` so the decomposition refactor doesn't silently |
| 3571 | hide the fake-wrapper marker assertions. |
| 3572 | |
| 3573 | ### Issue triage |
| 3574 | - 10 issues closed with verification commits cited (#247, #235, #197, |
| 3575 | #250, #234, #243, #238, #236, #239, #195). |
| 3576 | |
| 3577 | ## [0.8.2] - 2026-05-01 |
| 3578 | |
| 3579 | ### Fixed |
| 3580 | - **Windows release build (LNK1104)** — drop the `deepseek` shim binary in |
| 3581 | `crates/tui` that 0.8.1 introduced for the bundled `cargo install`. It |
| 3582 | produced a second `target/release/deepseek.exe` that collided with the |
| 3583 | `deepseek-tui-cli` artifact during workspace builds; the second linker |
| 3584 | invocation hit `LNK1104: cannot open file deepseek.exe` on Windows. The |
| 3585 | cli crate is now the single source of `deepseek`; workspace default |
| 3586 | members still produce both binaries (one per crate). |
| 3587 | - **npm wrapper offline robustness** — `bin/deepseek(-tui).js` no longer |
| 3588 | re-fetches the GitHub-hosted SHA-256 checksum manifest on every invocation. |
| 3589 | When the binary is already installed and its `.version` marker matches the |
| 3590 | package version, the wrapper trusts the local file. The manifest is fetched |
| 3591 | lazily on actual download (first install or `DEEPSEEK_TUI_FORCE_DOWNLOAD=1`), |
| 3592 | so GitHub flakes, captive portals, corporate proxies, and offline state no |
| 3593 | longer break every command. |
| 3594 | |
| 3595 | ### Added |
| 3596 | - **Model-visible skills block** — installed skills (name, description, file |
| 3597 | path) are now exposed in the agent's system prompt under a `## Skills` |
| 3598 | section, with progressive disclosure: bodies stay on disk, the model opens a |
| 3599 | specific `SKILL.md` only when it decides to use that skill. Capped at a 12k |
| 3600 | prompt budget with 512-char per-description truncation. Threaded through |
| 3601 | `EngineConfig.skills_dir` so the TUI app, exec agent, and runtime thread |
| 3602 | manager all populate it from `Config::skills_dir()`. |
| 3603 | - **Simplified Chinese README** (`README.zh-CN.md`) with cross-link from the |
| 3604 | English README. |
| 3605 | |
| 3606 | ### Changed |
| 3607 | - **`cargo install` UX** — to install the canonical `deepseek` command, |
| 3608 | `cargo install deepseek-tui-cli` (the historical path). The 0.8.1 |
| 3609 | one-command flow (`cargo install deepseek-tui` providing both binaries) is |
| 3610 | reverted because it broke Windows release builds; install both packages |
| 3611 | separately if you want the TUI binary too. |
| 3612 | |
| 3613 | ## [0.8.1] - 2026-05-01 |
| 3614 | |
| 3615 | ### Fixed |
| 3616 | - **One-command Cargo install** — `cargo install deepseek-tui --locked` now |
| 3617 | provides both the canonical `deepseek` dispatcher and the `deepseek-tui` |
| 3618 | companion binary from the main `deepseek-tui` package, so dispatcher |
| 3619 | subcommands such as `deepseek doctor --json` work without installing |
| 3620 | `deepseek-tui-cli` separately. |
| 3621 | |
| 3622 | ## [0.8.0] - 2026-05-01 |
| 3623 | |
| 3624 | ### Fixed |
| 3625 | - **Shell FD leak / post-send lag** — completed background shell jobs now release |
| 3626 | their process, stdin, stdout, and stderr handles as soon as completion is |
| 3627 | observed, while keeping the job record inspectable. This prevents long-running |
| 3628 | TUI sessions from hitting `Too many open files (os error 24)`, which could |
| 3629 | make checkpoint saves fail and cause shell spawning, message send, close, and |
| 3630 | Esc/cancel paths to lag or fail. |
| 3631 | - **Windows REPL runtime CI startup** — Windows gets a longer Python bootstrap |
| 3632 | readiness timeout for the REPL runtime tests, matching GitHub runner startup |
| 3633 | contention without weakening bootstrap failures on other platforms. |
| 3634 | |
| 3635 | ### Added |
| 3636 | - **China / mirror-friendly Cargo install docs** — README now documents |
| 3637 | installing through the TUNA Cargo mirror and direct release assets for users |
| 3638 | with slow GitHub/npm access. |
| 3639 | |
| 3640 | ### Tests |
| 3641 | - Added a regression test proving completed background shell jobs drop their |
| 3642 | live process handles after `exec_shell_wait`. |
| 3643 | - Re-ran the focused shell cancellation and Python REPL runtime slices. |
| 3644 | |
| 3645 | ## [0.7.9] - 2026-05-02 |
| 3646 | |
| 3647 | ### Fixed |
| 3648 | - **Post-turn freeze** — the checkpoint-restart cycle boundary (`maybe_advance_cycle`) now runs *before* `TurnComplete` emission instead of after, so the terminal is immediately responsive when the UI receives the completion event. The status chip ("↻ context refreshing…") remains visible during the cycle wait. (#234) |
| 3649 | - **Enter during streaming no longer corrupts the turn** — a new `QueueFollowUp` submit disposition parks the draft on `queued_messages` when the model is actively streaming text. Previously, pressing Enter during streaming would forward the message as a mid-turn steer, which could interfere with the in-flight response. The message now dispatches as a normal user message after `TurnComplete`. (#234) |
| 3650 | - **Idempotent Esc during fanout** — `finalize_active_cell_as_interrupted` and `finalize_streaming_assistant_as_interrupted` are now guarded by `Option::take()`. When Esc cancels a turn and the engine later delivers `TurnComplete(Interrupted)`, the second call is a no-op — no double `[interrupted]` prefix, no corrupted cell state. Regression test locks in the contract. (#243) |
| 3651 | |
| 3652 | ### Tests |
| 3653 | - 2 new tests: `submit_disposition_queue_follow_up_when_streaming` (Enter/steering fix), `turn_complete_after_esc_is_idempotent` (Esc fanout double-call hardening) |
| 3654 | - 1 expanded test: `submit_disposition_queue_when_offline_and_busy` now covers streaming state |
| 3655 | |
| 3656 | ## [0.7.8] - 2026-05-01 |
| 3657 | |
| 3658 | ### Added |
| 3659 | - **`exec_shell_cancel` tool** — cancel a running background shell task by id, or cancel all running tasks with `all: true`. Requires approval. (#248) |
| 3660 | - **Foreground-to-background shell detach** — press `Ctrl+B` while a foreground command is running to open shell controls and either detach the command to the background (where it can be polled via `exec_shell_wait`) or cancel the current turn. (#248) |
| 3661 | - **`exec_shell_wait` turn-cancellation awareness** — canceling a turn while `exec_shell_wait` is blocking now stops the wait but leaves the background task running, with `wait_canceled: true` in metadata. (#248) |
| 3662 | - **`ShellControlView` modal** (Ctrl+B) — two-option dialog (Background / Cancel) rendered as a popup over the transcript. (#248) |
| 3663 | |
| 3664 | ### Changed |
| 3665 | - **`exec_shell` foreground path** now spawns all foreground commands through the background job table, enabling the detach-to-background flow. Metadata now includes `backgrounded: true/false`. (#248) |
| 3666 | - **`exec_shell_interact`** poll loop now observes the turn cancel token so stalled interactive sessions don't block turn cancellation. (#248) |
| 3667 | - **Transcript running-tool hint** — executing shell cells now show "Ctrl+B opens shell controls" while running. (#248) |
| 3668 | - **Keybinding registry** now includes `Ctrl+B` (opens shell controls) next to `Ctrl+C` (cancel/exits). (#248) |
| 3669 | - **Deferred swarm card creation** — `agent_swarm` no longer pre-seeds an all-pending FanoutCard from `ToolCallStarted`; the card is created only when the first `SwarmProgress` event carries real worker state. Until then the sidebar uses the declared task count as a pending dispatch placeholder. (#236, #238) |
| 3670 | - **Swarm wording normalized** — fanout-family fallback labels now render as `swarm`, matching the canonical `agent_swarm` / `rlm` model and avoiding mixed `fanout` / `swarm` terminology in the transcript. (#236, #238) |
| 3671 | - **OPERATIONS_RUNBOOK** and **TOOL_SURFACE** updated with new shell control paths and `exec_shell_cancel` documentation. |
| 3672 | |
| 3673 | ### Fixed |
| 3674 | - **Nonblocking swarm state drift** — the sidebar no longer falls back to `0` or a contradictory seeded placeholder before the first progress event arrives, which removes the visible `pending` vs `running/done` mismatch during early `agent_swarm` dispatch. (#236, #238) |
| 3675 | - **Unicode-safe search globbing** — search wildcard matching now iterates on UTF-8 char boundaries instead of raw byte offsets, preventing panics on filenames like `dialogue_line__冰糖.mp3`. (#249) |
| 3676 | |
| 3677 | ### Tests |
| 3678 | - 7 new integration tests: foreground-to-background detach, wait-cancel-leaves-process, single-task cancel, bulk cancel (kill-all), foreground-cancel-kills, ShellControlView default/select states |
| 3679 | - Expanded swarm/sidebar regression coverage for deferred card creation and pending-count fallback before first `SwarmProgress`. (#236, #238) |
| 3680 | - Added a Unicode filename regression test for wildcard search matching. (#249) |
| 3681 | |
| 3682 | ## [0.7.7] - 2026-04-30 |
| 3683 | |
| 3684 | ### Added |
| 3685 | - **Checklist card rendering** — `checklist_write` / `todo_*` results now render as a purpose-built card with completed/total + percent header, per-item status markers (✅ / `●` / `○`), and a collapsing affordance for long lists. Plumbed through `GenericToolCell` so no new variant threading is needed. (#241) |
| 3686 | - **Context menu for transcript operations** — right-click or `Ctrl+M` opens a context-sensitive menu with Copy, Copy All, and selection-aware actions. (`crates/tui/src/tui/context_menu.rs`) |
