| 1 | #![allow(dead_code)] |
| 2 | //! System prompts for different modes. |
| 3 | //! |
| 4 | //! Prompts are assembled from composable layers loaded at compile time from |
| 5 | //! the single [`text`] module: |
| 6 | //! constitution + personality overlay → `message[0]` (byte-stable). |
| 7 | //! mode delta + approval policy → request-time runtime metadata. |
| 8 | //! Tool availability comes only from the per-turn model catalog. |
| 9 | //! |
| 10 | //! Keeping every layer's text in one module makes prompt tuning a |
| 11 | //! single-file operation. |
| 12 | |
| 13 | use crate::models::{SystemBlock, SystemPrompt}; |
| 14 | use crate::project_context::load_project_context_with_parents; |
| 15 | use std::path::{Path, PathBuf}; |
| 16 | use std::sync::{LazyLock, Mutex}; |
| 17 | |
| 18 | pub mod base_preview; |
| 19 | pub(crate) mod text; |
| 20 | |
| 21 | #[derive(Debug, Clone)] |
| 22 | pub struct PromptSessionContext<'a> { |
| 23 | pub user_memory_block: Option<&'a str>, |
| 24 | pub goal_objective: Option<&'a str>, |
| 25 | pub project_context_pack_enabled: bool, |
| 26 | /// Resolved BCP-47 locale tag for the `## Environment` block in |
| 27 | /// the system prompt (e.g. `"en"`, `"zh-Hans"`, `"ja"`). The |
| 28 | /// caller is responsible for resolving this from `Settings`; no |
| 29 | /// disk I/O happens inside the prompt builder, so the workspace- |
| 30 | /// static portion of the system prompt stays cache-friendly. |
| 31 | pub locale_tag: &'a str, |
| 32 | /// When true, a ## Language Output Requirement block is appended |
| 33 | /// to the system prompt instructing the model to respond in |
| 34 | /// the resolved session locale. |
| 35 | pub translation_enabled: bool, |
| 36 | /// Active model identifier. The bundled constitution is model-agnostic, |
| 37 | /// but embedders may still provide a prompt override containing |
| 38 | /// `{model_id}`. Defaults to `"codewhale"` when the caller doesn't supply one. |
| 39 | pub model_id: &'a str, |
| 40 | /// Route-effective context window, when known. Prompt composition no |
| 41 | /// longer prints context-window facts, but the field remains part of the |
| 42 | /// session context contract for embedders and future runtime metadata. |
| 43 | pub context_window_override: Option<u32>, |
| 44 | /// Optional output-verbosity mode. `concise` appends a short output |
| 45 | /// discipline block; unset keeps the normal conversational prompt. |
| 46 | pub verbosity: Option<&'a str>, |
| 47 | /// Restrict skill discovery to Codewhale-owned roots plus explicit |
| 48 | /// `skills_dir` configuration. |
| 49 | pub skills_scan_codewhale_only: bool, |
| 50 | /// Immutable plugin snapshot owned by this App/Engine workspace context. |
| 51 | /// Never sourced from process-global mutable state. |
| 52 | pub plugin_registry: Option<&'a crate::plugins::PluginRegistry>, |
| 53 | /// Active mode. Its doctrine overlay ships once here, in the stable |
| 54 | /// prefix, rather than being re-asserted in `<turn_meta>` on every user |
| 55 | /// message (#4780) — repetition-per-turn out-shouts the constitution and |
| 56 | /// makes the model perform compliance instead of exercising judgment. |
| 57 | /// |
| 58 | /// Changing mode does invalidate the prefix cache, which is the intended |
| 59 | /// trade: mode changes are rare, user messages are not. |
| 60 | pub mode: crate::tui::app::AppMode, |
| 61 | } |
| 62 | |
| 63 | impl Default for PromptSessionContext<'_> { |
| 64 | fn default() -> Self { |
| 65 | Self { |
| 66 | user_memory_block: None, |
| 67 | goal_objective: None, |
| 68 | project_context_pack_enabled: false, |
| 69 | locale_tag: "en", |
| 70 | translation_enabled: false, |
| 71 | model_id: "codewhale", |
| 72 | context_window_override: None, |
| 73 | verbosity: None, |
| 74 | skills_scan_codewhale_only: false, |
| 75 | plugin_registry: None, |
| 76 | mode: crate::tui::app::AppMode::Agent, |
| 77 | } |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | /// Conventional location for the structured session relay artifact (#32). |
| 82 | /// A previous session writes it on exit / `/compact`; the next session reads |
| 83 | /// it back on startup and prepends it to the system prompt so a fresh agent |
| 84 | /// doesn't have to re-discover open blockers from scratch. |
| 85 | pub const HANDOFF_RELATIVE_PATH: &str = ".codewhale/handoff.md"; |
| 86 | /// Legacy handoff path for reading from existing installs. |
| 87 | const LEGACY_HANDOFF_RELATIVE_PATH: &str = ".deepseek/handoff.md"; |
| 88 | |
| 89 | /// Per-file size cap for `instructions = [...]` entries (#454). Mirrors |
| 90 | /// the existing project-context cap in `project_context::load_context_file` |
| 91 | /// so a malicious / oversized include can't blow the prompt budget on |
| 92 | /// its own. Files larger than this are truncated with an explicit `[…truncated: N bytes omitted]` |
| 93 | /// marker rather than skipped entirely so the model still sees the head. |
| 94 | const INSTRUCTIONS_FILE_MAX_BYTES: usize = 100 * 1024; |
| 95 | |
| 96 | /// System prompt block appended when `translation_enabled` is true. |
| 97 | /// Instructs the model to respond in the resolved session locale for all |
| 98 | /// natural-language output — explanations, summaries, conversation. |
| 99 | /// Code identifiers, untranslatable technical terms, and explicitly |
| 100 | /// requested English code blocks are exempt. |
| 101 | fn translation_output_instruction(locale_tag: &str) -> String { |
| 102 | let target_language = translation_target_language_for_tag(locale_tag); |
| 103 | format!( |
| 104 | "\ |
| 105 | ## Language Output Requirement\n\ |
| 106 | \n\ |
| 107 | The user requires all responses in {target_language}. \ |
| 108 | Always respond in {target_language} — use natural, professional language for all \ |
| 109 | explanations, code comments, summaries, and conversational turns. \ |
| 110 | Only output English for:\n\ |
| 111 | - Code identifiers (variable names, function names, file paths)\n\ |
| 112 | - Technical terms that lack a standard translation in {target_language}\n\ |
| 113 | - Code blocks the user explicitly requests in English\n\n\ |
| 114 | This is a hard display requirement: the user does not read English, \ |
| 115 | so any English prose in your response will block their decision-making." |
| 116 | ) |
| 117 | } |
| 118 | |
| 119 | fn concise_output_discipline_instruction() -> &'static str { |
| 120 | "\ |
| 121 | ## Concise Output Discipline |
| 122 | |
| 123 | To minimize token usage and optimize speed: |
| 124 | - Output only direct, actionable code, technical steps, or final answers. |
| 125 | - Eliminate all conversational filler, fluff, introductions, transitions, or summarizing conclusions. |
| 126 | - Do NOT explain what you are about to do or what you have just completed. |
| 127 | - Do NOT provide conversational status updates before or after running tools. |
| 128 | - Keep explanations and comments extremely brief and technical, explaining only non-obvious reasoning." |
| 129 | } |
| 130 | |
| 131 | fn is_concise_verbosity(value: Option<&str>) -> bool { |
| 132 | value.is_some_and(|v| v.trim().eq_ignore_ascii_case("concise")) |
| 133 | } |
| 134 | |
| 135 | fn translation_target_language_for_tag(locale_tag: &str) -> &'static str { |
| 136 | let normalized = locale_tag.trim().to_ascii_lowercase(); |
| 137 | if normalized.starts_with("ja") { |
| 138 | "Japanese (日本語)" |
| 139 | } else if normalized.starts_with("zh-hant") |
| 140 | || normalized.contains("-tw") |
| 141 | || normalized.contains("-hk") |
| 142 | || normalized.contains("-mo") |
| 143 | { |
| 144 | "Traditional Chinese (繁體中文)" |
| 145 | } else if normalized.starts_with("zh") { |
| 146 | "Simplified Chinese (简体中文)" |
| 147 | } else if normalized.starts_with("pt") { |
| 148 | "Brazilian Portuguese (Português do Brasil)" |
| 149 | } else if normalized.starts_with("es") { |
| 150 | "Latin American Spanish (Español latinoamericano)" |
| 151 | } else if normalized.starts_with("vi") { |
| 152 | "Vietnamese (Tiếng Việt)" |
| 153 | } else if normalized.starts_with("ko") { |
| 154 | "Korean (한국어)" |
| 155 | } else if normalized.starts_with("ca") { |
| 156 | "Catalan (Català)" |
| 157 | } else if normalized.starts_with("de") { |
| 158 | "German (Deutsch)" |
| 159 | } else if normalized.starts_with("fr") { |
| 160 | "French (Français)" |
| 161 | } else if normalized.starts_with("id") { |
| 162 | "Indonesian (Bahasa Indonesia)" |
| 163 | } else if normalized.starts_with("hi") { |
| 164 | "Hindi (हिन्दी)" |
| 165 | } else if normalized.starts_with("ru") { |
| 166 | "Russian (Русский)" |
| 167 | } else if normalized.starts_with("uk") { |
| 168 | "Ukrainian (Українська)" |
| 169 | } else { |
| 170 | "English" |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | /// Render a `## Environment` block listing the resolved locale tag and the |
| 175 | /// actionable host facts that affect command syntax. |
| 176 | /// |
| 177 | /// The block is appended to the workspace-static portion of the system |
| 178 | /// prompt (after mode prompt + project context, before configured |
| 179 | /// instructions / skills). `locale_tag` is resolved by the caller from |
| 180 | /// `Settings` so this function stays I/O-free. |
| 181 | /// |
| 182 | /// `platform` and `shell` remain because they change how commands must be |
| 183 | /// written and are stable for the life of the process. The release version was |
| 184 | /// removed by the turn-meta diet: it is telemetry the model cannot act on and |
| 185 | /// churned the otherwise-static prefix on every release. The live workspace |
| 186 | /// path is delivered per-turn via `<turn_meta>` (see `turn_metadata_block`). |
| 187 | pub(crate) fn render_environment_block(_workspace: &Path, locale_tag: &str) -> String { |
| 188 | let platform = std::env::consts::OS; |
| 189 | let shell = crate::shell_dispatcher::global_dispatcher() |
| 190 | .kind() |
| 191 | .binary() |
| 192 | .to_string(); |
| 193 | |
| 194 | format!( |
| 195 | "## Environment\n\ |
| 196 | \n\ |
| 197 | - lang: {locale_tag}\n\ |
| 198 | - platform: {platform}\n\ |
| 199 | - shell: {shell}" |
| 200 | ) |
| 201 | } |
| 202 | |
| 203 | /// Source for an `EngineConfig.instructions` entry. Either a disk file (loaded |
| 204 | /// at render time, original semantics) or an inline string (content baked into |
| 205 | /// `EngineConfig`, no disk I/O at render time). |
| 206 | /// |
| 207 | /// The inline variant is useful for embedders that compute instructions at |
| 208 | /// runtime (e.g. rendering a template with workspace-specific substitutions) |
| 209 | /// and don't want to stage the content to a disk file just to satisfy a path |
| 210 | /// API. Staging adds two problems the inline path avoids: |
| 211 | /// |
| 212 | /// 1. The disk file looks like editable config but gets overwritten on |
| 213 | /// every launch — confusing for users browsing the install dir. |
| 214 | /// 2. Multi-engine setups need per-engine paths to avoid `rehydrate` |
