返回 DeepSeek-Reasonix
CHANGELOG.md
根目录 / CHANGELOG.md
1 # Changelog
2
3 All notable changes to the Go line (Reasonix 1.0+) are recorded here. The legacy
4 `0.x` TypeScript history lives on the [`v1`](https://github.com/esengine/DeepSeek-Reasonix/tree/v1)
5 branch.
6
7 ## Unreleased
8
9 ### Added
10
11 - **Live file observations:** structured file tools now protect mutations with
12 a host-owned current-version observation. Any successful text window is
13 sufficient, successful writes refresh the version, and external changes
14 produce `FS_STALE_VERSION` without blocking unrelated tools.
15
16 - **MCP 2026-07-28 protocol:** multi-round-trip form/URL elicitation across
17 Desktop, CLI TUI, and serve; headless entries stay on the core surface and
18 cancel unanswered requests instead of guessing.
19 - **MCP Apps 2026-01-26 (Desktop):** inline app surfaces in tool cards behind
20 a per-server double-iframe sandbox, app-tool visibility metadata, bounded
21 aggregate local presentations, tab-bound AppBridge routing and teardown,
22 immutable digest-bound resource snapshots, and confirmed external links;
23 local rich results, instance-gated app tool calls, and the four-layer
24 capability matrix in MCP status.
25 - **Profile-scoped MCP schema caches:** capability-declaring hosts keep their
26 own `v3` cache files so catalogs negotiated under different client
27 capabilities never cross-read.
28
29 ### Changed
30
31 - **Persistent bash PTY:** ordinary foreground `bash` calls in a session now
32 share one PTY, so `cd`, exported variables, and shell functions survive
33 across calls. Output stays byte-identical to one-shot execution, stdin stays
34 detached, and a timeout or cancel reports partial output and says the shell
35 was reset. Background jobs, commands that background a child, per-call
36 write-root escalations, host terminals, and PowerShell hosts stay one-shot.
37 The bash tool schema and description are unchanged.
38
39 - **CLI YOLO shortcuts:** the CLI displays the unrestricted preset as YOLO;
40 `Ctrl+Y` toggles the canonical `danger-full-access` permission, while
41 `Shift+Tab` cycles Read only → Workspace write → YOLO → Plan.
42
43 - **Harness-style scheduling and recovery:** calls take effect in execution
44 order, including same-batch read/edit sequences. Bounded reads create no
45 completion debt. Unknown external effects are durable advisory facts and no
46 longer block tools or trigger replay. Proof/settlement tools, Auto Guard,
47 recovery actions, and repeat-call rejection are retired; identical calls get
48 non-blocking reminders at counts 3, 5, and 8.
49
50 - **Fact-driven execution:** Ordinary requests always enter the executor.
51 There is no automatic simple / light / full task mode and no per-turn
52 `TaskPolicy` classification. The planner runs only for an explicit Plan,
53 an approval boundary, or Goal start. The host builds verification
54 obligations from concrete tool effects and receipts. Plan, Goal,
55 permission, and sandbox stay independent. Tool schemas and the executor
56 system prefix stay byte-stable. Historical `<execution-policy>` tags remain
57 readable on old sessions and are stripped from new provider context.
58 Old `--preset`/`--profile` compatibility no-ops are unchanged.
59
60 - **Remote connect wizard host picker:** Step 1's host field now opens the
61 saved SSH connections through an explicit chevron dropdown on the input's
62 right edge instead of the old focus-triggered popup. The dropdown lists
63 every saved connection unfiltered, appends non-standard ports to each row,
64 leads with a "saved SSH connections" caption, and closes on pick, arrow
65 toggle, Escape (before the Escape that exits the wizard), or an outside
66 pointer press. The arrow is hidden while no hosts are saved and disabled
67 while a connection is busy.