| 3687 | - **Windows .exe sibling lookup** — `locate_sibling_tui_binary` in the CLI dispatcher finds `deepseek-tui.exe` on Windows, honours `DEEPSEEK_TUI_BIN` override, and falls back to suffix-less lookup. Tests lock in platform-correct name resolution and env override. (#247) |
| 3688 | |
| 3689 | ### Changed |
| 3690 | - **Swarm/sub-agent canonical data model** — `SwarmTaskOutcome` and `SwarmOutcome` are now the single source of truth. Every UI surface (sidebar, transcript FanoutCard, footer) reads from `swarm_jobs` rather than maintaining parallel projections. (#236, #238) |
| 3691 | - **`swarm_card_index`** binds each swarm to its own FanoutCard by `swarm_id`, so overlapping fanouts no longer have one swarm's late progress clobber another's card. (#236, #238) |
| 3692 | - **Fanout-class tools suppressed from footer** — `agent_swarm`, `spawn_agents_on_csv`, `rlm`, and `agent_spawn` no longer appear as active tools in the status strip; sidebar and FanoutCard show the actual worker counts. (#236, #238) |
| 3693 | - **Esc clears active tool entries optimistically** — the active cell is finalized immediately on cancel rather than waiting for the engine's `TurnComplete` echo. Background `block:false` swarms remain durable and tracked through `swarm_jobs`. (#243) |
| 3694 | - **Post-turn workspace snapshot detached** — the snapshot still runs on `spawn_blocking` but the engine no longer awaits its `JoinHandle`, so the UI accepts input immediately after `TurnComplete`. (#234) |
| 3695 | - **Shell output preserves Cargo/test summaries under truncation** — high-signal tail lines (`test result:`, `failures:`, `error[E…]`, `Finished`, `Compiling`, panic markers) survive truncation so the agent doesn't re-run gates. (#242) |
| 3696 | - **Monotonic spend display** — `displayed_session_cost` + `displayed_cost_high_water` ensure the visible session+sub-agent total never decreases across reconciliation events (cache discounts, provisional → final). (#244) |
| 3697 | - Clipboard module expanded with additional platform-aware copy/paste paths. (`crates/tui/src/tui/clipboard.rs`) |
| 3698 | - Context inspector enriched with additional metadata columns and session-scoped agent state. (`crates/tui/src/tui/context_inspector.rs`) |
| 3699 | - Configuration documentation updated for v0.7.7 settings. (`docs/CONFIGURATION.md`, `docs/MODES.md`) |
| 3700 | |
| 3701 | ### Fixed |
| 3702 | - **Windows npm install path** — the npm-distributed `deepseek` dispatcher now locates the platform-correct `deepseek-tui` binary (`.exe` suffix on Windows), fixing runtime failures for Windows users. (#247) |
| 3703 | - **Sidebar/transcript/footer agreement** — all three surfaces now agree on agent counts and status because they share the canonical `swarm_jobs` store. (#236, #238) |
| 3704 | - **Fanout card clobbering** — overlapping swarms no longer overwrite each other's progress cards. (#238) |
| 3705 | - **Cost display regression** — negative reconciliation events (cache-hit discount applied after provisional count) no longer briefly drop the displayed cost. (#244) |
| 3706 | |
| 3707 | ### Tests |
| 3708 | - 65+ new/expanded tests: checklist card rendering, swarm card index binding, fanout tool suppression, Esc cancel contract, monotonic spend under reconciliation, shell summary preservation, Windows sibling binary lookup, clipboard platform paths, context menu state transitions |
| 3709 | |
| 3710 | ### Added |
| 3711 | - **UI Localization registry** — `locale` setting in `settings.toml` (`auto`, `en`, `ja`, `zh-Hans`, `pt-BR`) with `LC_ALL`/`LC_MESSAGES`/`LANG` auto-detection. Core packs shipped for English, Japanese, Chinese Simplified, and Brazilian Portuguese covering composer placeholder, history search, `/config` chrome, and help overlay. Missing/unsupported locales fall back to English. (`crates/tui/src/localization.rs`, `docs/CONFIGURATION.md`) |
| 3712 | - **Grouped, searchable `/config` editor** — settings organized by section (Model, Permissions, Display, Composer, Sidebar, History, MCP) with live substring filter. Typing `j`/`k` navigates when the filter is empty; otherwise they enter the filter. (`crates/tui/src/tui/views/mod.rs`) |
| 3713 | - **Pending input preview widget** — while a turn is running, queued messages, pending steers, rejected steers, and context chips render above the composer. Three-row-per-message truncation with ellipsis overflow. (`crates/tui/src/tui/widgets/pending_input_preview.rs`) |
| 3714 | - **Alt+↑ edit-last-queued** — pops the most recently queued message back into the composer for editing. No-op when the composer is dirty. (`crates/tui/src/tui/app.rs`) |
| 3715 | - **Composer history search and draft recovery** — `Alt+R` opens a live substring search across `input_history` and `draft_history` (max 50 entries). `Enter` accepts, `Esc` restores the pre-search draft. Unicode case-insensitive matching. (`crates/tui/src/tui/app.rs`) |
| 3716 | - **Paste-burst detection** — fallback rapid-key paste detection independent of terminal bracketed-paste mode. Configurable via `paste_burst_detection` setting (default on). CRLF normalization (`\r\n` → `\n`, `\r` → `\n`). (`crates/tui/src/tui/paste_burst.rs`) |
| 3717 | - **Composer attachment management** — `↑` at the composer start selects the attachment row; `Backspace`/`Delete` removes it without editing placeholder text. (`crates/tui/src/tui/app.rs`) |
| 3718 | - **Searchable help overlay** — live substring filter across slash commands and keybindings, multi-term AND matching, localized chrome. (`crates/tui/src/tui/views/help.rs`) |
| 3719 | - **Keyboard-binding documentation catalog** — single source of truth for help overlay rendering. Documents 38+ keyboard chords across Navigation, Editing, Submission, Modes, Sessions, Clipboard, and Help sections. (`crates/tui/src/tui/keybindings.rs`) |
| 3720 | - **Legacy Rust deprecation audit** — non-destructive compatibility audit covering legacy MCP sync API, prompt constants, `/compact`, `todo_*` aliases, sub-agent aliases, provider `api_key` compatibility, model alias canonicalization, and palette aliases. Tracked by #218–#221. (`docs/LEGACY_RUST_AUDIT_0_7_6.md`) |
| 3721 | |
| 3722 | ### Changed |
| 3723 | - **Shift+Tab cycles reasoning-effort** through Off → High → Max (three behaviorally distinct tiers). Previously Tab cycled modes; Shift+Tab is now the reasoning-effort shortcut. (`crates/tui/src/tui/app.rs:1119`) |
| 3724 | - **Reasoning-effort `Off` now sends `"off"`** to the API (was `None`). Allows explicit thinking disable. (`crates/tui/src/tui/app.rs`) |
| 3725 | - **Media `@`-mentions now emit `<media-file>` hints** directing users to `/attach` instead of inlining binary bytes. Tests lock in the contract. (`crates/tui/src/tui/file_mention.rs`) |
| 3726 | - **`/attach` rejects non-media files** with a descriptive error pointing to `@path` for text. (`crates/tui/src/commands/attachment.rs`) |
| 3727 | - **Configuration reference updated** to cover all v0.7.6 settings: `locale`, `paste_burst_detection`, `reasoning_effort`, `composer_density`, `sidebar_focus`, and more. (`docs/CONFIGURATION.md`) |
| 3728 | |
| 3729 | ### Fixed |
| 3730 | - **Unicode-safe truncation** in pending-input preview and view text — no more mid-character breaks on multi-byte UTF-8. (`crates/tui/src/tui/widgets/pending_input_preview.rs`, `crates/tui/src/tui/views/mod.rs`) |
| 3731 | - **CJK/emoji display-width handling** in locale tests and config view rendering. (`crates/tui/src/localization.rs`) |
| 3732 | - **Context preview distinguishes `@media`, `/attach`, missing, and included files** with separate kind labels and inclusion status. (`crates/tui/src/tui/file_mention.rs`) |
| 3733 | - **Config view filter accept `j`/`k` only when filter is empty** — typing `j` or `k` into the filter field no longer navigates away. (`crates/tui/src/tui/views/mod.rs`) |
| 3734 | |
| 3735 | ### Tests |
| 3736 | - 7 localization tests (tag normalization, env resolution, shipped pack completeness, missing-key fallback, Unicode width truncation) |
| 3737 | - 11 pending-input preview tests (context buckets, truncation, URL overflow, narrow-width) |
| 3738 | - 13 paste tests (burst detection, CRLF normalization, clipboard images, Unicode) |
| 3739 | - 9 draft/history search tests (match filter, unicode, accept/cancel, recovery) |
| 3740 | - 93 config tests (grouping, filter, edit, j/k, localization, escape/cancel) |
| 3741 | - 24 workspace tests (context refresh, scroll, mention completion) |
| 3742 | - 7 file-mention tests (context references, media/attach distinction, removability) |
| 3743 | |
| 3744 | ## [0.7.1] - 2026-04-28 |
| 3745 | |
| 3746 | ### Added |
| 3747 | - Grouped active tool-call cards with compact rails and a live working-status row while tools run. (#142, #149) |
| 3748 | - Selected-card-aware Alt+V details so the visible or selected tool card opens the matching detail payload. (#143) |
| 3749 | - Compact terminal-native session context inspector with persisted `@path` and `/attach` reference metadata for resumed transcripts. (#146, #150) |
| 3750 | |
| 3751 | ### Changed |
| 3752 | - Polished tool cards, diff summaries, and pending context previews for denser terminal-native scanning. (#141, #144, #145, #148) |
| 3753 | - Ranked Ctrl+P file-picker results with working-set relevance from modified files, recent `@file` mentions, and recent tool paths while keeping fuzzy filtering in memory. (#147) |
| 3754 | |
| 3755 | ## [0.7.0] - 2026-04-28 |
| 3756 | |
| 3757 | ### Added |
| 3758 | - OS keyring-backed auth storage with `deepseek auth` subcommands, migration from plaintext config, provider-aware key resolution, and doctor visibility. (#134) |