| 215 | /// reading another session's instructions; with inline sources the |
| 216 | /// content lives in the per-engine `EngineConfig` and the race |
| 217 | /// surface goes away. |
| 218 | /// |
| 219 | /// `From<PathBuf>` is provided so existing callers passing `Vec<PathBuf>` can |
| 220 | /// keep working with a `.into()` upgrade at the call site. |
| 221 | #[derive(Debug, Clone)] |
| 222 | pub enum InstructionSource { |
| 223 | /// Load this file from disk at prompt-render time. Original behavior: |
| 224 | /// missing files are skipped with a warning, oversized files are |
| 225 | /// truncated to `INSTRUCTIONS_FILE_MAX_BYTES` with an `[…elided]` |
| 226 | /// marker. |
| 227 | File(PathBuf), |
| 228 | /// Use the provided string directly. `name` becomes the |
| 229 | /// `<instructions source="…">` attribute (typically a synthetic |
| 230 | /// identifier like `embedded:my-template` or a logical path). |
| 231 | Inline { name: String, content: String }, |
| 232 | } |
| 233 | |
| 234 | impl From<PathBuf> for InstructionSource { |
| 235 | fn from(path: PathBuf) -> Self { |
| 236 | InstructionSource::File(path) |
| 237 | } |
| 238 | } |
| 239 | |
| 240 | impl From<&PathBuf> for InstructionSource { |
| 241 | fn from(path: &PathBuf) -> Self { |
| 242 | InstructionSource::File(path.clone()) |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | /// Render the `instructions = [...]` config array as a single |
| 247 | /// system-prompt block (#454). Each source is processed in declared order; |
| 248 | /// missing `File` sources are skipped with a tracing warning so a stale entry |
| 249 | /// doesn't fail the launch. Empty input (or all sources missing/empty) |
| 250 | /// returns `None` so callers append nothing. |
| 251 | fn render_instructions_block(sources: &[InstructionSource]) -> Option<String> { |
| 252 | let mut sections: Vec<String> = Vec::new(); |
| 253 | for source in sources { |
| 254 | let (raw_source_name, raw_content): (String, String) = match source { |
| 255 | InstructionSource::File(path) => match std::fs::read_to_string(path) { |
| 256 | Ok(raw) => (path.display().to_string(), raw), |
| 257 | Err(err) => { |
| 258 | tracing::warn!( |
| 259 | target: "instructions", |
| 260 | ?err, |
| 261 | ?path, |
| 262 | "skipping unreadable instructions file" |
| 263 | ); |
| 264 | continue; |
| 265 | } |
| 266 | }, |
| 267 | InstructionSource::Inline { name, content } => (name.clone(), content.clone()), |
| 268 | }; |
| 269 | let trimmed = raw_content.trim(); |
| 270 | if trimmed.is_empty() { |
| 271 | continue; |
| 272 | } |
| 273 | let body = if trimmed.len() > INSTRUCTIONS_FILE_MAX_BYTES { |
| 274 | let head_end = (0..=INSTRUCTIONS_FILE_MAX_BYTES) |
| 275 | .rev() |
| 276 | .find(|&i| trimmed.is_char_boundary(i)) |
| 277 | .unwrap_or(0); |
| 278 | format!( |
| 279 | "{}\n[…truncated: {} of {} bytes omitted — consider splitting this instructions file]", |
| 280 | &trimmed[..head_end], |
| 281 | trimmed.len() - head_end, |
| 282 | trimmed.len() |
| 283 | ) |
| 284 | } else { |
| 285 | trimmed.to_string() |
| 286 | }; |
| 287 | sections.push(format!( |
| 288 | "<instructions source=\"{raw_source_name}\">\n{body}\n</instructions>" |
| 289 | )); |
| 290 | } |
| 291 | if sections.is_empty() { |
| 292 | None |
| 293 | } else { |
| 294 | Some(sections.join("\n\n")) |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | /// Read the workspace-local relay artifact, if present, and format it as a |
| 299 | /// system-prompt block. Returns `None` when the file is absent or empty so |
| 300 | /// callers can keep the default-uncluttered prompt for fresh workspaces. |
| 301 | fn load_handoff_block(workspace: &Path) -> Option<String> { |
| 302 | let primary = workspace.join(HANDOFF_RELATIVE_PATH); |
| 303 | let path = if primary.exists() { |
| 304 | primary |
| 305 | } else { |
| 306 | workspace.join(LEGACY_HANDOFF_RELATIVE_PATH) |
| 307 | }; |
| 308 | let raw = std::fs::read_to_string(&path).ok()?; |
| 309 | let trimmed = raw.trim(); |
| 310 | if trimmed.is_empty() { |
| 311 | return None; |
| 312 | } |
| 313 | Some(format!( |
| 314 | "## Previous Session Relay\n\nThe previous session in this workspace left a relay artifact at `{HANDOFF_RELATIVE_PATH}`. Consider it the first artifact to read on this turn — open blockers, in-flight changes, and recent decisions live there. Update or rewrite it before exiting if state changes materially.\n\n{trimmed}" |
| 315 | )) |
| 316 | } |
| 317 | |
| 318 | /// Load the structured user-global constitution, if present, and render it as |
| 319 | /// its own model-facing block. |
| 320 | pub(crate) fn load_user_constitution_block() -> Option<String> { |
| 321 | if user_constitution_disabled_by_setup_state() { |
| 322 | return None; |
| 323 | } |
| 324 | |
| 325 | let path = match codewhale_config::UserConstitution::path() { |
| 326 | Ok(path) => path, |
| 327 | Err(err) => { |
| 328 | tracing::warn!( |
| 329 | target: "prompts", |
| 330 | "could not resolve user-global constitution path: {err:#}" |
| 331 | ); |
| 332 | return None; |
| 333 | } |
| 334 | }; |
| 335 | |
| 336 | match codewhale_config::UserConstitution::load_from(&path) { |
| 337 | codewhale_config::UserConstitutionLoad::Loaded(constitution) => { |
| 338 | constitution.render_block(None) |
| 339 | } |
| 340 | codewhale_config::UserConstitutionLoad::Missing |
| 341 | | codewhale_config::UserConstitutionLoad::Empty => None, |
| 342 | codewhale_config::UserConstitutionLoad::Invalid(err) => { |
| 343 | tracing::warn!( |
| 344 | target: "prompts", |
| 345 | "skipping invalid user-global constitution {}: {err}", |
| 346 | path.display() |
| 347 | ); |
| 348 | None |
| 349 | } |
| 350 | codewhale_config::UserConstitutionLoad::Unreadable(err) => { |
| 351 | tracing::warn!( |
| 352 | target: "prompts", |
| 353 | "skipping unreadable user-global constitution {}: {err}", |
| 354 | path.display() |
| 355 | ); |
| 356 | None |
| 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | fn user_constitution_disabled_by_setup_state() -> bool { |
| 362 | match codewhale_config::SetupState::load() { |
| 363 | Ok(Some(state)) => matches!( |
| 364 | state.constitution_choice, |
| 365 | codewhale_config::ConstitutionChoice::Bundled |
| 366 | | codewhale_config::ConstitutionChoice::Deferred |
| 367 | | codewhale_config::ConstitutionChoice::ExpertOverride |
| 368 | ), |
| 369 | Ok(None) => false, |
| 370 | Err(err) => { |
| 371 | tracing::warn!( |
| 372 | target: "prompts", |
| 373 | "could not resolve setup-state path while loading user constitution: {err:#}" |
| 374 | ); |
| 375 | false |
| 376 | } |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | // ── Prompt layers loaded at compile time ────────────────────────────── |
| 381 | // |
| 382 | // Every bundled prompt layer lives in `prompts/text.rs` as a compile-time |
| 383 | // constant (consolidated from the retired per-layer `prompts/*.md` files; |
| 384 | // each constant is byte-identical to the file it replaced, trailing newline |
| 385 | // included). The constants are re-exported here so the existing |
| 386 | // `crate::prompts::NAME` paths used across the crate are unchanged. Edit |
| 387 | // prompt text in `text.rs` directly; the test suite below guards content |
| 388 | // and ordering invariants (constitution structure and binding gates #4032, |
| 389 | // byte-stable prefix ordering, prefix privacy #4632). |
| 390 | #[cfg(test)] |
| 391 | use text::CALM_PERSONALITY; |
| 392 | pub use text::{ |
| 393 | AGENT_MODE, BASE_PROMPT, COMPACT_TEMPLATE, CORE_EXECUTION_PROFILE_PROMPT, |
| 394 | GOAL_CONTINUATION_PROMPT, LANGUAGE_PROMPT, MEMORY_GUIDANCE, OPERATE_MODE, OUTPUT_PROMPT, |
| 395 | PLAN_MODE, |
| 396 | }; |
| 397 | |
| 398 | // ── Embedder prompt overrides ── |
| 399 | // Let an embedder replace these compile-time prompt constants at startup, |
| 400 | // so brand / slimming customizations live in the embedder crate instead of |
| 401 | // editing these files in-tree. Unset → the bundled constant (fully |
| 402 | // backward compatible). Intended to be set once at process start, before |
| 403 | // any engine spawns; later sets return the rejected override string. |
| 404 | static BASE_PROMPT_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 405 | static LOCALE_PREAMBLE_ZH_HANS_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 406 | static LOCALE_PREAMBLE_JA_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 407 | static LOCALE_PREAMBLE_PT_BR_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 408 | static LOCALE_PREAMBLE_VI_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 409 | static LOCALE_CLOSER_ZH_HANS_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 410 | static LOCALE_CLOSER_JA_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 411 | static LOCALE_CLOSER_PT_BR_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 412 | static LOCALE_CLOSER_VI_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 413 | static AUTHORITY_RECAP_OVERRIDE: std::sync::OnceLock<String> = std::sync::OnceLock::new(); |
| 414 | static STATIC_PROMPT_COMPOSER: std::sync::OnceLock<Box<StaticPromptComposer>> = |
| 415 | std::sync::OnceLock::new(); |
| 416 | static PROMPT_OVERRIDE_NOTICES: LazyLock<Mutex<Vec<String>>> = |
| 417 | LazyLock::new(|| Mutex::new(Vec::new())); |
| 418 | |
| 419 | /// Context passed to an embedder-provided static prompt composer. |
| 420 | /// |
| 421 | /// This hook only replaces the byte-stable base/personality prompt segment. |
| 422 | /// Mode deltas, approval policy, Core Execution, and action-specific relay |
| 423 | /// formatting stay owned by Codewhale. |
| 424 | #[non_exhaustive] |
| 425 | #[derive(Debug)] |
| 426 | pub struct StaticPromptCtx<'a> { |
| 427 | /// Active model identifier after caller-side routing. |
| 428 | pub model_id: &'a str, |
| 429 | /// Personality overlay requested for the base static prompt. |
| 430 | pub personality: Personality, |
| 431 | /// Default base/personality prompt layers that would be used without an |
| 432 | /// override. |
| 433 | pub default_layers: &'a str, |
| 434 | } |
| 435 | |
| 436 | /// Embedder hook for replacing Codewhale's byte-stable base/personality prompt |
| 437 | /// segment. |
| 438 | pub type StaticPromptComposer = dyn Fn(&StaticPromptCtx<'_>) -> String + Send + Sync + 'static; |
| 439 | |
| 440 | /// Replace `BASE_PROMPT` for all subsequent prompt composition. First call |
| 441 | /// wins; later calls return the rejected string. Set before spawning any |
| 442 | /// engine. |
| 443 | pub fn set_base_prompt_override(s: String) -> Result<(), String> { |
| 444 | set_prompt_override(&BASE_PROMPT_OVERRIDE, s) |
| 445 | } |
| 446 | |
| 447 | // ── Config-directory prompt overrides (issue #3638) ── |
| 448 | // Bridge the embedder override hooks above to a user-facing source: an |
| 449 | // optional file in the Codewhale config directory. This lets users repurpose |
| 450 | // the TUI for non-software use cases (e.g. long-form writing) by swapping the |