68
69 ### Fixed
70
71 - **Read evidence recovery:** partial reads no longer freeze independent work
72 or ordinary final answers. Explicit full reads retain bounded completion
73 checks. Rejected edits track operation/version requirements so successful
74 retries, fresh versions and confirmed deletion retire obsolete blocks.
75 - **File and shell boundaries:** guard raced creates/overwrites and move-source
76 changes, recognize `git --no-pager` inspections, and retain structured recovery
77 diagnostics without changing provider tool schemas.
78
79 - **Relay image input:** ID-only or invalid model metadata now stays unknown.
80 Both Desktop model editors expose per-model Auto / On / Off overrides, with
81 official protocol limits retained. A separate V2 discovery cache rejects stale
82 results; saved settings and runtime image serialization share one resolver and
83 apply at Controller rebuild boundaries. Legacy configuration remains readable.
84 - **中转站图片输入:** 缺失或无效的模型能力显示“图片能力未识别”,两个编辑入口
85 均可逐模型选择“自动 / 开启 / 关闭”。独立 V2 缓存隔离旧错误声明并防止陈旧结果
86 覆盖;保存设置与实际图片请求统一解析,在 Controller 重建边界生效,兼容旧配置。
87
88 - **Deterministic natural-turn completion:** removed the extra completion
89 validator model request. Clean model stops now finish from provider/tool state;
90 true zero-content responses retry the frozen request at the Agent step
91 boundary, while explicit host-owned readiness and safety gates remain active.
92 Legacy completion-validator configuration and `completion_uncertain` event
93 values remain readable for compatibility but are no longer produced by the
94 validator path.
95
96 - **serve Host-header allowlist:** `reasonix serve` now rejects requests whose
97 `Host` is neither loopback nor the actual listen address (HTTP 421), closing
98 the DNS-rebinding bypass of the JSON content-type CSRF guard — a rebind page
99 becomes same-origin with the loopback listener and could previously drive
100 `/bypass`, `/submit`, and read `/history`. `behind_proxy` deployments and
101 wildcard/non-loopback binds are exempt. The non-loopback plaintext-HTTP
102 startup warning now also fires — loudest — for the unauthenticated `auth =
103 none` case that used to stay silent.
104
105 - **Preview read confinement:** `write_file` / `edit_file` / `multi_edit`
106 previews now apply the same `confinePreview` boundary as `delete_range` /
107 `delete_symbol`. A model-supplied absolute path outside the workspace roots
108 previously read the file (rendering its contents into the approval card and
109 session log) even though Execute would refuse the write.
110
111 - **Clean-filter hardening on internal diffs:** gitcmd diff invocations now
112 neutralize every `filter.<driver>` defined in the repository's local
113 `.git/config` (`clean=` emptied, `required` forced off), so viewing a changed
114 file's diff can no longer execute a repository-configured clean filter via
115 `.gitattributes`. Emptied filters are identity pass-throughs: the diff still
116 renders the real working-tree change.
117
118 - **install_source proxy SSRF parity:** the install_source SSRF dial guard now
119 also validates the request destination (IP literals) at the RoundTripper
120 boundary, so a configured HTTP/HTTPS proxy can no longer forward a blocked
121 target (cloud metadata, RFC1918, link-local, CGNAT) that the dial-time check
122 never sees — matching web_fetch's proxy-path behavior.
123
124 - **awk approval classification:** the bash indirect-execution classifier now
125 treats `awk`/`gawk`/`mawk`/`nawk` with an inline program (anything not read
126 via `-f`/`--file`) like `python -c`: it always requires human approval and
127 can never be covered by a remembered reusable prefix rule. `awk
128 'BEGIN{system("…")}'` previously fell through to the reusable class.
129
130 - **cargo check/doc read-only correction:** the legacy read-only command table
131 no longer lists `cargo check` / `cargo doc` as permission readers — cargo
132 executes the crate's `build.rs` for both. The effect classifier already
133 billed them as code-executing writers; the stale table entry (and its test)
134 now agree. Only `cargo search` remains read-only.