| 3759 | - Egress network policy with allow/deny/prompt decisions, deny-wins matching, audit logging, and enforcement hooks for network-capable tools. (#135) |
| 3760 | - LSP diagnostics auto-injection after edits so compile feedback can be reinjected into the next agent turn. (#136) |
| 3761 | - Side-git workspace snapshots, `/restore`, and `revert_turn` so agent edits can be rolled back without moving the user's repository HEAD. (#137) |
| 3762 | - Esc-Esc backtrack over prior user turns, desktop turn-complete notifications, Alt+V tool-details access, safer command-prefix auto-allow matching, bundled `skill-creator`, and `/skill install` management for community skills. (#131, #132, #133, #138, #139, #140) |
| 3763 | |
| 3764 | ### Changed |
| 3765 | - Split more engine/tool primitives into focused modules and workspace crates, including shared tool result primitives and extracted turn/capacity flow. (#67, #74) |
| 3766 | |
| 3767 | ### Tests |
| 3768 | - Added mock LLM and skill-install integration coverage for streaming turns, reasoning replay, tool-call loops, network policy, and skill validation. (#69, #140) |
| 3769 | |
| 3770 | ## [0.6.5] - 2026-04-27 |
| 3771 | |
| 3772 | ### Added |
| 3773 | - **`rlm_process` tool — recursive language model as a tool call.** The previous `/rlm` slash command had a UI rendering gap (the answer never made it back to the model's view) and required the user to remember to invoke it manually. `rlm_process` exposes the full RLM loop as a structured tool the model itself can choose, the same way it reaches for `agent_spawn` or `rlm_query`. Inputs: `task` (small instruction, shown to the root LLM each iteration) plus exactly one of `file_path` (workspace-relative, preferred — keeps the long input out of the model's context entirely) or `content` (inline, capped at 200k chars). Optional `child_model` (default `deepseek-v4-flash`) and `max_depth` (default 1, paper experiments). Returns the synthesized answer with metadata (iterations, duration, tokens, termination reason). Loaded across Plan / Agent / YOLO; never deferred via ToolSearch. (`crates/tui/src/tools/rlm_process.rs`) |
| 3774 | - **Reference-aligned REPL surface.** Aligned the in-REPL Python helpers with the canonical reference RLM (alexzhang13/rlm). The sub-agent now sees `context` (the full input, not `PROMPT`), `llm_query`, `llm_query_batched`, `rlm_query` (was `sub_rlm`), `rlm_query_batched`, `SHOW_VARS()`, `FINAL(...)`, `FINAL_VAR(...)`, plus `repl_get`/`repl_set`. Same prompt patterns and decomposition strategies from the paper now apply verbatim. (`crates/tui/src/repl/runtime.rs`) |
| 3775 | - **Concurrent fanout from inside the REPL.** `llm_query_batched(prompts, model=None)` runs up to 16 child completions in parallel via a new `POST /llm_batch` sidecar endpoint — much faster than serial `[llm_query(p) for p in prompts]`. `rlm_query_batched(prompts)` does the same for recursive RLM sub-calls via `POST /rlm_batch`. (`crates/tui/src/rlm/sidecar.rs`) |
| 3776 | - **`SHOW_VARS()`** — returns `{name: type-name}` for every user variable in the REPL. Lets the model inspect what it has accumulated across rounds before deciding whether to call `FINAL_VAR(name)`. |
| 3777 | - **Auto-persistence of REPL variables across rounds.** Any top-level JSON-serializable variable the sub-agent creates in a `repl` block now persists to the next round automatically — no `repl_set` ceremony needed unless you want explicit control. Matches the in-process reference REPL semantics. |
| 3778 | |
| 3779 | ### Changed |
| 3780 | - **Code fence is `repl`, not `python`.** Matches the reference RLM language identifier so the same prompts and few-shot examples work here. Backward-compat fallback to `python` / `py` retained for older model behaviors. |
| 3781 | - **`FINAL` / `FINAL_VAR` parseable from raw response text.** The reference RLM lets the model write `FINAL(value)` on its own line outside any code block to terminate the loop. Added `parse_text_final()` so that path works alongside the existing in-REPL Python sentinel mechanism. Code-fenced occurrences of `FINAL(...)` are correctly ignored to avoid false positives. |
| 3782 | - **Strict termination loop.** The sub-agent must emit a ```repl block (or text-level FINAL) to make progress. One fence-less round triggers a reminder; two consecutive trigger a `RlmTermination::DirectAnswer` exit so we don't loop forever. |
| 3783 | - **`rlm_process` separates `task` (root_prompt) from `file_path`/`content` (context).** The `task` rides along as `root_prompt` and is shown to the root LLM each iteration; the big input lives only in the REPL as `context`. Mirrors the reference's `completion(prompt, root_prompt=...)` API. |
| 3784 | - **System prompt rewritten** with the reference's strategy patterns (PREVIEW → CHUNK + map-reduce via `llm_query_batched` → RECURSIVE decomposition via `rlm_query` → programmatic computation + LLM interpretation). |
| 3785 | - The `/rlm` slash command stays for manual experimentation but is no longer the recommended path; the description in `commands/mod.rs` now points the model toward `rlm_process` for the in-agent flow. |
| 3786 | |
| 3787 | ### Reference |
| 3788 | - Zhang, Kraska, Khattab. "Recursive Language Models." arXiv:2512.24601. |
| 3789 | - alexzhang13/rlm — reference implementation by the paper authors. Variable names, helper surface, and code-fence convention align with that repo so prompts and patterns transfer. |
| 3790 | |
| 3791 | |
| 3792 | ### Fixed |
| 3793 | - **`/rlm` actually recurses now (Algorithm 1 substrate, paper-faithful).** The v0.6.3 RLM loop had the right *shape* but its recursive substrate was non-functional: `llm_query()` was a Python stub that returned a hardcoded string, and `child_model` was bound with an underscore prefix and silently dropped. The loop ran but the sub-LLM never fired. v0.6.4 fixes this end-to-end: |
| 3794 | - **HTTP sidecar.** Each RLM turn spins up a localhost-only axum server on a kernel-assigned port for the duration of the turn. Python's `llm_query()` and `sub_rlm()` are real `urllib.request.urlopen` POSTs; Rust services them via the existing DeepSeek client and returns the completion text. No long-lived python process, no FIFOs, no two-pass replay — Python blocks on HTTP, Rust answers it. (`crates/tui/src/rlm/sidecar.rs`) |
| 3795 | - **`child_model` is plumbed through.** `Op::RlmQuery` and `AppAction::RlmQuery` carry the configured child model (default `deepseek-v4-flash`) all the way to the sidecar, where every `llm_query()` call uses it. Token usage is folded into `RlmTurnResult.usage` so cost tracking works. |
| 3796 | - **`sub_rlm()` is exposed as a paper-faithful recursive RLM call.** The Python REPL gets a real `sub_rlm(prompt)` function that runs another full Algorithm-1 turn at depth-1 inside the same process (different sidecar route, decremented recursion budget). Default `max_depth = 2` from the `/rlm` command — the model can recurse twice before the budget hits zero. The recursive opaque-future cycle (`run_rlm_turn_inner` → `start_sidecar` → `sub_rlm_handler` → `run_rlm_turn_inner`) is broken by returning a concrete `Pin<Box<dyn Future + Send>>` from `run_rlm_turn_inner`. |
| 3797 | - **Strict termination.** The loop only ends via `FINAL(value)` (or the iteration cap). The previous "no fence = direct answer, end loop" early-exit deviated from the paper and could short-circuit on iteration 1 with a chatty model that never saw `PROMPT`. The new behavior tolerates one fence-less round (with a reminder appended), then falls back to a `RlmTermination::DirectAnswer` exit. `RlmTurnResult` now carries a `termination: RlmTermination` enum (`Final | DirectAnswer | Exhausted | Error`) so callers can tell what happened. |
| 3798 | - **Richer `Metadata(state)`.** The metadata message the root LLM sees now includes paper-required *access patterns* (`repl_get`, slicing, `splitlines`, `repl_set`, `llm_query`, `sub_rlm`, `FINAL`) and a live list of variable keys currently in the REPL state file — so the model can see what it's accumulated across rounds without us shipping the values themselves. |
| 3799 | - **Unicode-safe truncation.** `truncate_text` now counts Unicode codepoints (was mixing `text.len()` bytes with `chars().take(n)`), so multi-byte previews can no longer mis-count. Per-turn temp state files are cleaned up on completion. `ROOM_TEMPERATURE` typo → `ROOT_TEMPERATURE`. |
| 3800 | - **End-to-end smoke test.** `rlm::turn::tests::sidecar_url_is_exported_to_python_env` stands up a stand-in axum server that always replies `{"text":"pong-from-sidecar"}`, runs `print(llm_query('hello'))` in the real `PythonRuntime`, and asserts the reply round-trips. This catches future regressions in the sidecar URL passthrough. |
| 3801 | |
| 3802 | ### Reference |
| 3803 | - Zhang, Kraska, Khattab. "Recursive Language Models." arXiv:2512.24601 (Algorithm 1). |
| 3804 | |
| 3805 | |
| 3806 | ### Added |
| 3807 | - **Sub-agents surface in the footer status strip.** When N > 0 sub-agents are in flight, the footer grows a "1 agent" / "N agents" chip in DeepSeek-sky color matching the model badge. Hides entirely at zero. (`footer_agents_chip` in `widgets/footer.rs`) |
| 3808 | - **`@`-mention popup is fully wired in the composer.** Previously only the App state fields existed (`mention_menu_selected`, `mention_menu_hidden`). The popup now renders below the input mirror-style with the slash menu, with `@`-prefixed entries; Up/Down navigates, Enter / Tab apply the selection, Esc hides until the next input edit. Mention takes precedence over slash because the positional check is stricter. (`visible_mention_menu_entries` + `apply_mention_menu_selection` in `file_mention.rs`) |