| 451 | // constitutional base prompt, without editing in-tree files or shipping a |
| 452 | // custom embedder build. |
| 453 | // |
| 454 | // Scope is deliberately narrow: only the byte-stable base prompt segment is |
| 455 | // user-overridable. Mode deltas, approval policy, Core Execution, and |
| 456 | // action-specific relay formatting stay owned by the runtime assembly (see |
| 457 | // `StaticPromptCtx`), so an override cannot strip safety-relevant guidance. |
| 458 | // A missing or empty file is a no-op — the bundled constant is used — so this |
| 459 | // is fully backward compatible. |
| 460 | // |
| 461 | // Because replacing the base prompt is a trust-boundary action (per maintainer |
| 462 | // review on #3638), the override file alone is NOT sufficient: the user must |
| 463 | // also set an explicit opt-in flag (`CODEWHALE_ALLOW_BASE_PROMPT_OVERRIDE`). |
| 464 | // This keeps replacing the global Constitution a deliberate, auditable act |
| 465 | // rather than something a stray file can do. |
| 466 | |
| 467 | /// Relative path, under the config directory, of the optional base-prompt |
| 468 | /// (constitution) override file. |
| 469 | pub const CONSTITUTION_OVERRIDE_FILE: &str = "prompts/constitution.md"; |
| 470 | |
| 471 | /// Env flag that must be set (`1`/`true`/`on`/`yes`) to enable config-dir base |
| 472 | /// prompt overrides. Required in addition to the override file so the global |
| 473 | /// base prompt can never be replaced by file presence alone. |
| 474 | pub const BASE_PROMPT_OVERRIDE_OPT_IN_ENV: &str = "CODEWHALE_ALLOW_BASE_PROMPT_OVERRIDE"; |
| 475 | |
| 476 | /// Whether the user has explicitly opted in to base-prompt overrides. |
| 477 | pub(crate) fn base_prompt_override_opt_in() -> bool { |
| 478 | match std::env::var(BASE_PROMPT_OVERRIDE_OPT_IN_ENV) { |
| 479 | Ok(v) => matches!( |
| 480 | v.trim().to_ascii_lowercase().as_str(), |
| 481 | "1" | "true" | "on" | "yes" |
| 482 | ), |
| 483 | Err(_) => false, |
| 484 | } |
| 485 | } |
| 486 | |
| 487 | /// Read an optional prompt-override file rooted at `config_dir`. |
| 488 | /// |
| 489 | /// Returns the file contents when it exists and is non-empty after trimming; |
| 490 | /// otherwise `None` so the caller falls back to the embedded default. Pure |
| 491 | /// over `config_dir`, so it is unit-testable without touching the global |
| 492 | /// override cells. |
| 493 | fn read_prompt_override_file(config_dir: &Path, relative: &str) -> Option<String> { |
| 494 | let path = config_dir.join(relative); |
| 495 | let raw = std::fs::read_to_string(&path).ok()?; |
| 496 | if raw.trim().is_empty() { |
| 497 | tracing::warn!( |
| 498 | target: "prompts", |
| 499 | "ignoring empty prompt override file {}", |
| 500 | path.display(), |
| 501 | ); |
| 502 | return None; |
| 503 | } |
| 504 | tracing::info!( |
| 505 | target: "prompts", |
| 506 | "loaded prompt override from {}", |
| 507 | path.display(), |
| 508 | ); |
| 509 | Some(raw) |
| 510 | } |
| 511 | |
| 512 | fn push_prompt_override_notice(message: String) { |
| 513 | if let Ok(mut notices) = PROMPT_OVERRIDE_NOTICES.lock() { |
| 514 | notices.push(message); |
| 515 | } |
| 516 | } |
| 517 | |
| 518 | pub fn take_prompt_override_notices() -> Vec<String> { |
| 519 | PROMPT_OVERRIDE_NOTICES |
| 520 | .lock() |
| 521 | .map(|mut notices| std::mem::take(&mut *notices)) |
| 522 | .unwrap_or_default() |
| 523 | } |
| 524 | |
| 525 | /// Load user prompt overrides from `config_dir` and install them through the |
| 526 | /// existing override hooks. Returns the names of the overrides that were |
| 527 | /// applied (for logging/diagnostics). |
| 528 | /// |
| 529 | /// Call once at startup, before any engine spawns, because the underlying |
| 530 | /// override cells are first-call-wins. Missing files are a no-op, preserving |
| 531 | /// the bundled defaults. |
| 532 | pub fn load_config_dir_prompt_overrides(config_dir: &Path) -> Vec<&'static str> { |
| 533 | let mut applied = Vec::new(); |
| 534 | if let Some(text) = read_prompt_override_file(config_dir, CONSTITUTION_OVERRIDE_FILE) { |
| 535 | if !base_prompt_override_opt_in() { |
| 536 | // A file exists but the user hasn't opted in. Don't silently |
| 537 | // replace the base prompt — surface the gate instead. |
| 538 | let warning = format!( |
| 539 | "Custom Constitution override found at {}/{} but {} is not set; using the bundled Constitution. Set {}=1 to opt in.", |
| 540 | config_dir.display(), |
| 541 | CONSTITUTION_OVERRIDE_FILE, |
| 542 | BASE_PROMPT_OVERRIDE_OPT_IN_ENV, |
| 543 | BASE_PROMPT_OVERRIDE_OPT_IN_ENV, |
| 544 | ); |
| 545 | tracing::warn!( |
| 546 | target: "prompts", |
| 547 | "{warning}", |
| 548 | ); |
| 549 | push_prompt_override_notice(warning); |
| 550 | } else if set_base_prompt_override(text).is_ok() { |
| 551 | applied.push("constitution"); |
| 552 | } |
| 553 | } |
| 554 | applied |
| 555 | } |
| 556 | |
| 557 | /// Resolve the Codewhale config directory and load any prompt overrides found |
| 558 | /// there. Convenience wrapper around [`load_config_dir_prompt_overrides`] for |
| 559 | /// startup wiring; silently does nothing when the config home cannot be |
| 560 | /// resolved. |
| 561 | pub fn load_prompt_overrides_from_config_home() { |
| 562 | let Ok(home) = codewhale_config::codewhale_home() else { |
| 563 | return; |
| 564 | }; |
| 565 | let applied = load_config_dir_prompt_overrides(&home); |
| 566 | if !applied.is_empty() { |
| 567 | tracing::info!( |
| 568 | target: "prompts", |
| 569 | "applied {} config-directory prompt override(s): {}", |
| 570 | applied.len(), |
| 571 | applied.join(", "), |
| 572 | ); |
| 573 | } |
| 574 | } |
| 575 | |
| 576 | fn set_prompt_override(cell: &std::sync::OnceLock<String>, s: String) -> Result<(), String> { |
| 577 | cell.set(s) |
| 578 | } |
| 579 | |
| 580 | fn effective_prompt_override<'a>( |
| 581 | cell: &'a std::sync::OnceLock<String>, |
| 582 | fallback: &'static str, |
| 583 | ) -> &'a str { |
| 584 | cell.get().map(String::as_str).unwrap_or(fallback) |
| 585 | } |
| 586 | |
| 587 | fn effective_base_prompt() -> &'static str { |
| 588 | effective_prompt_override(&BASE_PROMPT_OVERRIDE, BASE_PROMPT) |
| 589 | } |
| 590 | |
| 591 | /// Where the base-prompt bytes used by this process actually came from. |
| 592 | /// |
| 593 | /// #3928: diagnostics used to cite `crates/tui/src/prompts/text.rs`, which is |
| 594 | /// a source-tree path that does not exist on an installed binary and says |
| 595 | /// nothing about whether an override replaced the constant at startup. This |
| 596 | /// reports the runtime truth instead. |
| 597 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 598 | pub(crate) enum BasePromptOrigin { |
| 599 | /// The `BASE_PROMPT` constant compiled into this binary. |
| 600 | Bundled, |
| 601 | /// An opted-in `prompts/constitution.md` override installed at startup. |
| 602 | ConfigOverride, |
| 603 | } |
| 604 | |
| 605 | impl BasePromptOrigin { |
| 606 | /// Short, user-facing provenance label. Contains no filesystem paths. |
| 607 | pub(crate) fn label(self) -> &'static str { |
| 608 | match self { |
| 609 | Self::Bundled => "bundled in this codewhale-tui build (BASE_PROMPT, compiled in)", |
| 610 | Self::ConfigOverride => concat!( |
| 611 | "config-directory override installed at startup ", |
| 612 | "(prompts/constitution.md, opt-in enabled)" |
| 613 | ), |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | /// Runtime provenance of the base prompt for this process. |
| 619 | pub(crate) fn base_prompt_origin() -> BasePromptOrigin { |
| 620 | if BASE_PROMPT_OVERRIDE.get().is_some() { |
| 621 | BasePromptOrigin::ConfigOverride |
| 622 | } else { |
| 623 | BasePromptOrigin::Bundled |
| 624 | } |
| 625 | } |
| 626 | |
| 627 | /// The exact base-prompt bytes this process will compose into the system |
| 628 | /// prompt — the override when one is installed, the bundled constant |
| 629 | /// otherwise. |
| 630 | pub(crate) fn effective_base_prompt_text() -> &'static str { |
| 631 | effective_base_prompt() |
| 632 | } |
| 633 | |
| 634 | /// Where the effective base prompt actually comes from right now (#3928). |
| 635 | /// |
| 636 | /// Reads the same cells composition reads, so a preview cannot claim "bundled" |
| 637 | /// while an override is live. `config_dir` only supplies the path shown in the |
| 638 | /// override label; it does not decide whether an override is in effect. |
| 639 | #[must_use] |
| 640 | pub fn effective_base_prompt_source(config_dir: Option<&Path>) -> base_preview::BasePromptSource { |
| 641 | if STATIC_PROMPT_COMPOSER.get().is_some() { |
| 642 | // An embedder composer wraps or replaces the whole static layer set, so |
| 643 | // it outranks the base-prompt cell as the honest answer. |
| 644 | return base_preview::BasePromptSource::EmbedderComposer; |
| 645 | } |
| 646 | if BASE_PROMPT_OVERRIDE.get().is_some() { |
| 647 | return base_preview::BasePromptSource::ConfigOverride { |
| 648 | path: config_dir.map_or_else( |
| 649 | || CONSTITUTION_OVERRIDE_FILE.to_string(), |
| 650 | |dir| dir.join(CONSTITUTION_OVERRIDE_FILE).display().to_string(), |
| 651 | ), |
| 652 | }; |
| 653 | } |
| 654 | base_preview::BasePromptSource::Bundled |
| 655 | } |
| 656 | |
| 657 | fn effective_static_prompt_composer() -> Option<&'static StaticPromptComposer> { |
| 658 | STATIC_PROMPT_COMPOSER.get().map(Box::as_ref) |
| 659 | } |
| 660 | |
| 661 | fn effective_locale_preamble_zh_hans() -> &'static str { |
| 662 | effective_prompt_override(&LOCALE_PREAMBLE_ZH_HANS_OVERRIDE, LOCALE_PREAMBLE_ZH_HANS) |
| 663 | } |
| 664 | |
| 665 | fn effective_locale_preamble_ja() -> &'static str { |
| 666 | effective_prompt_override(&LOCALE_PREAMBLE_JA_OVERRIDE, LOCALE_PREAMBLE_JA) |
| 667 | } |
| 668 | |
| 669 | fn effective_locale_preamble_pt_br() -> &'static str { |
| 670 | effective_prompt_override(&LOCALE_PREAMBLE_PT_BR_OVERRIDE, LOCALE_PREAMBLE_PT_BR) |
| 671 | } |
| 672 | |
| 673 | fn effective_locale_preamble_vi() -> &'static str { |
| 674 | effective_prompt_override(&LOCALE_PREAMBLE_VI_OVERRIDE, LOCALE_PREAMBLE_VI) |
| 675 | } |
| 676 | |
| 677 | fn effective_locale_closer_zh_hans() -> &'static str { |
| 678 | effective_prompt_override(&LOCALE_CLOSER_ZH_HANS_OVERRIDE, LOCALE_CLOSER_ZH_HANS) |
| 679 | } |
| 680 | |
| 681 | fn effective_locale_closer_ja() -> &'static str { |
| 682 | effective_prompt_override(&LOCALE_CLOSER_JA_OVERRIDE, LOCALE_CLOSER_JA) |
| 683 | } |
| 684 | |
| 685 | fn effective_locale_closer_pt_br() -> &'static str { |
| 686 | effective_prompt_override(&LOCALE_CLOSER_PT_BR_OVERRIDE, LOCALE_CLOSER_PT_BR) |
| 687 | } |
| 688 | |
| 689 | fn effective_locale_closer_vi() -> &'static str { |
| 690 | effective_prompt_override(&LOCALE_CLOSER_VI_OVERRIDE, LOCALE_CLOSER_VI) |
| 691 | } |
| 692 | |
| 693 | pub(crate) fn effective_authority_recap() -> &'static str { |
| 694 | effective_prompt_override(&AUTHORITY_RECAP_OVERRIDE, AUTHORITY_RECAP) |
| 695 | } |
| 696 | |