135
136 - **Compact MCP discovery:** `use_capability(action=list)` now returns one
137 compact summary per configured MCP server instead of expanding every cached
138 tool description, including tools from disabled servers. Inspecting one
139 enabled `mcp-server:<name>` still returns its live or cached directory
140 without starting it, while direct known-ID calls, routing, authorization,
141 and the fixed provider-visible tool schema remain unchanged.
142
143 - **Project MCP session reliability:** The MCP client now uses the official Go
144 SDK for stdio, legacy SSE, and Streamable HTTP while retaining Reasonix's
145 existing configuration, OAuth, process isolation, and schema-cache contracts.
146 Streamable HTTP opens its long-lived GET/SSE listener immediately after
147 initialization, so JetBrains project-level `.mcp.json` servers no longer lose
148 their pending session before the first tool call. Lost sessions converge on
149 one bounded rebuild and one replay, read-only surfaces consume every cursor
150 page, prompts/resources share the tool session, and shutdown terminates HTTP
151 sessions and local processes. MCP calls also accept a single JSON-object
152 string in `use_capability.arguments`, while rejecting arrays, scalars, invalid
153 JSON, and nested encoded strings. `/mcp` and Desktop expose redacted protocol,
154 listening, reconnect, and error-category diagnostics without session IDs.
155
156 - **v1.24.2 session snapshot & recovery root fix:** Keep PR #7982's WAL/CAS/lease
157 safety foundation, but replace process-level "I hold a lease" ownership with a
158 generation-bound `SessionWriteAuthority`. Same-revision tool-preview/load
159 reshapes no longer false-diverge; recovery files are bounded to one path per
160 writer/lineage; empty checkpoints heal from their own WAL; projection lineage
161 rebinds across upgrade/model switch and inherits across recovery forks without
162 changing provider-visible prompt bytes. Catalog upgrades to disposable
163 `session-catalog/v3.sqlite` with recovery lineage roles
164 (`normal|covered_copy|adopted|diverged`); covered idle copies move to the
165 recoverable `.trash` using a 15-minute idle threshold applied on two early
166 sweeps (at startup and ~20 minutes later), then a 24-hour threshold on the
167 6-hour background ticker; independent diverged branches stay and are listed
168 for user choice. v1/v2 catalogs are
169 left byte-unchanged for coexistence/downgrade.
170 **v1.24.1** only hid/reclaimed already-created covered copies and fixed Windows
171 flash-window startup; **v1.24.2** stops the misclassification source and repairs
172 existing user directories without rewriting authoritative JSONL/WAL/sidecar data.
173
174 - Goal now runs continuously by default: the former 16-round per-Run boundary,
175 10/20/40 cross-Run quotas, default wall-clock budget, and numeric
176 no-progress/Todo-stall pauses no longer stop valid work. Progress guards still
177 detect repeated host outcomes and zero-evidence work, but redirect the model
178 to re-plan instead of producing `goal_run_budget` or `goal_stuck`. Explicit
179 `[agent].goal_token_budget`, `--max-steps`, positive time/cost budgets, manual
180 pause/stop, genuine user/external blockers, and evaluator fail-closed behavior
181 remain available. The Goal token budget defaults to `0` (off); resuming its
182 `budget_spend` pause grants a fresh slice without clearing cumulative usage.
183 Goal status reports turns, provider requests, tokens, the optional configured
184 token threshold, and cumulative active work time. Bot `max_steps` also
185 defaults to `0` (continuous), while positive user configuration is enforced.
186
187 - Removed numeric Goal pauses in existing sidecars automatically normalize to
188 `running` without sending a model request. Active Goal sidecars write
189 `turnsLimit: -1` as a downgrade-safe unlimited sentinel while public runtime
190 APIs retain deprecated limit fields as `0`. The migration preserves unknown
191 fields, todos, checkpoints, usage, evidence, and historical metadata.