| 3809 | |
| 3810 | ### Fixed |
| 3811 | - **Tool-call cells no longer flash `<command>` / `<file>` placeholders.** The engine used to emit `ToolCallStarted` from `ContentBlockStart` with `input: {}` — before any `InputJsonDelta` had streamed in — which baked the placeholder into the cell at creation time. The emission is now deferred to `ContentBlockStop` and routed through `final_tool_input`, so the cell is created with the parsed args already in hand. (engine.rs `final_tool_input`; engine/tests.rs `final_tool_input_*`) |
| 3812 | - **`parse_invocation_count` flake.** Two `markdown_render` tests both read the global PARSE_INVOCATIONS atomic and raced when other tests called `parse()` in parallel. Switched the counter to `thread_local!<Cell<u64>>`, so each test thread sees only its own invocations. Tested 8 sequential full-suite runs: 8/8 green (was ~40% green). |
| 3813 | |
| 3814 | ### Changed |
| 3815 | - **System prompts redesigned with decomposition-first philosophy.** All four prompt tiers (base, agent, plan, yolo) now teach the model to decompose tasks before acting — `todo_write` first for granular task tracking, `update_plan` for high-level strategy, and sub-agents for parallelizable work. Inspired by the "mismanaged geniuses hypothesis" (Zhang et al., 2026): frontier LMs are already capable enough; the bottleneck is how we scaffold their self-management. The prompts now make work visible through the sidebar (Plan / Todos / Tasks / Agents) instead of letting the model work invisibly. |
| 3816 | - **Tool labels use progressive verbs.** "Read foo.rs" → "Reading foo.rs", "List X" → "Listing X", "Search pattern" → "Searching for `pattern`", "List files" → "Listing files". Past-tense labels read wrong while a tool is still in flight; the new forms match what the user actually sees. |
| 3817 | - **Long-running tools grow an elapsed badge.** From 3 s onward the `running` status segment becomes `running (3s)`, `running (4s)`, … so the user can tell a tool isn't stuck. The status-animation tick (360 ms) drives the redraw; below 3 s the badge stays hidden so quick reads/greps don't churn. (history.rs `running_status_label_with_elapsed`) |
| 3818 | - **Spinner pulse is twice as fast** — `TOOL_STATUS_SYMBOL_MS` 1800 ms → 720 ms per glyph (full 4-glyph heartbeat in ~2.88 s instead of ~7.2 s). |
| 3819 | - **`tools/subagent.rs` is now a folder module.** Tests live in `tools/subagent/tests.rs`; runtime + manager + tool implementations stay in `tools/subagent/mod.rs`. Public API unchanged. The runtime / tool-impl split was deferred — `SubAgentTask`, `run_subagent_task`, `build_allowed_tools`, the agent prompt constants, and `normalize_role_alias` are referenced from both layers and need a small API design pass before they cleanly separate. |
| 3820 | |
| 3821 | ### Test hygiene |
| 3822 | - **5 regression tests pin auto-scroll churn contract.** `mark_history_updated` does not scroll; tool-cell handlers only `mark_history_updated`; `add_message` and `flush_active_cell` gate on `user_scrolled_during_stream`; the per-stream lock clears at TurnComplete and when the user returns to the live tail. (P2.4) |
| 3823 | |
| 3824 | ## [0.6.1] - 2026-04-26 |
| 3825 | |
| 3826 | ### Changed |
| 3827 | - **V4 cache-hit input prices cut to 1/10th per DeepSeek's pricing update.** Pro promo 0.03625→0.003625, Pro base 0.145→0.0145, Flash 0.028→0.0028 per 1M tokens. Cache-miss and output rates unchanged. |
| 3828 | - **Removed the "light" theme option.** It was never tested, looked bad, and the dark/whale palettes are the supported targets. Theme validation now accepts only `default`, `dark`, and `whale`. |
| 3829 | - **System prompts redesigned with decomposition-first philosophy.** All five prompt tiers teach the model to `todo_write` before acting, `update_plan` for strategy, and sub-agents for parallel work. Inspired by the mismanaged-geniuses hypothesis (Zhang et al., 2026). |
| 3830 | |
| 3831 | ## [0.6.0] - 2026-04-25 |
| 3832 | |
| 3833 | ### Added |
| 3834 | - **`rlm_query` tool — recursive language models as a first-class structured tool.** Inspired by [Alex Zhang's RLM work](https://github.com/alexzhang13/rlm) and Sakana AI's published novelty-search research, but trimmed to what an agent loop actually needs. The model calls `rlm_query` with one prompt or up to 16 concurrent prompts; children run on `deepseek-v4-flash` by default and can be promoted to Pro per-call. Children dispatch concurrently via `tokio::join_all` against the existing DeepSeek client — no external runtime, no fenced-block DSL, no Python sandbox. Returns plain text for one prompt, indexed `[0] ...\n\n---\n\n[1] ...` blocks for many. Available in Plan / Agent / YOLO. Cost is folded into the session's running total automatically. |
| 3835 | |
| 3836 | ### Changed |
| 3837 | - **Scroll position survives content rewrites (#56).** `TranscriptScroll::resolve_top` and `scrolled_by` no longer teleport to bottom when the anchor cell vanishes. Three-level fallback chain: same line → same cell, line 0 → nearest surviving cell at-or-before. Previously, any rewrite of the assistant message (e.g. tool-result replacement) silently dropped the user back to the live tail mid-scroll. |
| 3838 | - **Looser command-safety chains (#57).** `cargo build && cargo test`, `git fetch && git rebase`, and similar chains of known-safe commands now escalate to `RequiresApproval` instead of being hard-blocked as `Dangerous`. Chains containing unknown commands still block. |
| 3839 | - **`GettingCrowded` no longer surfaces a footer chip.** The context-percent header already covers conversation pressure; the chip now only fires for active engine interventions (`refreshing context`, `verifying`, `resetting plan`). |
| 3840 | |
| 3841 | ## [0.5.2] - 2026-04-25 |
| 3842 | |
| 3843 | ### Added |
| 3844 | - **`/model` opens a Pro/Flash + thinking-effort picker (#39).** Typing `/model` with no argument now pops a two-pane modal: model on the left (`deepseek-v4-pro` flagship, `deepseek-v4-flash` fast/cheap, plus a "current (custom)" row when the active id isn't one of the listed defaults), and thinking effort on the right. Tab/←/→ swaps panes, ↑/↓ moves within the focused pane, Enter applies both selections, Esc cancels. The effort pane intentionally exposes only **Off / High / Max** because [DeepSeek's Thinking Mode docs](https://api-docs.deepseek.com/guides/reasoning_model) state `low`/`medium` are mapped to `high` server-side and `xhigh` is mapped to `max` — the legacy variants stay valid in `~/.deepseek/settings.toml` for back-compat, the picker just doesn't surface them. Apply path persists `default_model` and `reasoning_effort` to settings, forwards `Op::SetModel` + `Op::SetCompaction` to the running engine so the next turn picks up the change without a restart, and resets the per-turn token gauges (cache, replay) so the footer numbers reflect the new model. `/model <id>` keeps working unchanged for power users. |
| 3845 | |
| 3846 | ## [0.5.1] - 2026-04-25 |
| 3847 | |
| 3848 | ### Added |
| 3849 | - **`fetch_url` tool** for direct HTTP GET on a known URL — complements `web_search` for cases where the link is already known. Supports `format` (`markdown` / `text` / `raw`), `max_bytes` (default 1 MB, hard cap 10 MB), `timeout_ms` (default 15 s, max 60 s), redirect following, and structured `{url, status, content_type, content, truncated}` responses. 4xx/5xx bodies are returned (with `success: false`) so the caller can read JSON error envelopes. (#33) |
| 3850 | - **PDF support in `read_file`.** PDFs are auto-detected by extension or `%PDF-` magic bytes and extracted via `pdftotext -layout` (poppler) when available. New optional `pages` arg (`"5"` or `"1-10"`) reads page slices. Without `pdftotext`, returns a structured `{type: "binary_unavailable", kind: "pdf", reason, hint}` with install commands for macOS/Debian. (#34) |
| 3851 | - **Reasoning-content replay telemetry, end-to-end (#30).** The chat-completions sanitizer now estimates replayed `reasoning_content` tokens (~4 chars/token), threads the value through the streaming `Usage` payload, stores it on the App, and renders an `rsn N.Nk` chip in the footer next to the cache hit-rate. The chip turns warning-coloured when replay tokens exceed 50% of the input budget, so users on long thinking-mode loops can see at a glance how much of their context window is going to V4's "Interleaved Thinking" replay (paper §5.1.1). Logged at `RUST_LOG=deepseek_tui=info` for tail-friendly diagnosis. |
| 3852 | - **`@file` Tab-completion (#28).** Typing `@<partial>` and pressing Tab now resolves the mention against the workspace using the existing `ignore::WalkBuilder`. A unique match is spliced into the input; multiple matches with a longer common prefix extend the partial; remaining ambiguity is surfaced via the status line. The mention-expansion path that ships file contents to the model is unchanged — this is purely a discovery aid for typing the path. Inline-contents and a fuzzy popup picker are queued for v0.5.2. |
| 3853 | - **Per-workspace external trust list (#29).** `~/.deepseek/workspace-trust.json` now records, for each workspace, the absolute paths the user has opted into reading/writing from outside that workspace. The new `/trust` slash command supports `add <path>`, `remove <path>`, `list`, `on`, `off`, and a status read with no args; the engine consults the list when constructing every `ToolContext` so changes apply on the next tool call without restart. `/diagnostics` surfaces the list. The interactive "Allow once / Always allow / Deny" approval prompt is deferred — for now grant access ahead of the turn with `/trust add <path>`. |