| 697 | /// Optional locale-native reinforcement preamble prepended to the system |
| 698 | /// prompt when the user's UI locale is non-English. |
| 699 | /// |
| 700 | /// `constitution.md` itself stays English (single source of truth, model is |
| 701 | /// natively multilingual, prefix-cache stable across users in the same |
| 702 | /// locale). For non-English locales we prepend a short locale-native |
| 703 | /// passage so the model's first exposure to the prompt overrides the |
| 704 | /// "match user message language" English directive with an explicit |
| 705 | /// "use {locale}" instruction in the user's own writing system. Reduces |
| 706 | /// the model's reliance on inferring intent from `## Environment.lang` |
| 707 | /// — which previously got overpowered by overwhelmingly English task |
| 708 | /// context, the symptom reported in #1118 and visible in the WeChat |
| 709 | /// screenshot that prompted this change. |
| 710 | /// |
| 711 | /// The list is intentionally short (`zh-Hans`, `ja`, `pt-BR`, `vi`) even |
| 712 | /// though the TUI ships UI packs for many more locales. Other locales fall |
| 713 | /// through to `None` and get the English-only directive, which is the same |
| 714 | /// behavior as before this change; the test |
| 715 | /// `v092_locales_add_no_prompt_bookends_so_prompt_bytes_stay_stable` locks |
| 716 | /// that set so adding a UI pack never silently changes prompt bytes. |
| 717 | /// |
| 718 | /// ## Design philosophy: why a bookend, not a full translation |
| 719 | /// |
| 720 | /// Community feedback on the WeChat thread that prompted this work |
| 721 | /// pointed out — correctly — that DeepSeek V4 is a Chinese-first |
| 722 | /// multilingual model, not an English-only model with multilingual |
| 723 | /// veneer. Its tokenizer is co-trained on Chinese; `你好` typically |
| 724 | /// encodes to ~1 token, not 2 — the "Chinese is expensive in tokens" |
| 725 | /// folk wisdom from Western-LLM commentary doesn't apply here. |
| 726 | /// |
| 727 | /// The naïve translation of that argument would be: ship a fully |
| 728 | /// translated `constitution.md` per locale. We deliberately stop short of |
| 729 | /// that for v0.8.29. The reasons, ranked: |
| 730 | /// |
| 731 | /// 1. **Drift risk.** A 200+ line technical prompt has subtle |
| 732 | /// phrasing that drives subtle behavior. Every rule change has |
| 733 | /// to land in N translated copies, kept in lockstep. The class |
| 734 | /// of bug that arises (Chinese users see slightly different |
| 735 | /// agent behavior than English users) is hard to reproduce and |
| 736 | /// hard to triage from bug reports. |
| 737 | /// 2. **Cache stability.** With one English `constitution.md` and a |
| 738 | /// per-locale preamble+closer, the largest cacheable chunk |
| 739 | /// (mode prompt + project context + environment) stays |
| 740 | /// byte-stable within a session and across users in the same |
| 741 | /// locale. A fully translated per-locale `constitution.md` keeps cache |
| 742 | /// per-locale but doesn't share with English users. |
| 743 | /// 3. **Translation QA is expensive.** Each prompt-language pair |
| 744 | /// needs a native speaker reviewing tone, register, and rule |
| 745 | /// preservation. Getting it 95% right is bad, because the |
| 746 | /// missing 5% becomes silent behavior divergence. |
| 747 | /// |
| 748 | /// What we DO instead — the bookend pattern @MuMu described from |
| 749 | /// their other project — is reinforce the locale directive in |
| 750 | /// native script at BOTH ends of the prompt. The opening anchors |
| 751 | /// behavior at session start; the closing reinforcement |
| 752 | /// (`locale_reinforcement_closer`) sits at the maximum-recency |
| 753 | /// position right before the user's next message. Empirically this |
| 754 | /// is sufficient to keep `reasoning_content` in the target locale |
| 755 | /// even as English code accumulates in context turn-over-turn. |
| 756 | /// |
| 757 | /// If at some future point the bookend proves insufficient — or if |
| 758 | /// the maintenance cost of per-locale `constitution.md` files becomes |
| 759 | /// preferable to whatever's blocking it — full translation is the |
| 760 | /// natural next step. The locale tags here, the test invariants, |
| 761 | /// and the closer position would all carry over unchanged. |
| 762 | pub(crate) fn locale_reinforcement_preamble(locale_tag: &str) -> Option<&'static str> { |
| 763 | match locale_tag { |
| 764 | "zh-Hans" | "zh-CN" | "zh" => Some(effective_locale_preamble_zh_hans()), |
| 765 | "ja" | "ja-JP" => Some(effective_locale_preamble_ja()), |
| 766 | "pt-BR" | "pt" => Some(effective_locale_preamble_pt_br()), |
| 767 | "vi" | "vi-VN" => Some(effective_locale_preamble_vi()), |
| 768 | _ => None, |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | /// Locale-native closing reinforcement appended to the very end of the |
| 773 | /// system prompt — the bookend MuMu described in the WeChat thread that |
| 774 | /// prompted #1118 follow-up work. |
| 775 | /// |
| 776 | /// The opening preamble alone is not enough: as the model accumulates |
| 777 | /// English context turn-over-turn (code, error logs, search results, |
| 778 | /// file listings), the recency bias of the transformer's attention |
| 779 | /// drifts thinking back toward English even when the user keeps writing |
| 780 | /// in their own language. A closing native-script reinforcement sits at |
| 781 | /// the position closest to the user's next message — where attention |
| 782 | /// weight is highest — and re-asserts the language rule right before |
| 783 | /// the model generates `reasoning_content` for the turn. |
| 784 | /// |
| 785 | /// Like the opening preamble, English (and unknown) locales return |
| 786 | /// `None` and the system prompt is byte-identical to the pre-bookend |
| 787 | /// behavior. |
| 788 | pub(crate) fn locale_reinforcement_closer(locale_tag: &str) -> Option<&'static str> { |
| 789 | match locale_tag { |
| 790 | "zh-Hans" | "zh-CN" | "zh" => Some(effective_locale_closer_zh_hans()), |
| 791 | "ja" | "ja-JP" => Some(effective_locale_closer_ja()), |
| 792 | "pt-BR" | "pt" => Some(effective_locale_closer_pt_br()), |
| 793 | "vi" | "vi-VN" => Some(effective_locale_closer_vi()), |
| 794 | _ => None, |
| 795 | } |
| 796 | } |
| 797 | |
| 798 | const LOCALE_PREAMBLE_ZH_HANS: &str = "## 语言要求\n\n\ |
| 799 | 你正在 codewhale 中运行。无论任务上下文（代码、错误日志、文件名）\ |
| 800 | 是英文，无论系统提示的其余部分是英文，你都必须用简体中文进行 \ |
| 801 | `reasoning_content`（内部思考）和最终回复。代码、文件路径、工具名称\ |
| 802 | （例如 `File`、`Bash`）、环境变量、命令行参数和 URL \ |
| 803 | 保持原样 —— 只有自然语言散文要切换到简体中文。\n\n\ |
| 804 | 如果用户在会话中切换到另一种语言，从下一轮开始跟随切换。\ |
| 805 | 如果用户明确要求（例如 \"think in English\"），则覆盖此规则。"; |
| 806 | |
| 807 | const LOCALE_PREAMBLE_JA: &str = "## 言語要件\n\n\ |
| 808 | codewhale を実行しています。タスクコンテキスト（コード、エラーログ、\ |
| 809 | ファイル名）が英語であっても、システムプロンプトの他の部分が英語で\ |
| 810 | あっても、`reasoning_content`（内部思考）と最終的な返信は日本語で\ |
| 811 | 行ってください。コード、ファイルパス、ツール名（例：`File`、\ |
| 812 | `Bash`）、環境変数、コマンドライン引数、URL は元のまま —— \ |
| 813 | 自然言語の文章のみ日本語に切り替えます。\n\n\ |
| 814 | ユーザーがセッション中に別の言語に切り替えた場合は、次のターンから\ |
| 815 | それに従ってください。ユーザーが明示的に要求した場合（例：\ |
| 816 | \"think in English\"）はこのルールを上書きします。"; |
| 817 | |
| 818 | const LOCALE_PREAMBLE_PT_BR: &str = "## Requisito de Idioma\n\n\ |
| 819 | Você está rodando dentro do codewhale. Escreva tanto \ |
| 820 | `reasoning_content` (seu pensamento interno) quanto a resposta final \ |
| 821 | em português do Brasil, mesmo quando o contexto da tarefa (código, \ |
| 822 | logs de erro, nomes de arquivos) estiver em inglês e mesmo quando o \ |
| 823 | resto do system prompt for em inglês. Mantenha código, caminhos de \ |
| 824 | arquivos, nomes de ferramentas (por exemplo `File`, \ |
| 825 | `Bash`), variáveis de ambiente, flags de linha de comando e \ |
| 826 | URLs no formato original — apenas a prosa em linguagem natural muda \ |
| 827 | para português do Brasil.\n\n\ |
| 828 | Se o usuário mudar de idioma no meio da sessão, mude no próximo turno. \ |
| 829 | Se o usuário pedir explicitamente (por exemplo, \"think in English\"), \ |
| 830 | isso sobrescreve esta regra."; |
| 831 | |
| 832 | // ── Closing bookends (appended to the very end of the system prompt) ── |
| 833 | |
| 834 | const LOCALE_CLOSER_ZH_HANS: &str = "## 语言再次提醒\n\n\ |
| 835 | **重要：你的 `reasoning_content`（内部思考）和最终回复必须保持简体中文。** \ |
| 836 | 无论你在这次会话中读到了多少英文代码、错误日志或文档，无论项目上下文 \ |
| 837 | 是英文，思考过程不能漂移到英文。这是会话级硬性要求 —— 用户的语言决定 \ |
| 838 | 你的语言，与上下文中累积的英文内容无关。除非用户明确要求切换（例如 \ |
| 839 | \"think in English\"），否则继续用简体中文思考和回答。"; |
| 840 | |
| 841 | const LOCALE_CLOSER_JA: &str = "## 言語再確認\n\n\ |
| 842 | **重要：`reasoning_content`（内部思考）と最終的な返信は日本語で行ってください。** \ |
| 843 | このセッションで読み込んだ英語のコード、エラーログ、ドキュメントの量に \ |
| 844 | 関係なく、プロジェクトコンテキストが英語であっても、思考プロセスを \ |
| 845 | 英語に逸らさないでください。これはセッションレベルの厳格な要件であり、 \ |
| 846 | ユーザーの言語があなたの言語を決定します。ユーザーが明示的に切り替えを \ |
| 847 | 要求しない限り（例：\"think in English\"）、日本語で思考し、回答し続けて \ |
| 848 | ください。"; |
| 849 | |
| 850 | const LOCALE_CLOSER_PT_BR: &str = "## Reforço de Idioma\n\n\ |
| 851 | **Importante: seu `reasoning_content` (pensamento interno) e a resposta \ |
| 852 | final devem permanecer em português do Brasil.** Independentemente de \ |
| 853 | quanto código em inglês, logs de erro ou documentação você ler nesta \ |
| 854 | sessão, e independentemente de o contexto do projeto ser em inglês, o \ |
| 855 | processo de pensamento não pode derivar para o inglês. Este é um \ |
| 856 | requisito rígido em nível de sessão — o idioma do usuário define seu \ |
| 857 | idioma. A menos que o usuário peça explicitamente a troca (por exemplo, \ |
| 858 | \"think in English\"), continue pensando e respondendo em português do \ |
| 859 | Brasil."; |
| 860 | |
| 861 | const LOCALE_PREAMBLE_VI: &str = "## Yêu cầu ngôn ngữ\n\n\ |
| 862 | Bạn đang chạy trong codewhale. Cho dù ngữ cảnh tác vụ (mã nguồn, nhật ký lỗi, tên tệp) \ |
| 863 | là tiếng Anh, cho dù phần còn lại của system prompt là tiếng Anh, bạn đều phải sử dụng \ |
| 864 | tiếng Việt cho phần `reasoning_content` (suy nghĩ nội bộ) và câu trả lời cuối cùng. Các từ \ |
| 865 | mã nguồn, đường dẫn tệp, tên công cụ (ví dụ `File`, `Bash`), biến môi trường, \ |
| 866 | tham số dòng lệnh và URL giữ nguyên dạng gốc —— chỉ các văn bản giải thích bằng ngôn ngữ \ |
| 867 | tự nhiên mới được chuyển sang tiếng Việt.\n\n\ |
| 868 | Nếu người dùng chuyển sang ngôn ngữ khác trong phiên làm việc, hãy chuyển theo từ lượt tiếp theo. \ |
| 869 | Nếu người dùng yêu cầu rõ ràng (ví dụ \"think in English\"), hãy ghi đè quy tắc này."; |
| 870 | |
| 871 | const LOCALE_CLOSER_VI: &str = "## Nhắc nhở ngôn ngữ một lần nữa\n\n\ |
| 872 | **Quan trọng: phần `reasoning_content` (suy nghĩ nội bộ) và phản hồi cuối cùng của bạn phải được viết bằng tiếng Việt.** \ |
| 873 | Dù bạn có đọc bao nhiêu mã nguồn tiếng Anh, nhật ký lỗi hay tài liệu trong phiên làm việc này, và dù ngữ cảnh \ |
| 874 | dự án có là tiếng Anh, quá trình suy nghĩ của bạn cũng không được chuyển sang tiếng Anh. Đây là yêu cầu cứng \ |