192
193 - Goal is now the sole long-task runtime. Historical AutoResearch sidecars
194 migrate transactionally into Goals with research compatibility metadata. Invalid archives block
195 fail closed and remain read-only, retaining the task id and compatibility mode
196 for a restart or `/goal resume` retry; successful Goal-only sidecars omit the
197 old task id and write an explicit downgrade fence so previous readers cannot
198 reactivate the removed runtime.
199
200 - Context-dependent workflow tools now share one host-side execution boundary.
201 Goal, Plan sign-off, and background-job calls cannot reach permissions,
202 hooks, leases, or Execute outside their owning context; mixed batches execute
203 valid calls once and stop safely after one repair. Child agents also isolate
204 inherited Goal, Jobs, and live memory queues, while persisted tool identity
205 records the effective child schema projection.
206
207 - **Issue #7575:** Linux Bash under bubblewrap no longer mounts a fresh empty
208 `--tmpfs /tmp` on every call. Consecutive commands in the same logical session
209 now share a private temporary directory (bound at `/tmp` on Linux, exported via
210 `TMPDIR`/`TMP`/`TEMP` on all platforms) without exposing the host public
211 temporary root. `/new`, `/clear`, resume of another session, and branch
212 switches rotate the directory; model/settings hot rebuilds keep it. Sub-agent
213 runs get independent directories. Temporary files are not durable across process
214 restarts.
215
216 ### Added
217
218 - Added `[ui].show_turn_usage` so CLI/TUI users can hide per-request token and
219 cost receipts from transcript scrollback without disabling usage accounting.
220
221 ## [1.20.0] — 2026-08-05
222
223 Extension kernel, Task Monitor, and safer Goal completion.
224
225 Compact decision surfaces, local decision receipts, unified extension kernel,
226 native Task Monitor, bounded sub-agent progress, Goal fail-closed completion,
227 MiMo and DashScope Responses fixes, SSH remote access simplification, and
228 multiple Desktop stability improvements.
229
230 ### Highlights
231
232 - **Unified Extension Kernel and Extension Protocol v1**: Immutable runtime
233 snapshots, fail-atomic reload, Plugin Manifest v1 (prompts, themes, full-trust
234 code runtimes), stable JSON-RPC sidecar protocol, interceptor dispatch,
235 streaming provider adapter, structured UI, and Go SDK.
236 - **Native Task Monitor**: Monitor agent tasks natively in CLI and Desktop with
237 lifecycle semantics and session-scoped summary view.
238 - **Bounded Sub-agent Progress Forwarding**: Forward structured progress for
239 `task`, `parallel_tasks`, and `fleet` without flooding the parent stream.
240 Renders nested lifecycle cards in Desktop and stable per-child transcript
241 slots in CLI.
242 - **Goal Completion Fail-Closed**: Replace free-form Goal footer markers with a
243 stable `update_goal` tool and epoch-scoped per-turn reports. Centralized
244 completion logic with bounded evaluator, progress-aware budgets, and
245 pause/resume controls.
246 - **Ablation Subsystem Switches**: Switch subsystems off behind one shared
247 vocabulary for controlled experiments. Includes planner, subagent, retrieval,
248 evidence, and compaction.
249 - **Benchmark Cost per Solved Task**: Report cost per solved task, tokens per
250 solved, median wall time, and failure-class breakdown in e2e reports.
251 - **Compact Decision Surfaces and Local Receipts**: Compact footer decision-card
252 layout with bounded scroll, dense action rows, and overflow disclosure.
253 Record bounded Ask, approval, and recovery decisions as local transcript
254 receipts.
255 - **Simplified SSH Remote Access**: Remove Remote Workbench protocol and
256 stacks; reuse CLI/Serve remote model. Desktop opens per-host native web child
257 windows via SSH. Keyless remote Serve setup with loopback-only page.
258 - **Model Usage Charts with Primer Palette**: Replace monochrome accent ramp
259 with GitHub Primer data-viz two-set categorical palette. Fix donut overflow
260 on hover and keyboard accessibility.