| 3854 | |
| 3855 | ### Fixed |
| 3856 | - **TUI sidebar gutter bleed regression test (#36).** Snapshot tests now lock in that long single-line tool results — including a `todo_write` echo of a multi-kilobyte JSON payload — never write any cells outside `chat_area` at the widths reported in the bug (80, 120, 165, 200 cols). A second test verifies the scrollbar coexists with content along the right edge instead of overdrawing the penultimate column. |
| 3857 | - **Version drift caught in CI.** New `versions` job in `.github/workflows/ci.yml` runs `scripts/release/check-versions.sh` on every push/PR, verifying every per-crate `Cargo.toml` inherits the workspace version, the npm wrapper matches the workspace version, and `Cargo.lock` is in sync. The release runbook now lists `check-versions.sh` as the first preflight step. (#31) |
| 3858 | - **Per-mode soft context budget for V4 compaction trigger** (#27). |
| 3859 | - **Phantom `web.run` references stripped** from prompts and the `web_search` tool surface (#25). |
| 3860 | - **Unused import + `cargo fmt` drift** that landed with `feat(#27)` and broke Build / Test / npm wrapper smoke under `-Dwarnings`. |
| 3861 | |
| 3862 | ## [0.5.0] - 2026-04-25 |
| 3863 | |
| 3864 | ### Fixed |
| 3865 | - Multi-turn tool calls on thinking-mode models no longer return HTTP 400. Every assistant message in the conversation now carries `reasoning_content` when thinking is enabled — not just tool-call rounds — matching DeepSeek's actual API validation, which rejects any assistant message missing the field even though the docs describe non-tool-call reasoning as "ignored". |
| 3866 | - Added a final-pass wire-payload sanitizer in the chat-completions client that forces a non-empty `reasoning_content` placeholder onto any assistant message still missing one at request time. This is the last line of defense after engine-side and build-side substitution, so sessions restored from older checkpoints, sub-agents that append messages directly, and cached prefix mismatches all produce a valid request. |
| 3867 | - On a `reasoning_content`-related 400, the client now logs the offending message indices to make future regressions diagnosable. |
| 3868 | - Stripped phantom `web.run` references from prompts and the `web_search` tool surface ([#25](https://github.com/Hmbown/CodeWhale/issues/25)). |
| 3869 | |
| 3870 | ### Changed |
| 3871 | - Header/UI widget refactor in the TUI (`crates/tui/src/tui/ui.rs`, `widgets/header.rs`) — internal cleanup, no user-visible behavior change. |
| 3872 | |
| 3873 | ## [0.4.9] - 2026-04-27 |
| 3874 | |
| 3875 | ### Fixed |
| 3876 | - DeepSeek thinking-mode tool-call rounds now always replay `reasoning_content` in all subsequent requests (including across new user turns), matching DeepSeek's documented API contract that assistant messages with tool calls must retain their reasoning content forever. |
| 3877 | - Missing `reasoning_content` on a tool-call assistant message now substitutes a safe placeholder (`"(reasoning omitted)"`) instead of dropping the tool calls and their matching tool results, preventing orphaned conversation chains and API 400 errors. |
| 3878 | - Session checkpoint now persists a Thinking-block placeholder for tool-call turns that produced no streamed reasoning text, keeping on-disk sessions structurally correct so subsequent requests avoid HTTP 400 rejections. |
| 3879 | - Token estimation for compaction now counts thinking tokens across all tool-call rounds (not just the current user turn), aligning with the updated reasoning_content replay rule. |
| 3880 | |
| 3881 | ## [0.4.8] - 2026-04-25 |
| 3882 | |
| 3883 | ### Fixed |
| 3884 | - DeepSeek V4 Pro cost estimates now use DeepSeek's current limited-time 75% discount until 2026-05-05 15:59 UTC, then automatically fall back to the base Pro rates. |
| 3885 | |
| 3886 | ## [0.4.5] - 2026-04-24 |
| 3887 | |
| 3888 | ### Fixed |
| 3889 | - Alternate-screen TUI sessions now capture mouse input by default so wheel scrolling moves the transcript instead of exposing terminal scrollback from before the TUI started. Use `--no-mouse-capture` or `tui.mouse_capture = false` when terminal-native drag selection is preferred. |
| 3890 | |
| 3891 | ## [0.4.2] - 2026-04-24 |
| 3892 | |
| 3893 | ### Fixed |
| 3894 | - DeepSeek V4 thinking-mode tool turns now checkpoint the engine's authoritative API transcript, including assistant `reasoning_content` on reasoning-to-tool-call turns with no visible assistant text. |
| 3895 | - Chat Completions request building now drops stale V4 tool-call rounds that are missing required `reasoning_content`, preventing old corrupted checkpoints from triggering DeepSeek HTTP 400 replay errors. |
| 3896 | - Web search now falls back to Bing HTML results when DuckDuckGo returns a bot challenge or otherwise yields no parseable results. |
| 3897 | |
| 3898 | ## [0.4.1] - 2026-04-24 |
| 3899 | |
| 3900 | ### Fixed |
| 3901 | - DeepSeek V4 tool-result context now preserves large file reads and command outputs instead of compacting noisy tools to a 900-character snippet after 2k characters. |
| 3902 | - Capacity guardrail refresh no longer performs destructive summary compaction unless the normal model-aware compaction thresholds are actually crossed. |
| 3903 | - V4 compaction summaries retain larger tool-result excerpts and summary input when compaction is genuinely needed. |
| 3904 | - The transcript now follows the bottom again when sending a new message, shows an in-app scrollbar when internally scrolled, and leaves mouse capture off in `--no-alt-screen` mode so terminal-native scrolling can work. |
| 3905 | |
| 3906 | ## [0.4.0] - 2026-04-23 |
| 3907 | |
| 3908 | ### Added |
| 3909 | - **DeepSeek V4 support**: `deepseek-v4-pro` (flagship) and `deepseek-v4-flash` (fast/cheap) are now first-class model IDs with 1M context windows. |
| 3910 | - **Reasoning-effort tier**: new `reasoning_effort` config field (`off | low | medium | high | max`) mapped to DeepSeek's `reasoning_effort` + `thinking` request fields. Defaults to `max`. |
| 3911 | - **Shift+Tab cycles reasoning-effort** through the three behaviorally distinct tiers (`off → high → max`). The current tier is shown as a ⚡ chip in the header. |
| 3912 | - Per-model pricing table: `deepseek-v4-pro` priced at $0.145/$1.74/$3.48 per 1M tokens (cache-hit/miss/output); `deepseek-v4-flash` and legacy aliases at $0.028/$0.14/$0.28. |
| 3913 | |
| 3914 | ### Changed |
| 3915 | - **Default model flipped to `deepseek-v4-pro`** (from `deepseek-reasoner`). |
| 3916 | - `deepseek-chat` / `deepseek-reasoner` remain as silent aliases of `deepseek-v4-flash` for API compatibility; priced identically. |
| 3917 | - **Context compaction**: 1M-context V4 models now compact at 800k input tokens or 2,000 messages, so short/tool-heavy sessions do not compact as if they were 128k-context runs. |
| 3918 | - Cycling modes is now Tab-only; Shift+Tab is repurposed for reasoning-effort (reverse-mode cycle was low-value with only three modes). |
| 3919 | - Updated help/hint strings, validator error messages, and the model picker to reference V4 IDs. |
| 3920 | |
| 3921 | ### Fixed |
| 3922 | - `requires_reasoning_content` now recognizes `deepseek-v4*` so thinking streams render correctly on V4 models. |
| 3923 | - DeepSeek V4 thinking-mode tool calls now preserve prior assistant `reasoning_content` whenever a tool call is replayed, matching DeepSeek's multi-turn contract and avoiding HTTP 400 rejections on later turns. |
| 3924 | - Raw Chat Completions requests now send DeepSeek's top-level `thinking` parameter instead of the OpenAI SDK-only `extra_body` wrapper. |
| 3925 | - Config, env, and UI model selection now normalize legacy DeepSeek aliases to `deepseek-v4-flash` instead of preserving old model labels. |
| 3926 | - npm wrapper first-run downloads now use process-unique temp files so concurrent `deepseek` / `deepseek-tui` invocations do not race on `*.download` files. |
| 3927 | |
| 3928 | ## [0.3.33] - 2026-04-11 |
| 3929 | |
| 3930 | ### Changed |
| 3931 | - Footer polish: simplified footer rendering, removed footer clock label, updated status line layout |
| 3932 | - Palette cleanup: removed `FOOTER_HINT` color constant |
| 3933 | |
| 3934 | ### Removed |
| 3935 | - `FOOTER_HINT` color constant from palette (use `TEXT_MUTED` or `TEXT_HINT` instead) |
| 3936 | |
| 3937 | ### Fixed |
| 3938 | - Test updates to align with simplified footer logic |
| 3939 | - Empty state placeholder text removed for cleaner UI |
| 3940 | |
| 3941 | ## [0.3.32] - 2026-04-11 |
| 3942 | |
| 3943 | ### Added |
| 3944 | - Finance tool: Yahoo Finance v8 quote endpoint with chart fallback, supporting stocks, ETFs, indices, forex, and crypto lookups. |
| 3945 | - Header widget redesign: proportional truncation, context-usage bar with gradient fill, streaming indicator, and graceful narrow-terminal degradation. |
| 3946 | - Expanded test coverage: 680+ tests including footer state, context spans, plan prompt lifecycle, workspace context refresh, header rendering, and finance tool integration tests with wiremock. |
| 3947 | - Workspace context refresh with configurable TTL and deferred initial fetch. |
| 3948 | - Config command additions for runtime settings management. |
| 3949 | |
| 3950 | ### Changed |
| 3951 | - Redesigned footer status strip with mode/model/status layout, context bar, and narrow-terminal fallback. |
| 3952 | - Plan prompt now uses numeric selection (1-4) instead of keyword input; old aliases are sent as regular messages. |
| 3953 | - Archived outdated docs (`workspace_migration_status.md` -> `docs/archive/`). |