| 875 | ở cấp phiên làm việc —— ngôn ngữ của người dùng quyết định ngôn ngữ của bạn, không phụ thuộc vào nội dung tiếng Anh \ |
| 876 | tích lũy trong ngữ cảnh. Trừ khi người dùng yêu cầu rõ ràng việc chuyển đổi (ví dụ \"think in English\"), \ |
| 877 | hãy tiếp tục suy nghĩ và trả lời bằng tiếng Việt."; |
| 878 | |
| 879 | // ── Personality selection ───────────────────────────────────────────── |
| 880 | |
| 881 | /// Which personality overlay to apply. Tone is folded into the constitutional |
| 882 | /// preamble, so this is a compile-time marker carried through the static-prompt |
| 883 | /// composer context rather than a separate overlay. |
| 884 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 885 | pub enum Personality { |
| 886 | /// Cool, spatial, reserved — the default and only shipped personality. |
| 887 | Calm, |
| 888 | } |
| 889 | |
| 890 | // ── Composition ─────────────────────────────────────────────────────── |
| 891 | |
| 892 | /// Substitute the model id for embedder-supplied prompt overrides that still |
| 893 | /// template it. The bundled constitution is deliberately model-agnostic and |
| 894 | /// carries no model-fact placeholders. |
| 895 | fn apply_model_template( |
| 896 | prompt: &str, |
| 897 | model_id: &str, |
| 898 | _context_window_override: Option<u32>, |
| 899 | ) -> String { |
| 900 | prompt.replace("{model_id}", model_id) |
| 901 | } |
| 902 | |
| 903 | /// Authority recap block — appended at the end of the system prompt, |
| 904 | /// just before the user's first message. Uses recency bias constructively |
| 905 | /// without restating ranks: precedence is stated only in `BASE_PROMPT` |
| 906 | /// § Whose word wins (#4777). |
| 907 | const AUTHORITY_RECAP: &str = "\ |
| 908 | ## Authority Recap |
| 909 | |
| 910 | Codewhale's constitution governs your behavior. Ground truth underlies the |
| 911 | whole list: the user may override a fact, but no one may invent one. When |
| 912 | guidance conflicts, consult ### Whose word wins — that is the only place |
| 913 | precedence is stated."; |
| 914 | |
| 915 | pub(crate) fn compose_prompt_with_approval_model_and_shell( |
| 916 | personality: Personality, |
| 917 | model_id: &str, |
| 918 | ) -> String { |
| 919 | let default_layers = compose_default_static_layers(personality, model_id); |
| 920 | apply_static_prompt_composer( |
| 921 | effective_static_prompt_composer(), |
| 922 | personality, |
| 923 | model_id, |
| 924 | &default_layers, |
| 925 | ) |
| 926 | } |
| 927 | |
| 928 | pub(crate) fn compose_default_static_layers(_personality: Personality, model_id: &str) -> String { |
| 929 | compose_default_static_layers_with_context(model_id, None) |
| 930 | } |
| 931 | |
| 932 | fn compose_default_static_layers_with_context( |
| 933 | model_id: &str, |
| 934 | context_window_override: Option<u32>, |
| 935 | ) -> String { |
| 936 | // Personality is folded into the constitutional preamble/articles — no |
| 937 | // separate overlay is appended. Language and output rules are split into |
| 938 | // their own static segments so the 0.9.0 constitution stays compact. |
| 939 | let layers = format!( |
| 940 | "{}\n\n{}\n\n{}", |
| 941 | effective_base_prompt().trim(), |
| 942 | LANGUAGE_PROMPT.trim(), |
| 943 | OUTPUT_PROMPT.trim() |
| 944 | ); |
| 945 | apply_model_template(&layers, model_id, context_window_override) |
| 946 | } |
| 947 | |
| 948 | /// Mode doctrine overlay for the stable prefix. |
| 949 | /// |
| 950 | /// Single mapping shared by prompt composition and the engine's mode-change |
| 951 | /// invalidation check, so the two can never disagree about which text a mode |
| 952 | /// ships (#4780). |
| 953 | pub(crate) fn mode_doctrine(mode: crate::tui::app::AppMode) -> &'static str { |
| 954 | use crate::tui::app::AppMode; |
| 955 | |
| 956 | match mode { |
| 957 | AppMode::Agent | AppMode::Auto | AppMode::Yolo => AGENT_MODE, |
| 958 | AppMode::Plan => PLAN_MODE, |
| 959 | AppMode::Operate => OPERATE_MODE, |
| 960 | } |
| 961 | } |
| 962 | |
| 963 | fn apply_static_prompt_composer( |
| 964 | composer: Option<&StaticPromptComposer>, |
| 965 | personality: Personality, |
| 966 | model_id: &str, |
| 967 | default_layers: &str, |
| 968 | ) -> String { |
| 969 | match composer { |
| 970 | Some(composer) => composer(&StaticPromptCtx { |
| 971 | model_id, |
| 972 | personality, |
| 973 | default_layers, |
| 974 | }), |
| 975 | None => default_layers.to_string(), |
| 976 | } |
| 977 | } |
| 978 | |
| 979 | // The full base prompt is always used; effective tool availability is enforced |
| 980 | // by the tool catalog and execution layer rather than by mutating message[0]. |
| 981 | |
| 982 | // ── Public API ──────────────────────────────────────────────────────── |
| 983 | |
| 984 | /// Get the system prompt for a specific mode with project context. |
| 985 | pub fn system_prompt_for_mode_with_context( |
| 986 | workspace: &Path, |
| 987 | working_set_summary: Option<&str>, |
| 988 | ) -> SystemPrompt { |
| 989 | system_prompt_for_mode_with_context_and_skills(workspace, working_set_summary, None, None, None) |
| 990 | } |
| 991 | |
| 992 | /// Get the system prompt for a specific mode with project and skills context. |
| 993 | /// |
| 994 | /// **Volatile-content-last invariant.** Blocks are appended in order from |
| 995 | /// most-static to most-volatile so DeepSeek's KV prefix cache hits the |
| 996 | /// longest possible byte prefix turn-over-turn: |
| 997 | /// |
| 998 | /// 1. mode prompt (compile-time constant) |
| 999 | /// 2. project context / fallback (workspace-static) |
| 1000 | /// 3. skills block (skills-dir-static) |
| 1001 | /// 4. `## Core Execution` (compile-time constant) |
| 1002 | /// 5. compaction relay template (compile-time constant) |
| 1003 | /// 6. relay block — file-backed; rewritten by `/compact` and on exit |
| 1004 | /// |
| 1005 | /// Anything appended after a volatile block forfeits the cache for the rest |
| 1006 | /// of the request. New blocks belong above the relay boundary unless they |
| 1007 | /// themselves are turn-volatile. Working-set metadata is now injected into the |
| 1008 | /// latest user message as per-turn metadata instead of this system prompt. |
| 1009 | pub fn system_prompt_for_mode_with_context_and_skills( |
| 1010 | workspace: &Path, |
| 1011 | working_set_summary: Option<&str>, |
| 1012 | skills_dir: Option<&Path>, |
| 1013 | instructions: Option<&[InstructionSource]>, |
| 1014 | user_memory_block: Option<&str>, |
| 1015 | ) -> SystemPrompt { |
| 1016 | system_prompt_for_mode_with_context_skills_and_session( |
| 1017 | workspace, |
| 1018 | working_set_summary, |
| 1019 | skills_dir, |
| 1020 | instructions, |
| 1021 | PromptSessionContext { |
| 1022 | user_memory_block, |
| 1023 | goal_objective: None, |
| 1024 | project_context_pack_enabled: false, |
| 1025 | locale_tag: "en", |
| 1026 | translation_enabled: false, |
| 1027 | model_id: "codewhale", |
| 1028 | context_window_override: None, |
| 1029 | verbosity: None, |
| 1030 | skills_scan_codewhale_only: false, |
| 1031 | plugin_registry: None, |
| 1032 | mode: crate::tui::app::AppMode::Agent, |
| 1033 | }, |
| 1034 | ) |
| 1035 | } |
| 1036 | |
| 1037 | pub fn system_prompt_for_mode_with_context_skills_and_session( |
| 1038 | workspace: &Path, |
| 1039 | _working_set_summary: Option<&str>, |
| 1040 | skills_dir: Option<&Path>, |
| 1041 | instructions: Option<&[InstructionSource]>, |
| 1042 | session_context: PromptSessionContext<'_>, |
| 1043 | ) -> SystemPrompt { |
| 1044 | system_prompt_for_mode_with_context_skills_session_and_approval( |
| 1045 | workspace, |
| 1046 | _working_set_summary, |
| 1047 | skills_dir, |
| 1048 | instructions, |
| 1049 | session_context, |
| 1050 | ) |
| 1051 | } |
| 1052 | |
| 1053 | pub fn system_prompt_for_mode_with_context_skills_session_and_approval( |
| 1054 | workspace: &Path, |
| 1055 | _working_set_summary: Option<&str>, |
| 1056 | skills_dir: Option<&Path>, |
| 1057 | instructions: Option<&[InstructionSource]>, |
| 1058 | session_context: PromptSessionContext<'_>, |
| 1059 | ) -> SystemPrompt { |
| 1060 | let default_layers = compose_default_static_layers_with_context( |
| 1061 | session_context.model_id, |
| 1062 | session_context.context_window_override, |
| 1063 | ); |
| 1064 | let composed = apply_static_prompt_composer( |
| 1065 | effective_static_prompt_composer(), |
| 1066 | Personality::Calm, |
| 1067 | session_context.model_id, |
| 1068 | &default_layers, |
| 1069 | ); |
| 1070 | |
| 1071 | // Mode doctrine is layer 1 of the stable prefix (#4780). It sits above the |
| 1072 | // constitution's own layers so the model reads "which mode am I in" before |
| 1073 | // the general rules it modulates, and it ships exactly once per prefix |
| 1074 | // rather than per user message. |
| 1075 | let mode_prompt = format!( |
| 1076 | "{}\n\n{}", |
| 1077 | mode_doctrine(session_context.mode).trim(), |
| 1078 | composed.trim_start() |
| 1079 | ); |
| 1080 | |
| 1081 | // Load project context from workspace |
| 1082 | let project_context = load_project_context_with_parents(workspace); |
| 1083 | |
| 1084 | // 0. Locale-native reinforcement preamble (#1118 follow-up). When the |
| 1085 | // user's UI locale is non-English we prepend a short native-script |
| 1086 | // passage so the model's first exposure to the prompt is an explicit |
| 1087 | // "think and reply in {locale}" directive in the user's own writing |
| 1088 | // system — defeats the "task context is English, so the model thinks |
| 1089 | // in English even though `lang: zh-Hans` is set" failure mode that |
| 1090 | // PR #1398 partially addressed. English (and unknown) locales get |
| 1091 | // `None` and keep the previous behavior unchanged. |
| 1092 | let preamble = locale_reinforcement_preamble(session_context.locale_tag); |
| 1093 | |
| 1094 | // 1–2. Mode prompt + project context. |
| 1095 | // `load_project_context_with_parents` generates an in-memory bounded |
| 1096 | // overview when no context file exists, so the fallback should usually be |
| 1097 | // available without writing project-local files. |
| 1098 | let mut full_prompt = if let Some(project_block) = project_context.as_system_block() { |
| 1099 | format!("{mode_prompt}\n\n{project_block}") |
| 1100 | } else { |
| 1101 | // Extremely unlikely: context generation failed (e.g. filesystem error). |
| 1102 | // Use mode prompt alone rather than panic. |
| 1103 | tracing::warn!("No project context available and auto-generation failed"); |
| 1104 | mode_prompt |
| 1105 | }; |
| 1106 | |
| 1107 | if let Some(preamble) = preamble { |
| 1108 | full_prompt = format!("{preamble}\n\n{full_prompt}"); |
| 1109 | } |
| 1110 | |
| 1111 | if let Some(user_constitution_block) = load_user_constitution_block() { |
| 1112 | full_prompt = format!("{full_prompt}\n\n{user_constitution_block}"); |
| 1113 | } |
| 1114 | |
| 1115 | if session_context.project_context_pack_enabled |