261 - **Cross-platform Extension and Task Monitor Reliability**: Make
262 content-reference eviction deterministic, reject Unix and Windows absolute
263 plugin paths consistently, stabilize parallel-task cancellation, and restore
264 reliable Windows validation for Task Monitor and remote provider setup.
265 - **MiMo and DashScope Responses Wire Alignment**: Fix multi-turn tool loops,
266 reasoning round-trip, JSON output for MiMo; fix DashScope second-turn 400
267 error, all-zero usage suppression, and vendor-aware cache TTL.
268 - **Desktop Stability Fixes**: Recover stuck updates and legacy WebKit, contain
269 macOS alias repair startup crashes, keep composer overflow stacks readable,
270 and harden account verification and community flows.
271 - **Remote Web Recovery After SSH Drops**: Add integration regression test for
272 SSH drop, forward recovery, and window reload. Document transient outage
273 behavior.
274 - **CI: Auto-minimize Activity-Farming Spam Comments**: Detect and minimize
275 template spam comments from non-contributor accounts based on structural
276 signals.
277
278 ### Added
279
280 - Added Extension Protocol v1 and the unified extension kernel: installed or
281 linked sidecars can contribute tools, skills, commands, hooks, MCP servers,
282 providers, interceptors, and structured UI surfaces through a versioned
283 NDJSON contract and the public Go SDK. CLI, Desktop, ACP, and Serve support
284 fail-atomic runtime reloads; Serve also renders extension surfaces and lists
285 extension-hosted providers without exposing credentials.
286 - Added the structured Goal completion protocol: the always-registered
287 `update_goal` tool (continue/complete/blocked with reason and next_action)
288 replaces the `[goal:*]` footer markers. The Goal FSM is now the exclusive
289 cross-turn decision point and validates every complete claim against Delivery
290 readiness; when the model submits no report, an independent bounded evaluator
291 (recovery_model → guardian_model → main model, no tools/history, usage
292 attributed to `goal-evaluator`) judges the turn once, and any evaluator
293 failure pauses the goal instead of continuing silently.
294 - Added Goal budget classes with safe pauses: simple 10 turns / 200k tokens,
295 write 20 turns / 400k tokens, AutoResearch 40 turns / 800k tokens, and a
296 4-turn no-host-verifiable-progress gate. Pauses keep all Goal state; `/goal
297 resume` continues and adds one slice of the current class when the pause was
298 budget-related. `/goal status` shows the full turn/token/no-progress runtime,
299 and `/goal pause` manually suspends a running Goal.
300 - Added the `goalRuntime` nested view to the desktop Meta, the remote protocol
301 (`session/goal/pause` operation, `goalRuntime` DTO on session meta), and the
302 ACP status payload; the desktop Composer goal menu shows the runtime summary
303 with distinct pause/end/resume actions.
304
305 ### Changed
306
307 - Delivery no longer retries final-answer readiness with hidden model messages:
308 a plain Delivery run ends on the first unsatisfied final answer and surfaces
309 the recovery card, while a Goal + Delivery run has the Goal FSM absorb the
310 failure and continue under budget with the missing requirements as the next
311 turn's prompt. Historical `[goal:*]` footers are stripped from old transcripts
312 for display only and never participate in state decisions.
313 - Added a **Remote SSH** module (VS Code Remote-SSH style): a user-global
314 `[remote]` host config, `reasonix remote` CLI (add/list/remove/import/test/
315 connect/status/forward/serve/fs) and `/remote` slash command, an SSH transport
316 with trust-on-first-use host-key verification, keepalive + exponential-backoff
317 reconnect, `-L`/`-R` port forwarding, and SFTP file access. `connect`
318 bootstraps a persistent `reasonix serve` on the remote host and tunnels its
319 loopback port so the full agent runs remotely. The desktop app adds a
320 **Settings -> Remote SSH** host manager, a remote file browser/editor, a
321 port-forwarding panel, and a status-bar connection chip. Linux/macOS remotes.