| 3954 | - Trimmed AGENTS.md boilerplate and updated task counts. |
| 3955 | - Clarified release-surface documentation: crates.io publication may lag the workspace/npm wrapper. |
| 3956 | |
| 3957 | ### Fixed |
| 3958 | - Header `metadata_spans` now uses `saturating_sub` to prevent underflow on narrow terminals. |
| 3959 | - Finance tool reuses a single HTTP client instead of rebuilding per request. |
| 3960 | - Finance tool tests no longer leak temp directories. |
| 3961 | |
| 3962 | ## [0.3.31] - 2026-03-08 |
| 3963 | |
| 3964 | ### Added |
| 3965 | - Replaced the finance tool backend with Yahoo Finance v8 + CoinGecko fallback for reliable real-time market data (stocks, ETFs, indices, forex, crypto). |
| 3966 | - Added compaction UX: status strip shows animated COMPACTING indicator during context summarization, footer reflects compaction state, and CompactionCompleted events now include message count statistics. |
| 3967 | - Added send flash: brief tinted background highlight on the last user message after sending. |
| 3968 | - Added braille typing indicator with smooth 10-frame animation cycle. |
| 3969 | |
| 3970 | ### Changed |
| 3971 | - Redesigned the footer status strip with mode/model/token/cost layout, quadrant separators, and a context-usage bar. |
| 3972 | - Added Unicode prefix indicators (▸ You, ◆ Answer, ● System) to chat history cells for visual distinction. |
| 3973 | - Improved thinking token delineation with labeled delimiters in transcript rendering. |
| 3974 | - Refactored source code into workspace crates for better modularity and dependency management. |
| 3975 | |
| 3976 | ### Fixed |
| 3977 | - Fixed Plan mode ESC key dismissing the prompt without clearing `plan_prompt_pending`, which prevented the prompt from reappearing on subsequent plan completions. |
| 3978 | - Fixed clippy lint (collapsible_if) in web browsing session management. |
| 3979 | |
| 3980 | ## [0.3.30] - 2026-03-06 |
| 3981 | |
| 3982 | ### Added |
| 3983 | - Added a release-ready local npm smoke path that builds binaries, serves release assets locally, packs the wrapper, installs the tarball, and checks both entrypoints before publish. |
| 3984 | - Added an opt-in full-matrix local release-asset fixture so `npm run release:check` can be exercised before GitHub release assets exist. |
| 3985 | |
| 3986 | ### Changed |
| 3987 | - Bumped the Rust workspace crates and npm wrapper to `0.3.30`. |
| 3988 | - Pointed the npm wrapper's default `deepseekBinaryVersion` at `0.3.30` for the next coordinated Rust + npm release. |
| 3989 | - Updated the crates dry-run helper to work from a dirty workspace and to preflight dependent workspace crates without requiring unpublished versions to already exist on crates.io. |
| 3990 | |
| 3991 | ## [0.3.29] - 2026-03-03 |
| 3992 | |
| 3993 | ### Added |
| 3994 | - Added npm publish-time release asset verification for the `deepseek-tui` package to fail fast when expected GitHub binaries are missing. |
| 3995 | - Added checksum manifests to GitHub release assets and checksum verification in the npm installer. |
| 3996 | - Added `npm pack` install-and-smoke CI coverage for the `deepseek-tui` wrapper package. |
| 3997 | - Added an end-to-end release runbook covering crates.io, GitHub Releases, and npm publication. |
| 3998 | |
| 3999 | ### Changed |
| 4000 | - Updated npm package documentation for clearer install modes, environment overrides, and release integrity behavior. |
| 4001 | - Improved installer support-matrix error messaging for unsupported platform/architecture combinations. |
| 4002 | - Decoupled npm package version from default binary artifact version via `deepseekBinaryVersion`, enabling packaging-only npm releases. |
| 4003 | - Moved the `deepseek-tui` binary target inside `crates/tui` so `cargo publish --dry-run -p deepseek-tui` works from the workspace package layout. |
| 4004 | - Replaced the root-level crates publish workflow with an ordered workspace publish flow. |
| 4005 | - Reworked first-run onboarding and README copy around primary workflows instead of shortcut memorization. |
| 4006 | - Relaxed onboarding API-key format heuristics so unusual keys warn instead of blocking setup. |
| 4007 | |
| 4008 | ## [0.3.28] - 2026-03-02 |
| 4009 | |
| 4010 | ### Added |
| 4011 | - Converted the project to a modular Cargo workspace using a `crates/` layout. |
| 4012 | - Added new crate boundaries mirroring a deepseek architecture (`agent`, `config`, `core`, `execpolicy`, `hooks`, `mcp`, `protocol`, `state`, `tools`, `tui-core`, `tui`, and `app-server`). |
| 4013 | |
| 4014 | ### Changed |
| 4015 | - Added parity CI coverage with protocol/state/snapshot checks. |
| 4016 | - Updated release workflow to build both `deepseek` and `deepseek-tui` binaries. |
| 4017 | |
| 4018 | ## [0.3.26] - 2026-03-02 |
| 4019 | |
| 4020 | ### Fixed |
| 4021 | - Resolved SSE stream corruption caused by byte/string position mismatch in streaming parse flow. |
| 4022 | - Hardened base URL validation to reject non-HTTP/HTTPS schemes. |
| 4023 | - Prevented multi-byte UTF-8 truncation panics in common-prefix and runtime thread summary paths. |
| 4024 | - Corrected context usage alert thresholds by separating warning and critical trigger levels. |
| 4025 | |
| 4026 | ### Changed |
| 4027 | - Removed non-code utility tools from the runtime tool registry (`calculator`, `weather`, `sports`, `finance`, `time`) and related wiring. |
| 4028 | - Consolidated duplicate URL encoding helpers by delegating to shared `crate::utils::url_encode`. |
| 4029 | - Replaced broad crate-level lint suppressions with targeted `#[allow(...)]` annotations where justified. |
| 4030 | - Cleaned up dead APIs, unused struct fields, unused builder helpers, and non-integrated modules. |
| 4031 | - Addressed clippy findings across the codebase (collapsible conditionals, defaults, indexing helpers, and API signature cleanup). |
| 4032 | |
| 4033 | ## [0.3.24] - 2026-02-25 |
| 4034 | |
| 4035 | ### Fixed |
| 4036 | - Preserve reasoning-only assistant turns for DeepSeek reasoning models (`deepseek-reasoner`, R-series markers) when rebuilding chat history. |
| 4037 | - Align SSE tool streaming indices so each tool block start/delta/stop uses the same block index. |
| 4038 | - Prevent transcript auto-scroll-to-bottom when a non-empty transcript selection is active. |
| 4039 | - Allow session picker search mode to accept the current selection with a single `Enter` press. |
| 4040 | - Preserve tool output whitespace/indentation while still wrapping long unbroken tokens. |
| 4041 | - Make transcript selection copy/highlighting display-width aware (wide chars and tabs). |
| 4042 | - Gate execpolicy behavior on the `exec_policy` feature flag across CLI/tool execution paths. |
| 4043 | - Run doctor API connectivity checks using the effective loaded config/profile (instead of reloading defaults). |
| 4044 | - Parse DeepSeek model context-window suffix hints such as `-32k` and `-256k`. |
| 4045 | - Update README config docs with key environment overrides and a direct link to full configuration docs. |
| 4046 | |
| 4047 | ## [0.3.23] - 2026-02-24 |
| 4048 | |
| 4049 | ### Changed |
| 4050 | - Updated project copy to describe the app as a terminal-native TUI/CLI for DeepSeek models (not pinned to a specific model generation). |
| 4051 | |
| 4052 | ### Fixed |
| 4053 | - Model selection and config validation now accept any valid `deepseek-*` model ID (including future releases), while still normalizing common aliases like `deepseek-v3.2` and `deepseek-r1`. |
| 4054 | - Tool-call recovery now auto-loads deferred tools when the model requests them directly, instead of failing with manual `tool_search_*` instructions. |
| 4055 | - YOLO mode now preloads tools by default (including deferred MCP tools), so model tool calls can run immediately without discovery indirection. |
| 4056 | - Unknown tool-call failures now include discovery guidance and nearest tool-name suggestions instead of generic availability errors. |
| 4057 | - Slash-command errors now suggest the closest known command (for example `/modle` -> `/model`) instead of only returning a generic unknown-command message. |
| 4058 | |
| 4059 | ## [0.3.22] - 2026-02-19 |
| 4060 | |
| 4061 | ### Added |
| 4062 | - Interactive `/config` editing modal for runtime settings updates. |
| 4063 | |
| 4064 | ### Changed |
| 4065 | - Retired user-facing `/set` command path (no longer reachable/discoverable). |
| 4066 | - Replaced `/deepseek` command behavior with `/links` (aliases: `dashboard`, `api`). |
| 4067 | |
| 4068 | ### Fixed |
| 4069 | - Legacy `/set` and `/deepseek` inputs now return migration guidance instead of generic unknown-command errors. |
| 4070 | |
| 4071 | ## [0.3.21] - 2026-02-19 |
| 4072 | |
| 4073 | ### Added |
| 4074 | - Parallel tool execution in `multi_tool_use.parallel` for independent task workflows. |
| 4075 | - Session resume-thread coverage in tests. |
| 4076 | |
| 4077 | ### Changed |
| 4078 | - Desktop and web parity polish across the TUI and runtime surfaces. |
| 4079 | - Onboarding and approval UX refinement from prior phase 3 iteration. |
| 4080 | |
| 4081 | ### Fixed |
| 4082 | - Runtime pre-release startup issues and config-path edge cases. |
| 4083 | - Clippy lint regressions introduced by the last parity pass. |
| 4084 | |
| 4085 | ### Security/Hardening |
| 4086 | - General pre-release hardening for runtime app behavior. |
| 4087 | |
| 4088 | ## [0.3.17] - 2026-02-16 |
| 4089 | |
| 4090 | ### Fixed |
| 4091 | - Config loading now expands `~` in `DEEPSEEK_CONFIG_PATH` and `--config` paths. |