| 1116 | && let Some(pack) = crate::project_context::generate_project_context_pack(workspace) |
| 1117 | { |
| 1118 | full_prompt = format!("{full_prompt}\n\n{pack}"); |
| 1119 | } |
| 1120 | |
| 1121 | // 2.3a. Translation output instruction — when enabled, instruct |
| 1122 | // the model to respond in the resolved session locale. Stays |
| 1123 | // above the volatile-content boundary because it's a per-session |
| 1124 | // flag, not a per-turn one: enabling `/translate` is a session |
| 1125 | // toggle, so the prompt-prefix bytes don't drift turn-over-turn. |
| 1126 | if session_context.translation_enabled { |
| 1127 | full_prompt = format!( |
| 1128 | "{full_prompt}\n\n{}", |
| 1129 | translation_output_instruction(session_context.locale_tag) |
| 1130 | ); |
| 1131 | } |
| 1132 | |
| 1133 | if is_concise_verbosity(session_context.verbosity) { |
| 1134 | full_prompt = format!( |
| 1135 | "{full_prompt}\n\n{}", |
| 1136 | concise_output_discipline_instruction() |
| 1137 | ); |
| 1138 | } |
| 1139 | |
| 1140 | // 3. Skills block. #432: default discovery walks every compatible |
| 1141 | // workspace/global skill directory so skills installed for other AI-tool |
| 1142 | // conventions show up in the catalogue. Users can opt into a Codewhale-only |
| 1143 | // scan with `[skills] scan_codewhale_only = true`. When an explicit |
| 1144 | // `skills_dir` is configured, union it with the workspace view instead of |
| 1145 | // treating it as a fallback; the workspace view often returns Some and |
| 1146 | // would otherwise shadow the configured directory entirely. |
| 1147 | let skill_discovery_mode = crate::skills::SkillDiscoveryMode::from_codewhale_only( |
| 1148 | session_context.skills_scan_codewhale_only, |
| 1149 | ); |
| 1150 | let skills_block = match skills_dir { |
| 1151 | Some(dir) => { |
| 1152 | crate::skills::render_available_skills_context_for_workspace_and_dir_with_mode_and_plugins( |
| 1153 | workspace, |
| 1154 | dir, |
| 1155 | skill_discovery_mode, |
| 1156 | session_context.locale_tag, |
| 1157 | session_context.plugin_registry, |
| 1158 | ) |
| 1159 | } |
| 1160 | None => crate::skills::render_available_skills_context_for_workspace_with_mode_and_plugins( |
| 1161 | workspace, |
| 1162 | skill_discovery_mode, |
| 1163 | session_context.locale_tag, |
| 1164 | session_context.plugin_registry, |
| 1165 | ), |
| 1166 | }; |
| 1167 | if let Some(block) = skills_block { |
| 1168 | full_prompt = format!("{full_prompt}\n\n{block}"); |
| 1169 | } |
| 1170 | |
| 1171 | // 4. Lean, runtime-only coding discipline. Context pressure, prompt-cache |
| 1172 | // accounting, footer presentation, and automatic compaction are host |
| 1173 | // responsibilities; teaching their UI to the model dilutes the task. |
| 1174 | full_prompt.push_str("\n\n"); |
| 1175 | full_prompt.push_str(CORE_EXECUTION_PROFILE_PROMPT.trim()); |
| 1176 | |
| 1177 | // The compaction/relay format is action-specific context. Automatic |
| 1178 | // compaction owns its structured successor brief, while `/relay` appends |
| 1179 | // `COMPACT_TEMPLATE` to that command's user message. Keeping the template |
| 1180 | // out of every fresh session saves a stable-prefix block without removing |
| 1181 | // the capability. |
| 1182 | |
| 1183 | // ── Volatile-content boundary → WorldState fragments ────────────────── |
| 1184 | // Constitution (`full_prompt`) stays the cache-stable Blocks[0] prefix. |
| 1185 | // Everything below drifts mid-session and is assembled as marked |
| 1186 | // WorldState fragments so an env/memory/goal/handoff change can |
| 1187 | // `render_diff` without rebuilding unrelated material. |
| 1188 | |
| 1189 | // Workspace fragment: environment + mid-session memory/goal facts. |
| 1190 | let mut workspace_parts = vec![render_environment_block( |
| 1191 | workspace, |
| 1192 | session_context.locale_tag, |
| 1193 | )]; |
| 1194 | if let Some(memory_block) = session_context.user_memory_block |
| 1195 | && !memory_block.trim().is_empty() |
| 1196 | { |
| 1197 | workspace_parts.push(format!("{memory_block}\n\n{MEMORY_GUIDANCE}")); |
| 1198 | } |
| 1199 | if let Some(harness_block) = crate::continual_harness::prompt_block(workspace) { |
| 1200 | workspace_parts.push(harness_block); |
| 1201 | } |
| 1202 | if let Some(goal_objective) = session_context.goal_objective |
| 1203 | && !goal_objective.trim().is_empty() |
| 1204 | { |
| 1205 | workspace_parts.push(format!( |
| 1206 | "## Current Goal\n\n<session_goal>\n{}\n</session_goal>", |
| 1207 | goal_objective.trim() |
| 1208 | )); |
| 1209 | } |
| 1210 | let workspace_body = workspace_parts.join("\n\n"); |
| 1211 | |
| 1212 | // Permissions fragment: configured `instructions = [...]` files (#454). |
| 1213 | let permissions_body = instructions.and_then(render_instructions_block); |
| 1214 | |
| 1215 | // Route fragment: verbosity / translation posture (the model id was |
| 1216 | // removed by the turn-meta diet — it is telemetry the model cannot act on). |
| 1217 | let route_body = render_route_fragment(&session_context); |
| 1218 | |
| 1219 | // Token-budget / continuity fragment: prior-session handoff relay. |
| 1220 | let token_budget_body = load_handoff_block(workspace); |
| 1221 | |
| 1222 | let world_state = world_state_from_session_facts( |
| 1223 | Some(workspace_body.as_str()), |
| 1224 | permissions_body.as_deref(), |
| 1225 | Some(route_body.as_str()), |
| 1226 | None, // AgentTopology is updated by runtime callers when available. |
| 1227 | None, // Skills stay in the constitution prefix (skills-dir-static). |
| 1228 | token_budget_body.as_deref(), |
| 1229 | ); |
| 1230 | |
| 1231 | let mut blocks = crate::model_context::WorldStateSnapshot { |
| 1232 | constitution: full_prompt, |
| 1233 | world_state, |
| 1234 | } |
| 1235 | .to_system_blocks(); |
| 1236 | |
| 1237 | // Trailers keep recency bias after WorldState: authority, then locale. |
| 1238 | blocks.push(SystemBlock { |
| 1239 | block_type: "text".to_string(), |
| 1240 | text: effective_authority_recap().trim().to_string(), |
| 1241 | cache_control: None, |
| 1242 | }); |
| 1243 | if let Some(closer) = locale_reinforcement_closer(session_context.locale_tag) { |
| 1244 | blocks.push(SystemBlock { |
| 1245 | block_type: "text".to_string(), |
| 1246 | text: closer.trim().to_string(), |
| 1247 | cache_control: None, |
| 1248 | }); |
| 1249 | } |
| 1250 | |
| 1251 | SystemPrompt::Blocks(blocks) |
| 1252 | } |
| 1253 | |
| 1254 | /// Flatten a system prompt to joined text (tests + debug inspectors). |
| 1255 | #[must_use] |
| 1256 | pub fn system_prompt_flat_text(prompt: &SystemPrompt) -> String { |
| 1257 | match prompt { |
| 1258 | SystemPrompt::Text(text) => text.clone(), |
| 1259 | SystemPrompt::Blocks(blocks) => blocks |
| 1260 | .iter() |
| 1261 | .map(|block| block.text.as_str()) |
| 1262 | .collect::<Vec<_>>() |
| 1263 | .join("\n\n"), |
| 1264 | } |
| 1265 | } |
| 1266 | |
| 1267 | fn render_route_fragment(session_context: &PromptSessionContext<'_>) -> String { |
| 1268 | let verbosity = session_context |
| 1269 | .verbosity |
| 1270 | .map(str::trim) |
| 1271 | .filter(|value| !value.is_empty()) |
| 1272 | .unwrap_or("default"); |
| 1273 | format!( |
| 1274 | "verbosity: {verbosity}\ntranslation: {}", |
| 1275 | if session_context.translation_enabled { |
| 1276 | "on" |
| 1277 | } else { |
| 1278 | "off" |
| 1279 | }, |
| 1280 | ) |
| 1281 | } |
| 1282 | |
| 1283 | /// Build a WorldState from the common volatile session facts. |
| 1284 | /// |
| 1285 | /// Does not load constitution — callers keep that as the stable base. |
| 1286 | pub fn world_state_from_session_facts( |
| 1287 | workspace_body: Option<&str>, |
| 1288 | permissions_body: Option<&str>, |
| 1289 | route_body: Option<&str>, |
| 1290 | agent_topology_body: Option<&str>, |
| 1291 | skills_tools_body: Option<&str>, |
| 1292 | token_budget_body: Option<&str>, |
| 1293 | ) -> crate::model_context::WorldState { |
| 1294 | let mut state = crate::model_context::WorldState::new(); |
| 1295 | if let Some(body) = workspace_body.filter(|s| !s.trim().is_empty()) { |
| 1296 | state = state.with_workspace(body); |
| 1297 | } |
| 1298 | if let Some(body) = permissions_body.filter(|s| !s.trim().is_empty()) { |
| 1299 | state = state.with_permissions(body); |
| 1300 | } |
| 1301 | if let Some(body) = route_body.filter(|s| !s.trim().is_empty()) { |
| 1302 | state = state.with_route(body); |
| 1303 | } |
| 1304 | if let Some(body) = agent_topology_body.filter(|s| !s.trim().is_empty()) { |
| 1305 | state = state.with_agent_topology(body); |
| 1306 | } |
| 1307 | if let Some(body) = skills_tools_body.filter(|s| !s.trim().is_empty()) { |
| 1308 | state = state.with_skills_tools(body); |
| 1309 | } |
| 1310 | if let Some(body) = token_budget_body.filter(|s| !s.trim().is_empty()) { |
| 1311 | state = state.with_token_budget(body); |
| 1312 | } |
| 1313 | state |
| 1314 | } |
| 1315 | |
| 1316 | #[cfg(test)] |
| 1317 | mod tests { |
| 1318 | // Don't assert on prose. If you wouldn't fail a code review for |
| 1319 | // changing the wording, don't fail a test for it. |
| 1320 | use super::*; |
| 1321 | use crate::tools::apply_patch::ApplyPatchTool; |
| 1322 | use crate::tools::file::{EditFileTool, WriteFileTool}; |
| 1323 | use crate::tools::handle::HandleReadTool; |
| 1324 | use crate::tools::rlm::RlmTool; |
| 1325 | use crate::tools::shell::BashTool; |
| 1326 | use crate::tools::spec::ToolSpec; |
| 1327 | use tempfile::tempdir; |
| 1328 | |
| 1329 | /// Discriminator unique to the injected relay block (not present in the |
| 1330 | /// agent prompt's own discussion of the convention). |
| 1331 | const HANDOFF_BLOCK_MARKER: &str = "left a relay artifact at `.codewhale/handoff.md`"; |
| 1332 | |
| 1333 | // Config-directory prompt override resolution (#3638). These exercise the |
| 1334 | // pure file resolver only; the global install path is intentionally not |
| 1335 | // unit-tested here because `set_base_prompt_override` writes a process-wide |
| 1336 | // `OnceLock` that would leak into sibling tests (same reason |
| 1337 | // `prompt_override_storage_reports_duplicate_sets` uses a local cell). |
| 1338 | |
| 1339 | #[test] |
| 1340 | fn config_override_reads_present_nonempty_file() { |
| 1341 | let tmp = tempdir().expect("tempdir"); |
| 1342 | let prompts_dir = tmp.path().join("prompts"); |
| 1343 | std::fs::create_dir_all(&prompts_dir).expect("mkdir"); |
| 1344 | std::fs::write( |
| 1345 | prompts_dir.join("constitution.md"), |
| 1346 | "You are a long-form writing companion.\n", |
| 1347 | ) |
| 1348 | .expect("write override"); |
| 1349 | |
| 1350 | let got = read_prompt_override_file(tmp.path(), CONSTITUTION_OVERRIDE_FILE); |
| 1351 | assert_eq!( |
| 1352 | got.as_deref(), |
| 1353 | Some("You are a long-form writing companion.\n") |
| 1354 | ); |
| 1355 | } |
| 1356 | |
| 1357 | #[test] |
| 1358 | fn config_override_absent_file_falls_back() { |
| 1359 | let tmp = tempdir().expect("tempdir"); |
| 1360 | // No prompts/ directory at all → None so the embedded constant is used. |