322 - Added `reasonix serve --port-file/--token-file/--pid-file` so a supervised
323 headless serve can bind an ephemeral port and read its auth token from a file
324 (keeping it out of `ps`).
325 - Added an authenticated, loopback-only Provider setup page for `reasonix
326 serve`. A Serve whose selected Provider is missing its API key now remains
327 reachable, stores the submitted key in that host's Reasonix credential file,
328 and rebuilds the active controller in place without restarting Serve.
329 - Added Claude Code-style searchable CLI pickers for models, providers, and
330 sessions, with arrow, Vim, and `Ctrl+P` / `Ctrl+N` navigation.
331 - Added `-p` / `--print`, `text`, `json`, and `stream-json` output modes for
332 one-shot use and automation.
333 - Added session-scoped `--allowed-tools`, repeatable `--add-dir`, Claude-compatible
334 permission modes, flexible `--resume [QUERY]`, and the `--copy` resume escape
335 hatch.
336 - Added `/status` details for the active model, effort, cache, Git state,
337 background jobs, work profile, and provider balance where available.
338 - Remote SSH workspaces now open as a standalone remote web window again.
339 Opening a workspace from the status bar or the Remote Server tab starts or
340 reuses the remote `reasonix serve`, tunnels its loopback port, and opens the
341 Serve web client in a dedicated per-host window. The remote web page uses
342 the provider configuration and API keys on the **remote** host; the desktop
343 no longer exposes its local providers to remote hosts. If the selected remote
344 Provider is missing its API key, the window opens a setup page that saves the
345 key only on that host and then opens the normal Serve UI. The Remote Workbench
346 protocol, its Provider Broker, and the same-window remote projection were
347 removed. Legacy mirror and provider-trust files are not deleted
348 automatically; Settings -> Remote SSH shows a cleanup card when they exist.
349 The hidden `remote attach-workspace`, `remote runtime-workbench`, and
350 `remote workbench-build-id` commands now fail with a pointer to
351 `reasonix remote connect <host> --open`.
352 - Automatic Plan Mode has been retired. Plan Mode is now always entered through
353 an explicit user choice, and the one-time config v5 upgrade removes legacy
354 `agent.auto_plan` and `agent.auto_plan_classifier` values so upgraded users
355 receive the same behavior as new users.
356 - `Shift+Tab` now cycles CLI safe modes from Ask to Auto to Plan, while YOLO
357 remains an independent `Ctrl+Y` toggle.
358 - Model, provider, resume, and approval menus now use consistent row selection;
359 slash completion, help, aliases, and dispatch share one command registry.
360 - The full-screen CLI composer now uses theme-accented borders and a slim bar
361 cursor by default, grows within the available terminal height, scrolls long
362 drafts independently, and preserves selections across explicit image paste.
363 - The persistent CLI footer now uses a responsive, theme-aware layout for
364 interaction state, model, effort, localized work mode, Git identity, cache,
365 context, compaction headroom, jobs, and balance. Narrow terminals move or
366 compact complete groups instead of clipping labels.
367 - CLI clipboard actions now separate terminal-native text paste from explicit
368 image paste: `Ctrl+V` on macOS/Linux, `Alt+V` on Windows, or `/paste-image`.
369 Local transcript copy verifies the native clipboard write, while SSH uses a
370 clearly labelled OSC 52 fallback.
371 - Runtime rebuilds after model, effort, or work-mode changes now preserve the
372 conversation, session permission overrides, additional directories, and
373 session lease ownership.
374 - Agent execution now monitors host-observed Todo progress automatically. A
375 stalled current item receives a recovery nudge after 8 tool-call rounds with
376 no new completion, unique read, command, or mutation, and pauses with saved
377 work after 16. Exact repeats do not renew the progress lease; real work does.
378 Two-level task lists keep the single in_progress contract: the active
379 sub-step is the only current item while its phase stays pending, and the
380 phase becomes in_progress to sign off only after all of its sub-steps are
381 completed. A level-1 sub-step with no phase header above it is rejected.