| 4092 | - When `DEEPSEEK_CONFIG_PATH` points to a missing file, config loading now falls back to `~/.deepseek/config.toml` if it exists. |
| 4093 | |
| 4094 | ### Changed |
| 4095 | - Removed committed transient runtime artifacts (`session_*.json`, `.deepseek/trusted`) and added ignore rules to prevent re-commit. |
| 4096 | |
| 4097 | ## [0.3.16] - 2026-02-15 |
| 4098 | |
| 4099 | ### Added |
| 4100 | - `deepseek models` CLI command to fetch and list models from the configured `/v1/models` endpoint (with `--json` output mode). |
| 4101 | - `/models` slash command to fetch and display live model IDs in the TUI. |
| 4102 | - Slash-command autocomplete hints in the composer plus `Tab` completion for `/` commands. |
| 4103 | - Command palette modal (`Ctrl+K`) for quick insertion of slash commands and skills. |
| 4104 | - Persistent right sidebar in wide terminals showing live plan/todo/sub-agent state. |
| 4105 | - Expandable tool payload views (`v` in transcript, `v` in approval modal) for full params/output inspection. |
| 4106 | - Runtime HTTP/SSE API (`deepseek serve --http`) with durable thread/turn/item lifecycle, interrupt/steer, and replayable event timeline. |
| 4107 | - Background task queue (`/task add|list|show|cancel` and `POST /v1/tasks`) with persistent storage, bounded worker pool, and timeline/artifact tracking. |
| 4108 | |
| 4109 | ### Changed |
| 4110 | - Centralized the default text model (`DEFAULT_TEXT_MODEL`) and shared common model list to reduce drift across runtime/config paths. |
| 4111 | - `/model` now clarifies that any valid DeepSeek model ID is accepted (including future releases), while still showing common model IDs. |
| 4112 | |
| 4113 | ### Fixed |
| 4114 | - Expanded reasoning-model detection for chat history reconstruction (supports R-series and reasoner-style naming without hardcoding single versions). |
| 4115 | - Aligned docs/config examples with the then-current runtime default model. |
| 4116 | |
| 4117 | ## [0.3.14] - 2026-02-05 |
| 4118 | |
| 4119 | ### Added |
| 4120 | - `web.run` now supports `image_query` (DuckDuckGo image search) |
| 4121 | - `multi_tool_use.parallel` now supports safe MCP meta tools (`list_mcp_resources`, `mcp_read_resource`, etc.) |
| 4122 | |
| 4123 | ### Fixed |
| 4124 | - Encode tool-call function names when rebuilding Chat Completions history (keeps dotted tool names API-safe) |
| 4125 | |
| 4126 | ### Changed |
| 4127 | - Prompts: stronger `web.run` citation placement and quote-limit guidance |
| 4128 | |
| 4129 | ## [0.3.13] - 2026-02-04 |
| 4130 | |
| 4131 | ### Fixed |
| 4132 | - Restore an in-app scrollbar for the transcript view |
| 4133 | |
| 4134 | ## [0.3.12] - 2026-02-04 |
| 4135 | |
| 4136 | ### Fixed |
| 4137 | - Map dotted tool names to API-safe identifiers for DeepSeek tool calls |
| 4138 | - Encode any invalid tool names for API tool lists while preserving internal names |
| 4139 | |
| 4140 | ## [0.3.11] - 2026-02-04 |
| 4141 | |
| 4142 | ### Fixed |
| 4143 | - Fix tool name mapping for DeepSeek API |
| 4144 | |
| 4145 | ## [0.3.10] - 2026-02-04 |
| 4146 | |
| 4147 | ### Fixed |
| 4148 | - Always enable mouse wheel scrolling in the TUI (even without alt screen) |
| 4149 | |
| 4150 | ## [0.3.9] - 2026-02-04 |
| 4151 | |
| 4152 | ### Removed |
| 4153 | - RLM mode, tools, and documentation pending a faithful implementation of the MIT RLM design |
| 4154 | - Duo mode tools and prompts pending a citable research spec |
| 4155 | |
| 4156 | ### Fixed |
| 4157 | - Footer context usage bar remains visible while status toasts are shown |
| 4158 | |
| 4159 | ### Changed |
| 4160 | - Updated prompts and docs to reflect the simplified mode/tool surface |
| 4161 | |
| 4162 | ## [0.3.8] - 2026-02-03 |
| 4163 | |
| 4164 | ### Fixed |
| 4165 | - Resolve clippy warnings (CI `-D warnings`) in new tool implementations |
| 4166 | |
| 4167 | ## [0.3.7] - 2026-02-03 |
| 4168 | |
| 4169 | ### Added |
| 4170 | - Tooling parity updates: `weather`, `finance`, `sports`, `time`, `calculator`, `request_user_input`, `multi_tool_use.parallel`, `web.run` |
| 4171 | - Shell streaming helpers: `exec_shell_wait` and `exec_shell_interact` |
| 4172 | - Sub-agent controls: `send_input` and `wait` (with aliases) |
| 4173 | - MCP resource helpers: `list_mcp_resources`, `list_mcp_resource_templates`, and `read_mcp_resource` alias |
| 4174 | |
| 4175 | ### Changed |
| 4176 | - Skills directory selection now prefers workspace `.agents/skills`, then `./skills`, then global |
| 4177 | - Docs and prompts updated to reflect new tool surface and parity notes |
| 4178 | |
| 4179 | ## [0.3.6] - 2026-02-02 |
| 4180 | |
| 4181 | ### Added |
| 4182 | - New welcome banner on startup showing "Welcome to DeepSeek TUI!" with directory, session ID, and model info |
| 4183 | - Visual context progress bar in footer showing usage with block characters [████░░░░░░] and percentage |
| 4184 | |
| 4185 | ### Changed |
| 4186 | - Removed custom block-character scrollbar from chat area - now uses terminal's native scroll |
| 4187 | - Simplified header bar: removed context percentage indicator (moved to footer as progress bar) |
| 4188 | |
| 4189 | ## [0.3.5] - 2026-01-30 |
| 4190 | |
| 4191 | ### Added |
| 4192 | - Intelligent context offloading: large tool results (>15k chars) are automatically moved to RLM memory to preserve the context window |
| 4193 | - Persistent history context: compacted messages are offloaded to RLM `history` variable for recall |
| 4194 | - Full MCP protocol support: SSE transport, Resources (`resources/list`, `resources/read`), and Prompts (`prompts/list`, `prompts/get`) |
| 4195 | - `mcp_read_resource` and `mcp_get_prompt` virtual tools exposed to the model |
| 4196 | - Dialectical Duo mode with specialized TUI rendering (`Player` / `Coach` history cells) |
| 4197 | - Dynamic system prompt refreshing at each turn for up-to-date RLM/Duo/working-set context |
| 4198 | - `project_map` tool for automatic codebase structure discovery |
| 4199 | - `delegate_to_agent` alias for streamlined sub-agent delegation |
| 4200 | |
| 4201 | ### Changed |
| 4202 | - Default theme changed to 'Whale' with updated color palette |
| 4203 | - `with_agent_tools` now includes `project_map`, `test_runner`, and conditionally RLM tools for all agent modes |
| 4204 | - MCP `McpServerConfig.command` is now `Option<String>` to support URL-only (SSE) servers |
| 4205 | |
| 4206 | ### Fixed |
| 4207 | - MCP test compilation errors for updated `McpServerConfig` struct shape |
| 4208 | |
| 4209 | ## [0.3.4] - 2026-01-29 |
| 4210 | |
| 4211 | ### Changed |
| 4212 | - Updated Cargo.lock dependencies |
| 4213 | |
| 4214 | ### Fixed |
| 4215 | - Compaction tool-call pairing: enforce bidirectional tool-call/tool-result integrity with fixpoint convergence |
| 4216 | - Safety net scanning to drop orphan tool results in the request builder |
| 4217 | - Double-dispatch race in parallel tool execution |
| 4218 | |
| 4219 | ## [0.3.3] - 2026-01-28 |
| 4220 | |
| 4221 | ### Added |
| 4222 | - TUI polish: Kimi-style footer with mode/model/token display |
| 4223 | - Streaming thinking blocks with dedicated rendering |
| 4224 | - Loading animation improvements |
| 4225 | |
| 4226 | ## [0.3.2] - 2026-01-28 |
| 4227 | |
| 4228 | ### Fixed |
| 4229 | - Preserve tool-call + tool-result pairing during compaction to avoid invalid tool message sequences |
| 4230 | - Drop orphan tool results in request builder as a safety net to prevent API 400s |
| 4231 | |
| 4232 | ## [0.3.1] - 2026-01-27 |
| 4233 | |
| 4234 | ### Added |
| 4235 | - `deepseek setup` to bootstrap MCP config and skills directories |
| 4236 | - `deepseek mcp init` to generate a template `mcp.json` at the configured path |
| 4237 | |
| 4238 | ### Changed |
| 4239 | - `deepseek doctor` now follows the resolved config path and config-derived MCP/skills locations |
| 4240 | |
| 4241 | ### Fixed |
| 4242 | - Doctor no longer reports missing MCP/skills when paths are overridden via config or env |
| 4243 | |
| 4244 | ## [0.3.0] - 2026-01-27 |
| 4245 | |
| 4246 | ### Added |
| 4247 | - Repo-aware working set tracking with prompt injection for active paths |
| 4248 | - Working set signals now pin relevant messages during auto-compaction |
| 4249 | - Offline eval harness (`deepseek eval`) with CI coverage in the test job |
| 4250 | - Shell tool now emits stdout/stderr summaries and truncation metadata |
| 4251 | - Dependency-aware `agent_swarm` tool for orchestrating multiple sub-agents |
| 4252 | - Expanded sub-agent tool access (apply_patch, web_search, file_search) |
| 4253 | |
| 4254 | ### Changed |
| 4255 | - Auto-compaction now accounts for pinned budget and preserves working-set context |
| 4256 | - Apply patch tool validates patch shape, reports per-file summaries, and improves hunk mismatch diagnostics |
| 4257 | - Eval harness shell step now uses a Windows-safe default command |
| 4258 | - Increased `max_subagents` clamp to `1..=20` |
| 4259 | |
| 4260 | ## [0.2.2] - 2026-01-22 |
| 4261 | |
| 4262 | ### Fixed |
| 4263 | - Session save no longer panics on serialization errors |
| 4264 | - Web search regex patterns are now cached for better performance |
| 4265 | - Improved panic messages for regex compilation failures |
| 4266 | |
| 4267 | ## [0.2.1] - 2026-01-22 |
| 4268 | |
| 4269 | ### Fixed |
| 4270 | - Resolve clippy warnings for Rust 1.92 |
| 4271 | |
| 4272 | ## [0.2.0] - 2026-01-20 |
| 4273 | |
| 4274 | ### Changed |
| 4275 | - Removed npm package distribution; now Cargo-only |
| 4276 | - Clean up for public release |
| 4277 | |
| 4278 | ### Fixed |
| 4279 | - Disabled automatic RLM mode switching; use /rlm or /aleph to enter RLM mode |
| 4280 | - Fixed cargo fmt formatting issues |
| 4281 | |
| 4282 | ## [0.0.2] - 2026-01-20 |