| 1361 | assert!(read_prompt_override_file(tmp.path(), CONSTITUTION_OVERRIDE_FILE).is_none()); |
| 1362 | } |
| 1363 | |
| 1364 | #[test] |
| 1365 | fn config_override_requires_explicit_opt_in() { |
| 1366 | // A present, non-empty override file must NOT replace the base prompt |
| 1367 | // unless the explicit opt-in flag is set. This test drains the shared |
| 1368 | // process-global PROMPT_OVERRIDE_NOTICES queue, so it must serialize |
| 1369 | // against the sibling test that also touches it |
| 1370 | // (`tui::ui::tests::prompt_override_notice_surfaces_in_transcript_and_toast`); |
| 1371 | // both take `lock_test_env()` for mutual exclusion under the multi- |
| 1372 | // threaded test binary. |
| 1373 | let _env_guard = crate::test_support::lock_test_env(); |
| 1374 | let tmp = tempdir().expect("tempdir"); |
| 1375 | let prompts_dir = tmp.path().join("prompts"); |
| 1376 | std::fs::create_dir_all(&prompts_dir).expect("mkdir"); |
| 1377 | std::fs::write( |
| 1378 | prompts_dir.join("constitution.md"), |
| 1379 | "You are a long-form writing companion.\n", |
| 1380 | ) |
| 1381 | .expect("write override"); |
| 1382 | |
| 1383 | // The resolver still finds the file... |
| 1384 | assert!(read_prompt_override_file(tmp.path(), CONSTITUTION_OVERRIDE_FILE).is_some()); |
| 1385 | // ...but without the opt-in flag, nothing is applied. |
| 1386 | if std::env::var(BASE_PROMPT_OVERRIDE_OPT_IN_ENV).is_err() { |
| 1387 | let _ = take_prompt_override_notices(); |
| 1388 | assert!( |
| 1389 | load_config_dir_prompt_overrides(tmp.path()).is_empty(), |
| 1390 | "override must require the explicit opt-in flag, not just a file" |
| 1391 | ); |
| 1392 | let notices = take_prompt_override_notices(); |
| 1393 | assert!( |
| 1394 | notices |
| 1395 | .iter() |
| 1396 | .any(|notice| notice.contains(BASE_PROMPT_OVERRIDE_OPT_IN_ENV) |
| 1397 | && notice.contains("using the bundled Constitution")), |
| 1398 | "gated override should record a visible notice, got {notices:?}" |
| 1399 | ); |
| 1400 | } |
| 1401 | } |
| 1402 | |
| 1403 | #[test] |
| 1404 | fn config_override_empty_file_is_ignored() { |
| 1405 | let tmp = tempdir().expect("tempdir"); |
| 1406 | let prompts_dir = tmp.path().join("prompts"); |
| 1407 | std::fs::create_dir_all(&prompts_dir).expect("mkdir"); |
| 1408 | std::fs::write(prompts_dir.join("constitution.md"), " \n\t\n").expect("write blank"); |
| 1409 | |
| 1410 | // Whitespace-only overrides are treated as absent so a stray empty file |
| 1411 | // can't silently blank the system prompt. |
| 1412 | assert!(read_prompt_override_file(tmp.path(), CONSTITUTION_OVERRIDE_FILE).is_none()); |
| 1413 | } |
| 1414 | |
| 1415 | #[test] |
| 1416 | fn prompt_override_storage_reports_duplicate_sets() { |
| 1417 | let cell = std::sync::OnceLock::new(); |
| 1418 | |
| 1419 | assert_eq!(effective_prompt_override(&cell, "fallback"), "fallback"); |
| 1420 | assert!(set_prompt_override(&cell, "first".to_string()).is_ok()); |
| 1421 | assert_eq!(effective_prompt_override(&cell, "fallback"), "first"); |
| 1422 | assert_eq!( |
| 1423 | set_prompt_override(&cell, "second".to_string()), |
| 1424 | Err("second".to_string()) |
| 1425 | ); |
| 1426 | assert_eq!(effective_prompt_override(&cell, "fallback"), "first"); |
| 1427 | } |
| 1428 | |
| 1429 | #[test] |
| 1430 | fn static_prompt_composer_unset_keeps_default_layers_byte_identical() { |
| 1431 | let default_layers = compose_default_static_layers(Personality::Calm, "deepseek-v4-flash"); |
| 1432 | let composed = apply_static_prompt_composer( |
| 1433 | None, |
| 1434 | Personality::Calm, |
| 1435 | "deepseek-v4-flash", |
| 1436 | &default_layers, |
| 1437 | ); |
| 1438 | |
| 1439 | assert_byte_identical("unset static prompt composer", &default_layers, &composed); |
| 1440 | } |
| 1441 | |
| 1442 | #[test] |
| 1443 | fn static_prompt_composer_receives_context_and_replaces_layers() { |
| 1444 | let default_layers = compose_default_static_layers(Personality::Calm, "deepseek-v4-pro"); |
| 1445 | let composer: Box<StaticPromptComposer> = Box::new(|ctx| { |
| 1446 | assert_eq!(ctx.model_id, "deepseek-v4-pro"); |
| 1447 | assert_eq!(ctx.personality, Personality::Calm); |
| 1448 | // The 0.9.0 core is model-agnostic ("You are Codewhale") and |
| 1449 | // folds tone in — no per-model id line, no separate personality |
| 1450 | // section in default_layers. |
| 1451 | assert!(ctx.default_layers.contains("You are Codewhale")); |
| 1452 | assert!( |
| 1453 | ctx.default_layers |
| 1454 | .contains("Take the work seriously. Don't take") |
| 1455 | ); |
| 1456 | assert!(!ctx.default_layers.contains("## Core Tool Taxonomy")); |
| 1457 | assert!(!ctx.default_layers.contains("Approval Policy")); |
| 1458 | "embedder static prompt".to_string() |
| 1459 | }); |
| 1460 | |
| 1461 | let composed = apply_static_prompt_composer( |
| 1462 | Some(composer.as_ref()), |
| 1463 | Personality::Calm, |
| 1464 | "deepseek-v4-pro", |
| 1465 | &default_layers, |
| 1466 | ); |
| 1467 | |
| 1468 | assert_eq!(composed, "embedder static prompt"); |
| 1469 | } |
| 1470 | |
| 1471 | fn contains_cjk(text: &str) -> bool { |
| 1472 | text.chars().any(|ch| { |
| 1473 | matches!( |
| 1474 | ch, |
| 1475 | '\u{3040}'..='\u{30ff}' |
| 1476 | | '\u{3400}'..='\u{4dbf}' |
| 1477 | | '\u{4e00}'..='\u{9fff}' |
| 1478 | | '\u{f900}'..='\u{faff}' |
| 1479 | ) |
| 1480 | }) |
| 1481 | } |
| 1482 | |
| 1483 | #[test] |
| 1484 | fn agent_mode_carries_execution_discipline_block() { |
| 1485 | for phrase in [ |
| 1486 | "Execute the user's task autonomously", |
| 1487 | "Keep `work_update` current", |
| 1488 | "present; otherwise", |
| 1489 | "verify load-bearing child", |
| 1490 | "never manufacture completion sentinels", |
| 1491 | "For substantial work", |
| 1492 | "session-persistent `repl` blocks", |
| 1493 | "retain source/transcript\nas data", |
| 1494 | "`workflow`, `agent`, goals, `harness`", |
| 1495 | ] { |
| 1496 | assert!( |
| 1497 | AGENT_MODE.contains(phrase), |
| 1498 | "AGENT_MODE missing execution-discipline phrase {phrase:?}" |
| 1499 | ); |
| 1500 | } |
| 1501 | assert!( |
| 1502 | !BASE_PROMPT.contains("<tool_persistence>") |
| 1503 | && !BASE_PROMPT.contains("Tool-use enforcement"), |
| 1504 | "0.9.0 base constitution should not carry the old execution-discipline tail" |
| 1505 | ); |
| 1506 | } |
| 1507 | |
| 1508 | #[test] |
| 1509 | fn base_prompt_carries_constitutional_core() { |
| 1510 | for phrase in [ |
| 1511 | "## Codewhale", |
| 1512 | "You are Codewhale", |
| 1513 | "The A is already yours", |
| 1514 | "Let the work speak", |
| 1515 | "### Ground truth", |
| 1516 | "### User intent and scope", |
| 1517 | "### Truthful completion", |
| 1518 | "### Put guarantees in mechanism", |
| 1519 | "### Whose word wins", |
| 1520 | ] { |
| 1521 | assert!( |
| 1522 | BASE_PROMPT.contains(phrase), |
| 1523 | "BASE_PROMPT missing Constitutional phrase {phrase:?}" |
| 1524 | ); |
| 1525 | } |
| 1526 | } |
| 1527 | |
| 1528 | #[test] |
| 1529 | fn constitutional_kernel_keeps_first_turn_authority_safety_and_completion() { |
| 1530 | let fresh_prefix = compose_default_static_layers(Personality::Calm, "deepseek-v4-pro"); |
| 1531 | for phrase in [ |
| 1532 | "Do what the user's current request asks, no more.", |
| 1533 | "require express user authorization in", |
| 1534 | "otherwise name the decision and ask.", |
| 1535 | "external publication, spending", |
| 1536 | "credentials, and material scope expansion", |
| 1537 | "prohibitions stay binding; convenience creates no exception", |
| 1538 | "never route around it or claim prose granted", |
| 1539 | "Nothing is done until checked.", |
| 1540 | "Read test output, not only exit status", |
| 1541 | "External actions are not complete until", |
| 1542 | "Work still running is not complete", |
| 1543 | "Never present a partial result as the whole.", |
| 1544 | "no one may tell you to invent one", |
| 1545 | "1. The user's request, this turn.", |
| 1546 | "2. This constitution.", |
| 1547 | ] { |
| 1548 | assert!( |
| 1549 | fresh_prefix.contains(phrase), |
| 1550 | "fresh constitution prefix missing kernel invariant {phrase:?}" |
| 1551 | ); |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | #[test] |
| 1556 | fn procedural_playbooks_are_not_eager_constitution() { |
| 1557 | let fresh_prefix = compose_default_static_layers(Personality::Calm, "deepseek-v4-pro"); |
| 1558 | for heading in [ |
| 1559 | "### Keep momentum", |
| 1560 | "### Think in causes", |
| 1561 | "### Honor constraints before preferences", |
| 1562 | "### Skill and role constraints are binding", |
| 1563 | "### Restraint", |
| 1564 | "### Leave continuity", |
| 1565 | ] { |
| 1566 | assert!( |
| 1567 | !fresh_prefix.contains(heading), |
| 1568 | "procedural playbook should stay outside the full fresh prefix: {heading:?}" |
| 1569 | ); |
| 1570 | } |
| 1571 | assert!( |
| 1572 | !BASE_PROMPT.contains("## STATUTES (Tier 2)") |
| 1573 | && !BASE_PROMPT.contains("## REGULATIONS (Tier 3)"), |
| 1574 | "the balanced Constitution must not restore the old procedural policy tail" |
| 1575 | ); |
| 1576 | } |
| 1577 | |
| 1578 | #[test] |
| 1579 | fn base_prompt_carries_verify_then_stop_completion_contract() { |
| 1580 | // The completion contract behind "Truthful completion": verify with real |
| 1581 | // evidence, keep running work visible, and hand back exactly what |
| 1582 | // changed. These phrases encode the contract's semantics, not its |
| 1583 | // prose — a rewording that keeps the contract should keep these, and |
| 1584 | // one that drops them is a real behavior change worth failing review |
| 1585 | // for. (Constitution kernel rewrite in #5077 renamed the section and |
| 1586 | // condensed the prose; the contract stands.) |
| 1587 | for phrase in [ |
| 1588 | "Nothing is done until checked.", |
| 1589 | "Read test output, not only exit status", |
| 1590 | "Work still running is not complete", |
| 1591 | "Never present a partial result as the whole.", |
| 1592 | ] { |
| 1593 | assert!( |
| 1594 | BASE_PROMPT.contains(phrase), |
| 1595 | "BASE_PROMPT missing completion-contract phrase {phrase:?}" |
| 1596 | ); |
| 1597 | } |
| 1598 | } |
| 1599 | |
| 1600 | #[test] |
| 1601 | fn yolo_mode_composed_prompt_carries_completion_contract() { |
| 1602 | // `codewhale exec --auto` runs AppMode::Yolo; the verify-then-stop |
| 1603 | // contract must survive composition into the prompt that mode ships. |
| 1604 | let tmp = tempdir().expect("tempdir"); |
| 1605 | let text = system_prompt_flat_text( |
| 1606 | &system_prompt_for_mode_with_context_skills_session_and_approval( |
| 1607 | tmp.path(), |
| 1608 | None, |
| 1609 | None, |
| 1610 | None, |
| 1611 | PromptSessionContext { |
| 1612 | user_memory_block: None, |
| 1613 | goal_objective: None, |
| 1614 | project_context_pack_enabled: false, |
| 1615 | locale_tag: "en", |
| 1616 | translation_enabled: false, |
| 1617 | model_id: "codewhale", |
| 1618 | context_window_override: None, |
| 1619 | verbosity: None, |