382 Executor and planner rounds now use automatic progress management. Retired
383 `[agent].max_steps` and `planner_max_steps` keys remain parseable for upgrades,
384 but are ignored and removed by a one-time migration so stale hidden limits
385 cannot truncate new behavior. One-off CLI and unattended bot limits remain.
386
387 ### Fixed
388
389 - Fixed long parallel sub-agent research being silently lost when combined
390 `parallel_tasks` or `fleet` answers exceeded the single-tool output limit.
391 Persisted sessions now keep each child transcript independently, return a
392 bounded fair preview plus stable reference for every result, and page full
393 answers through the conversation-scoped `read_subagent_result` tool.
394 - Fixed Remote Workbench failing with only `initialize: workbench-desktop:
395 connection closed` on fresh or cross-platform SSH hosts. Desktop now proves
396 the exact Host CLI Build ID, provisions the matching verified release without
397 requiring remote npm, runs the managed binary explicitly, and preserves a
398 safe structured bootstrap error when the remote command exits early.
399 - Hardened Bash permission reuse for dynamic and indirect execution. Parameter/arithmetic expansions,
400 assignments, redirects, heredocs, and globs can only be remembered as exact
401 `Bash=<literal>` rules, while still using Auto's normal fallback. Nested or
402 indirect execution now requires a human in interactive Ask/Auto and fails
403 closed in headless Ask/Auto/DontAsk. Broad Bash rules, Guardian/hook allows,
404 and the approved-plan window can no longer silently authorize that stricter
405 class; YOLO remains the explicit full-access bypass and sandbox enforcement
406 is unchanged.
407 - Fixed Desktop sessions incorrectly locking themselves during Goal + Delivery
408 mode changes, controller rebuilds, duplicate-tab restore, and background
409 reattachment. Desktop now keeps one process-local runtime owner per canonical
410 session, fences stale controller events by runtime epoch, blocks sends until
411 that runtime is ready, and scopes single-instance ownership to
412 `REASONIX_HOME` instead of the executable path. Switching saved sessions is
413 now transactional: a target build, restore, or lease failure leaves the
414 current controller, lease, path, mode profile, and runtime epoch untouched.
415 - Stabilized the desktop rich composer caret after skill and plugin invocation
416 tags. DOM→model and model→DOM selection mapping now treat invocation chips as
417 zero-length atoms while still counting user text that lands inside the NBSP
418 caret anchor (common on Windows WebView2), restore both selection ends, and
419 recover the insertion point from a `beforeinput` snapshot when the browser
420 temporarily loses selection — so mid-text edits no longer jump to the end.
421 - Isolated the Windows desktop WebView2 shell from stale system proxies, so an
422 exited proxy client cannot leave the embedded UI hidden during startup. If
423 WebView2 still does not reach DOM-ready within 15 seconds, Reasonix now shows
424 the native window with a recovery prompt instead of appearing not to launch.
425 Remote Markdown images are fetched by the backend with Reasonix's proxy
426 configuration instead of bypassing that proxy through the isolated WebView.
427 - Restored captured-mouse right-click text paste, made composer drag selection
428 copy through the verified native clipboard path, and kept non-Git footer
429 telemetry left-aligned without reserving an empty data band.
430 - Restored stateful MCP behavior after the v1.17.13 regression: user-added
431 servers work without extra trust settings (including delivery-mode on-demand
432 calls), repository-provided servers use one exact launch confirmation, and
433 stdio tools reuse one persistent process so browser sessions survive across
434 calls without repeated startup latency. The former trust/reverify/catalog
435 management UI and CLI are removed.
436 - Localized persistent-footer labels and displayed work-mode values in English,
437 Simplified Chinese, and Traditional Chinese, while keeping command arguments
438 stable.
439 - Restored the `0.53` content boundary: model output, tool output, session
440 transcripts, recovery branches, and background-job artifacts retain their
441 original text instead of being rewritten by heuristic secret redaction.