| 4283 | |
| 4284 | ### Fixed |
| 4285 | - Disabled automatic RLM mode switching; use /rlm or /aleph to enter RLM mode. |
| 4286 | |
| 4287 | ## [0.0.1] - 2026-01-19 |
| 4288 | |
| 4289 | ### Added |
| 4290 | - DeepSeek Responses API client with chat-completions fallback |
| 4291 | - CLI parity commands: login/logout, exec, review, apply, mcp, sandbox |
| 4292 | - Resume/fork session workflows with picker fallback |
| 4293 | - DeepSeek blue branding refresh + whale indicator |
| 4294 | - Responses API proxy subcommand for key-isolated forwarding |
| 4295 | - Execpolicy check tooling and feature flag CLI |
| 4296 | - Agentic exec mode (`deepseek exec --auto`) with auto-approvals |
| 4297 | |
| 4298 | ### Changed |
| 4299 | - Removed multimedia tooling and aligned prompts/docs for text-only DeepSeek API |
| 4300 | |
| 4301 | ## [0.1.9] - 2026-01-17 |
| 4302 | |
| 4303 | ### Added |
| 4304 | - API connectivity test in `deepseek doctor` command |
| 4305 | - Helpful error diagnostics for common API failures (invalid key, timeout, network issues) |
| 4306 | |
| 4307 | ## [0.1.8] - 2026-01-16 |
| 4308 | |
| 4309 | ### Added |
| 4310 | - Renderable widget abstraction and modal view stack for TUI composition |
| 4311 | - Parallel tool execution with lock-aware scheduling |
| 4312 | - Interactive shell mode with terminal pause/resume handling |
| 4313 | |
| 4314 | ### Changed |
| 4315 | - Tool approval requirements moved into tool specs |
| 4316 | - Tool results are recorded in original request order |
| 4317 | |
| 4318 | ## [0.1.7] - 2026-01-15 |
| 4319 | |
| 4320 | ### Added |
| 4321 | - Duo mode (player-coach autocoding workflow) |
| 4322 | - Character-level transcript selection |
| 4323 | |
| 4324 | ### Fixed |
| 4325 | - Approval flow tool use ID routing |
| 4326 | - Cursor position sync for transcript selection |
| 4327 | |
| 4328 | ## [0.1.6] - 2026-01-14 |
| 4329 | |
| 4330 | ### Added |
| 4331 | - Auto-RLM for large pasted blocks with context auto-load |
| 4332 | - `chunk_auto` and `rlm_query` `auto_chunks` for quick document sweeps |
| 4333 | - RLM usage badge with budget warnings in the footer |
| 4334 | |
| 4335 | ### Changed |
| 4336 | - Auto-RLM now honors explicit RLM file requests even for smaller files |
| 4337 | |
| 4338 | ## [0.1.5] - 2026-01-14 |
| 4339 | |
| 4340 | ### Added |
| 4341 | - RLM prompt with external-context guidance and REPL tooling |
| 4342 | - RLM tools for context loading, execution, status, and sub-queries (rlm_load, rlm_exec, rlm_status, rlm_query) |
| 4343 | - RLM query usage tracking and variable buffers |
| 4344 | - Workspace-relative `@path` support for RLM loads |
| 4345 | - Auto-switch to RLM when users request large file analysis (or the largest file) |
| 4346 | |
| 4347 | ### Changed |
| 4348 | - Removed Edit mode; RLM chat is default with /repl toggle |
| 4349 | |
| 4350 | ## [0.1.0] - 2026-01-12 |
| 4351 | |
| 4352 | ### Added |
| 4353 | - Initial alpha release of DeepSeek TUI |
| 4354 | - Interactive TUI chat interface |
| 4355 | - DeepSeek API integration (OpenAI-compatible Responses API) |
| 4356 | - Tool execution (shell, file ops) |
| 4357 | - MCP (Model Context Protocol) support |
| 4358 | - Session management with history |
| 4359 | - Skills/plugin system |
| 4360 | - Cost tracking and estimation |
| 4361 | - Hooks system and config profiles |
| 4362 | - Example skills and launch assets |
| 4363 | |
| 4364 | [0.8.39]: https://github.com/Hmbown/CodeWhale/compare/v0.8.38...v0.8.39 |
| 4365 | [0.8.38]: https://github.com/Hmbown/CodeWhale/compare/v0.8.37...v0.8.38 |
| 4366 | [0.8.37]: https://github.com/Hmbown/CodeWhale/compare/v0.8.36...v0.8.37 |
| 4367 | [0.8.36]: https://github.com/Hmbown/CodeWhale/compare/v0.8.35...v0.8.36 |
| 4368 | [0.8.35]: https://github.com/Hmbown/CodeWhale/compare/v0.8.34...v0.8.35 |
| 4369 | [0.8.34]: https://github.com/Hmbown/CodeWhale/compare/v0.8.33...v0.8.34 |
| 4370 | [0.8.33]: https://github.com/Hmbown/CodeWhale/compare/v0.8.32...v0.8.33 |
| 4371 | [0.8.32]: https://github.com/Hmbown/CodeWhale/compare/v0.8.31...v0.8.32 |
| 4372 | [0.8.31]: https://github.com/Hmbown/CodeWhale/compare/v0.8.30...v0.8.31 |
| 4373 | [0.8.30]: https://github.com/Hmbown/CodeWhale/compare/v0.8.29...v0.8.30 |
| 4374 | [0.8.29]: https://github.com/Hmbown/CodeWhale/compare/v0.8.28...v0.8.29 |
| 4375 | [0.8.28]: https://github.com/Hmbown/CodeWhale/compare/v0.8.27...v0.8.28 |
| 4376 | [0.8.27]: https://github.com/Hmbown/CodeWhale/compare/v0.8.26...v0.8.27 |
| 4377 | [0.8.26]: https://github.com/Hmbown/CodeWhale/compare/v0.8.25...v0.8.26 |
| 4378 | [0.8.25]: https://github.com/Hmbown/CodeWhale/compare/v0.8.24...v0.8.25 |
| 4379 | [0.8.24]: https://github.com/Hmbown/CodeWhale/compare/v0.8.23...v0.8.24 |
| 4380 | [0.8.23]: https://github.com/Hmbown/CodeWhale/compare/v0.8.22...v0.8.23 |
| 4381 | [0.8.22]: https://github.com/Hmbown/CodeWhale/compare/v0.8.21...v0.8.22 |
| 4382 | [0.8.21]: https://github.com/Hmbown/CodeWhale/compare/v0.8.20...v0.8.21 |
| 4383 | [0.8.20]: https://github.com/Hmbown/CodeWhale/compare/v0.8.19...v0.8.20 |
| 4384 | [0.8.19]: https://github.com/Hmbown/CodeWhale/compare/v0.8.18...v0.8.19 |
| 4385 | [0.8.18]: https://github.com/Hmbown/CodeWhale/compare/v0.8.17...v0.8.18 |
| 4386 | [0.8.17]: https://github.com/Hmbown/CodeWhale/compare/v0.8.16...v0.8.17 |
| 4387 | [0.8.16]: https://github.com/Hmbown/CodeWhale/compare/v0.8.15...v0.8.16 |
| 4388 | [0.8.15]: https://github.com/Hmbown/CodeWhale/compare/v0.8.13...v0.8.15 |
| 4389 | [0.8.13]: https://github.com/Hmbown/CodeWhale/compare/v0.8.12...v0.8.13 |
| 4390 | [0.8.12]: https://github.com/Hmbown/CodeWhale/compare/v0.8.11...v0.8.12 |
| 4391 | [0.8.11]: https://github.com/Hmbown/CodeWhale/compare/v0.8.10...v0.8.11 |
| 4392 | [0.8.10]: https://github.com/Hmbown/CodeWhale/compare/v0.8.8...v0.8.10 |
| 4393 | [0.8.8]: https://github.com/Hmbown/CodeWhale/compare/v0.8.7...v0.8.8 |
| 4394 | [0.8.7]: https://github.com/Hmbown/CodeWhale/compare/v0.8.6...v0.8.7 |
| 4395 | [0.8.6]: https://github.com/Hmbown/CodeWhale/compare/v0.8.5...v0.8.6 |
| 4396 | [0.8.5]: https://github.com/Hmbown/CodeWhale/compare/v0.8.4...v0.8.5 |
| 4397 | [0.8.4]: https://github.com/Hmbown/CodeWhale/compare/v0.8.3...v0.8.4 |
| 4398 | [0.8.3]: https://github.com/Hmbown/CodeWhale/compare/v0.8.2...v0.8.3 |
| 4399 | [0.8.2]: https://github.com/Hmbown/CodeWhale/compare/v0.8.1...v0.8.2 |
| 4400 | [0.8.1]: https://github.com/Hmbown/CodeWhale/compare/v0.8.0...v0.8.1 |
| 4401 | [0.8.0]: https://github.com/Hmbown/CodeWhale/compare/v0.7.9...v0.8.0 |
| 4402 | [0.7.9]: https://github.com/Hmbown/CodeWhale/compare/v0.7.8...v0.7.9 |
| 4403 | [0.7.8]: https://github.com/Hmbown/CodeWhale/compare/v0.7.7...v0.7.8 |
| 4404 | [0.7.7]: https://github.com/Hmbown/CodeWhale/compare/v0.7.6...v0.7.7 |
| 4405 | [0.7.6]: https://github.com/Hmbown/CodeWhale/compare/v0.7.5...v0.7.6 |
| 4406 | [0.6.1]: https://github.com/Hmbown/CodeWhale/compare/v0.6.0...v0.6.1 |
| 4407 | [0.6.0]: https://github.com/Hmbown/CodeWhale/compare/v0.4.9...v0.6.0 |
| 4408 | [0.4.9]: https://github.com/Hmbown/CodeWhale/compare/v0.4.8...v0.4.9 |
| 4409 | [0.4.8]: https://github.com/Hmbown/CodeWhale/compare/v0.3.33...v0.4.8 |
| 4410 | [0.3.33]: https://github.com/Hmbown/CodeWhale/compare/v0.3.32...v0.3.33 |
| 4411 | [0.3.32]: https://github.com/Hmbown/CodeWhale/compare/v0.3.31...v0.3.32 |
| 4412 | [0.3.31]: https://github.com/Hmbown/CodeWhale/compare/v0.3.28...v0.3.31 |
| 4413 | [0.3.28]: https://github.com/Hmbown/CodeWhale/compare/v0.3.27...v0.3.28 |
| 4414 | [0.3.23]: https://github.com/Hmbown/CodeWhale/compare/v0.3.22...v0.3.23 |
| 4415 | [0.3.22]: https://github.com/Hmbown/CodeWhale/compare/v0.3.21...v0.3.22 |
| 4416 | [0.3.21]: https://github.com/Hmbown/CodeWhale/compare/v0.3.17...v0.3.21 |
| 4417 | [0.3.17]: https://github.com/Hmbown/CodeWhale/compare/v0.3.16...v0.3.17 |
| 4418 | [0.3.16]: https://github.com/Hmbown/CodeWhale/compare/v0.3.14...v0.3.16 |
| 4419 | [0.3.14]: https://github.com/Hmbown/CodeWhale/compare/v0.3.13...v0.3.14 |
| 4420 | [0.3.13]: https://github.com/Hmbown/CodeWhale/compare/v0.3.12...v0.3.13 |
| 4421 | [0.3.12]: https://github.com/Hmbown/CodeWhale/compare/v0.3.11...v0.3.12 |
| 4422 | [0.3.11]: https://github.com/Hmbown/CodeWhale/compare/v0.3.10...v0.3.11 |
| 4423 | [0.3.10]: https://github.com/Hmbown/CodeWhale/compare/v0.3.6...v0.3.10 |
| 4424 | [0.3.6]: https://github.com/Hmbown/CodeWhale/compare/v0.3.5...v0.3.6 |
| 4425 | [0.3.5]: https://github.com/Hmbown/CodeWhale/compare/v0.3.4...v0.3.5 |
| 4426 | [0.3.4]: https://github.com/Hmbown/CodeWhale/compare/v0.3.3...v0.3.4 |
| 4427 | [0.3.3]: https://github.com/Hmbown/CodeWhale/compare/v0.3.2...v0.3.3 |
| 4428 | [0.3.2]: https://github.com/Hmbown/CodeWhale/compare/v0.3.1...v0.3.2 |
| 4429 | [0.3.1]: https://github.com/Hmbown/CodeWhale/compare/v0.3.0...v0.3.1 |
| 4430 | [0.3.0]: https://github.com/Hmbown/CodeWhale/compare/v0.2.2...v0.3.0 |
| 4431 | [0.2.2]: https://github.com/Hmbown/CodeWhale/compare/v0.2.0...v0.2.2 |
| 4432 | [0.2.0]: https://github.com/Hmbown/CodeWhale/releases/tag/v0.2.0 |
| 4433 | [0.0.2]: https://github.com/Hmbown/CodeWhale/releases/tag/v0.0.2 |
| 4434 | [0.0.1]: https://github.com/Hmbown/CodeWhale/releases/tag/v0.0.1 |
| 4435 | [0.1.9]: https://github.com/Hmbown/CodeWhale/compare/v0.1.8...v0.1.9 |
| 4436 | [0.1.8]: https://github.com/Hmbown/CodeWhale/compare/v0.1.7...v0.1.8 |
| 4437 | [0.1.7]: https://github.com/Hmbown/CodeWhale/compare/v0.1.6...v0.1.7 |
| 4438 | [0.1.6]: https://github.com/Hmbown/CodeWhale/compare/v0.1.5...v0.1.6 |
| 4439 | [0.1.5]: https://github.com/Hmbown/CodeWhale/compare/v0.1.0...v0.1.5 |
| 4440 | [0.1.0]: https://github.com/Hmbown/CodeWhale/releases/tag/v0.1.0 |
| 4441 |