442 Credential masking remains in key-entry summaries and explicit diagnostic or
443 session-cleanup paths. Transcript-bearing session/job sidecars are kept
444 private (`0600`, with private job directories), and the retired
445 `redact_tool_output` setting is removed with a one-time upgrade notice.
446
447 ### Notes
448
449 - Full bilingual release notes:
450 <https://reasonix.io/changelog/v1.20.0/> ·
451 [GitHub release](https://github.com/esengine/DeepSeek-Reasonix/releases/tag/desktop-v1.20.0).
452 - The detailed entries below accumulated on `main-v2` after 1.0.0 and shipped
453 across 1.1.0–1.20.0; per-version attribution lives in the per-version release
454 notes linked above.
455
456 ## 1.1.0 – 1.19.7
457
458 Per-version entries for the intermediate releases are published in the
459 [bilingual release notes](https://reasonix.io/changelog/) and on the
460 [GitHub releases page](https://github.com/esengine/DeepSeek-Reasonix/releases).
461
462 ## [1.0.0] — 2026-06-03
463
464 First stable release — a **ground-up rewrite in Go**. Not an upgrade of the `0.x`
465 TypeScript line; a new codebase that becomes the default (`main-v2`).
466
467 ### Highlights
468
469 - **Go kernel**: a single static binary (CGO-free), cross-compiled for
470 darwin/linux/windows on amd64 + arm64. Distributed via npm (the package wraps
471 the native binary), Homebrew (`esengine/reasonix` tap), and release archives;
472 no Node runtime needed to run it.
473 - **Agent core**: the loop, built-in tools (read/write/edit/multi_edit/glob/grep/
474 ls/bash/web_fetch/todo_write), permission gate, sandboxed bash, and the
475 DeepSeek prefix-cache–oriented design.
476 - **Subagents**: `task` plus explore/research/review/security_review skill agents.
477 - **Skills & hooks**: Claude-Code-style skills (`internal/skill`) and hooks
478 (`internal/hook`), symlink-aware and slash-integrated.
479 - **MCP client**: connect external servers over stdio / Streamable HTTP; reads
480 `[[plugins]]` and a Claude-Code `.mcp.json`.
481 - **Code intelligence via CodeGraph**: a tree-sitter symbol/call graph
482 (`codegraph_*` tools) replaces embedding semantic search — no embedding service
483 or API cost. Fetched into a local cache on first use (or `reasonix codegraph
484 install`) and indexed in the background, so installs and startup stay fast.
485 - **Plan mode** with evidence-backed step sign-off (`complete_step`).
486 - **Memory**: `REASONIX.md` hierarchy + auto-memory, folded into the cache-stable
487 prefix.
488 - **ACP** (`reasonix acp`) and an HTTP/SSE server frontend; desktop app (Wails).
489
490 ### Fixed
491
492 - **File encoding support restored** — GBK/GB18030 (and other non-UTF-8) files
493 can now be read, edited, and grepped correctly. The v2 rewrite had dropped
494 v1's encoding detection; files in CJK Windows charsets were silently misread
495 or rejected as binary. The read/edit/write round-trip now preserves the
496 original file encoding. (#2637)
497
498 ### Notes
499
500 - Versions: the legacy TypeScript line stays in `0.x`; the Go line starts at
501 `1.0.0`. See [docs/MIGRATING.md](docs/MIGRATING.md).
502 - Release archives ship a bare binary; CodeGraph is fetched on first use. Windows
503 support for the fetched runtime is unverified — install `codegraph` on PATH if
504 the auto-fetch doesn't resolve there.
505
506 [1.20.0]: https://github.com/esengine/DeepSeek-Reasonix/releases/tag/desktop-v1.20.0
507 [1.0.0]: https://github.com/esengine/DeepSeek-Reasonix/releases/tag/v1.0.0
508
508 lines MARKDOWN