| 1620 | skills_scan_codewhale_only: false, |
| 1621 | plugin_registry: None, |
| 1622 | mode: crate::tui::app::AppMode::Yolo, |
| 1623 | }, |
| 1624 | ), |
| 1625 | ); |
| 1626 | for phrase in [ |
| 1627 | "##### Mode: Agent", |
| 1628 | "### Truthful completion", |
| 1629 | "Nothing is done until checked.", |
| 1630 | "Never present a partial result as the whole.", |
| 1631 | ] { |
| 1632 | assert!( |
| 1633 | text.contains(phrase), |
| 1634 | "YOLO-mode composed prompt missing completion-contract phrase {phrase:?}" |
| 1635 | ); |
| 1636 | } |
| 1637 | } |
| 1638 | |
| 1639 | #[test] |
| 1640 | fn constitutional_hierarchy_keeps_user_turn_above_local_law() { |
| 1641 | let heading_at = BASE_PROMPT |
| 1642 | .find("### Whose word wins") |
| 1643 | .expect("Whose word wins heading present"); |
| 1644 | let user_at = BASE_PROMPT |
| 1645 | .find("1. The user's request, this turn.") |
| 1646 | .expect("user request tier present"); |
| 1647 | let constitution_at = BASE_PROMPT |
| 1648 | .find("2. This constitution.") |
| 1649 | .expect("constitution tier present"); |
| 1650 | let project_at = BASE_PROMPT |
| 1651 | .find("3. Project law and instructions") |
| 1652 | .expect("project tier present"); |
| 1653 | let preference_at = BASE_PROMPT |
| 1654 | .find("4. Your standing user-global preferences.") |
| 1655 | .expect("user-global preference tier present"); |
| 1656 | let memory_at = BASE_PROMPT |
| 1657 | .find("5. Memory and previous-session handoffs.") |
| 1658 | .expect("memory/handoff tier present"); |
| 1659 | |
| 1660 | assert!( |
| 1661 | heading_at < user_at |
| 1662 | && user_at < constitution_at |
| 1663 | && constitution_at < project_at |
| 1664 | && project_at < preference_at |
| 1665 | && preference_at < memory_at, |
| 1666 | "Whose word wins must rank the current user request above constitution, \ |
| 1667 | project law, standing user-global preferences, then memory/handoffs" |
| 1668 | ); |
| 1669 | assert!( |
| 1670 | BASE_PROMPT.contains("the user may override a fact, but no one may invent\none"), |
| 1671 | "Whose word wins must keep ground truth overridable but never inventable" |
| 1672 | ); |
| 1673 | assert!( |
| 1674 | BASE_PROMPT.contains("A tie you cannot break is not yours to break"), |
| 1675 | "Whose word wins must keep tie-break escalation" |
| 1676 | ); |
| 1677 | } |
| 1678 | |
| 1679 | #[test] |
| 1680 | fn base_prompt_is_model_fact_free() { |
| 1681 | for placeholder in [ |
| 1682 | "{model_id}", |
| 1683 | "{context_window_note}", |
| 1684 | "{subagent_economics}", |
| 1685 | "{model_thinking_note}", |
| 1686 | "{model_characteristics}", |
| 1687 | ] { |
| 1688 | assert!( |
| 1689 | !BASE_PROMPT.contains(placeholder), |
| 1690 | "0.9.0 BASE_PROMPT must not contain model-fact placeholder {placeholder}" |
| 1691 | ); |
| 1692 | } |
| 1693 | for forbidden in [ |
| 1694 | "Your V4 Characteristics", |
| 1695 | "Model Characteristics", |
| 1696 | "one-million-token context window", |
| 1697 | "provider-dependent and not known", |
| 1698 | ] { |
| 1699 | assert!( |
| 1700 | !BASE_PROMPT.contains(forbidden), |
| 1701 | "0.9.0 BASE_PROMPT must not contain model-specific fact {forbidden:?}" |
| 1702 | ); |
| 1703 | } |
| 1704 | } |
| 1705 | |
| 1706 | fn assert_no_unresolved_model_placeholders(prompt: &str) { |
| 1707 | for placeholder in [ |
| 1708 | "{model_id}", |
| 1709 | "{context_window_note}", |
| 1710 | "{subagent_economics}", |
| 1711 | "{model_thinking_note}", |
| 1712 | "{model_characteristics}", |
| 1713 | ] { |
| 1714 | assert!( |
| 1715 | !prompt.contains(placeholder), |
| 1716 | "composed prompt must not contain unresolved {placeholder}" |
| 1717 | ); |
| 1718 | } |
| 1719 | } |
| 1720 | |
| 1721 | #[test] |
| 1722 | fn compose_prompt_for_v4_model_stays_model_fact_free() { |
| 1723 | let prompt = |
| 1724 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "deepseek-v4-pro"); |
| 1725 | assert!(prompt.contains("You are Codewhale")); |
| 1726 | assert!(!prompt.contains("Your V4 Characteristics")); |
| 1727 | assert!(!prompt.contains("one-million-token context window")); |
| 1728 | assert_no_unresolved_model_placeholders(&prompt); |
| 1729 | } |
| 1730 | |
| 1731 | #[test] |
| 1732 | fn compose_prompt_for_kimi_stays_model_fact_free() { |
| 1733 | let prompt = |
| 1734 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "moonshotai/kimi-k2.6"); |
| 1735 | assert!(prompt.contains("You are Codewhale")); |
| 1736 | assert!(!prompt.contains("Your V4 Characteristics")); |
| 1737 | assert!(!prompt.contains("one-million")); |
| 1738 | assert!(!prompt.contains("$0.14")); |
| 1739 | assert!(!prompt.contains("262144-token context window")); |
| 1740 | assert!(!prompt.contains("Models may emit *thinking tokens*")); |
| 1741 | assert_no_unresolved_model_placeholders(&prompt); |
| 1742 | } |
| 1743 | |
| 1744 | #[test] |
| 1745 | fn compose_prompt_for_openai_api_gpt_55_stays_model_fact_free() { |
| 1746 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "gpt-5.5"); |
| 1747 | assert!(prompt.contains("You are Codewhale")); |
| 1748 | assert!(!prompt.contains("Your V4 Characteristics")); |
| 1749 | assert!(!prompt.contains("1050000-token context window")); |
| 1750 | assert!(!prompt.contains("Models may emit *thinking tokens*")); |
| 1751 | assert!(!prompt.contains("provider-dependent and not known")); |
| 1752 | assert_no_unresolved_model_placeholders(&prompt); |
| 1753 | } |
| 1754 | |
| 1755 | #[test] |
| 1756 | fn compose_prompt_for_unknown_model_stays_model_fact_free() { |
| 1757 | let prompt = |
| 1758 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "llama3.3:70b"); |
| 1759 | assert!(prompt.contains("You are Codewhale")); |
| 1760 | assert!(!prompt.contains("Your V4 Characteristics")); |
| 1761 | assert!(!prompt.contains("one-million")); |
| 1762 | assert!(!prompt.contains("$0.14")); |
| 1763 | assert!(!prompt.contains("provider-dependent and not known")); |
| 1764 | assert!(!prompt.contains("Models may emit *thinking tokens*")); |
| 1765 | assert_no_unresolved_model_placeholders(&prompt); |
| 1766 | } |
| 1767 | |
| 1768 | #[test] |
| 1769 | fn apply_model_template_replaces_placeholder() { |
| 1770 | let result = apply_model_template("You are {model_id}", "deepseek-v4-pro", None); |
| 1771 | assert_eq!(result, "You are deepseek-v4-pro"); |
| 1772 | assert!(!result.contains("{model_id}")); |
| 1773 | } |
| 1774 | |
| 1775 | #[test] |
| 1776 | fn apply_model_template_does_not_resolve_removed_model_fact_templates() { |
| 1777 | let result = apply_model_template("{context_window_note}", "gpt-5.5", Some(400_000)); |
| 1778 | assert_eq!(result, "{context_window_note}"); |
| 1779 | assert!(!result.contains("400000-token context window")); |
| 1780 | assert!(!result.contains("1050000-token context window")); |
| 1781 | } |
| 1782 | |
| 1783 | #[test] |
| 1784 | fn compose_prompt_is_model_agnostic_in_preamble() { |
| 1785 | // 0.9.0 keeps the preamble byte-for-byte the same regardless of |
| 1786 | // model id, and no {model_id} placeholder leaks. |
| 1787 | let flash = |
| 1788 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "deepseek-v4-flash"); |
| 1789 | let kimi = |
| 1790 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "moonshotai/kimi-k2.6"); |
| 1791 | assert!( |
| 1792 | flash.contains("You are Codewhale"), |
| 1793 | "0.9.0 preamble must open with the model-agnostic Codewhale stance" |
| 1794 | ); |
| 1795 | assert!( |
| 1796 | !flash.contains("You are deepseek-v4-flash") |
| 1797 | && !kimi.contains("You are moonshotai/kimi-k2.6"), |
| 1798 | "0.9.0 preamble must not inject a per-model identity line" |
| 1799 | ); |
| 1800 | assert!( |
| 1801 | !flash.contains("{model_id}") && !kimi.contains("{model_id}"), |
| 1802 | "composed prompt must not contain the raw {{model_id}} placeholder" |
| 1803 | ); |
| 1804 | } |
| 1805 | |
| 1806 | #[test] |
| 1807 | fn tool_descriptions_carry_edit_and_shell_guidance() { |
| 1808 | let write = WriteFileTool.description(); |
| 1809 | assert!( |
| 1810 | write.contains("instead of heredocs") |
| 1811 | && write.contains("`Bash`") |
| 1812 | && !write.contains("exec_shell"), |
| 1813 | "write guidance must name the live Bash tool and never the retired exec_shell name" |
| 1814 | ); |
| 1815 | |
| 1816 | let edit = EditFileTool.description(); |
| 1817 | // Every handler description must name the live `File` surface plus an |
| 1818 | // action. `read_file`/`write_file`/`apply_patch` are retired spellings |
| 1819 | // (crates/tui/src/tools/registry.rs:2066-2088). |
| 1820 | assert!(edit.contains("File `read`")); |
| 1821 | assert!(edit.contains("File `patch` or `write`")); |
| 1822 | assert!( |
| 1823 | !edit.contains("read_file") |
| 1824 | && !edit.contains("write_file") |
| 1825 | && !edit.contains("apply_patch"), |
| 1826 | "edit guidance must not teach a retired tool name: {edit:?}" |
| 1827 | ); |
| 1828 | |
| 1829 | let patch = ApplyPatchTool.description(); |
| 1830 | assert!(patch.contains("unified-diff") && patch.contains("transactional")); |
| 1831 | |
| 1832 | let shell_tool = BashTool::new("Bash"); |
| 1833 | let shell = shell_tool.description(); |
| 1834 | assert!(shell.contains("background=true")); |
| 1835 | assert!(shell.contains(">5 seconds")); |
| 1836 | } |
| 1837 | |
| 1838 | #[test] |
| 1839 | fn composed_prompt_does_not_claim_tool_availability() { |
| 1840 | let prompt = |
| 1841 | compose_prompt_with_approval_model_and_shell(Personality::Calm, "deepseek-v4-pro"); |
| 1842 | assert!(!prompt.contains("## Core Tool Taxonomy")); |
| 1843 | assert!(!prompt.contains("## Toolbox")); |
| 1844 | assert!(prompt.contains("You are Codewhale")); |
| 1845 | } |
| 1846 | |
| 1847 | #[test] |
| 1848 | fn authority_recap_appears_in_full_prompt() { |
| 1849 | let tmp = tempdir().expect("tempdir"); |
| 1850 | let text = system_prompt_flat_text( |
| 1851 | &system_prompt_for_mode_with_context_skills_session_and_approval( |
| 1852 | tmp.path(), |
| 1853 | None, |
| 1854 | None, |
| 1855 | None, |
| 1856 | PromptSessionContext::default(), |
| 1857 | ), |
| 1858 | ); |
| 1859 | assert!( |
| 1860 | text.contains("## Authority Recap"), |
| 1861 | "full system prompt must contain the authority recap" |
| 1862 | ); |
| 1863 | assert!( |
| 1864 | text.contains("Codewhale's constitution governs your behavior"), |
| 1865 | "authority recap must reference the Constitution" |
| 1866 | ); |
| 1867 | assert!( |
| 1868 | text.contains("consult ### Whose word wins"), |
| 1869 | "authority recap must point at 0.9.0's precedence section" |
| 1870 | ); |
| 1871 | } |
| 1872 | |
| 1873 | #[test] |
| 1874 | fn system_prompt_merges_workspace_and_configured_skills_dir() { |
| 1875 | let _env_guard = crate::test_support::lock_test_env(); |
| 1876 | let tmp = tempdir().expect("tempdir"); |
| 1877 | let _home = ScopedHome::set(tmp.path().join("home")); |
| 1878 | let workspace = tmp.path().join("workspace"); |
| 1879 | let configured_dir = tmp.path().join("configured-skills"); |
| 1880 | write_test_skill( |
| 1881 | &workspace.join(".claude").join("skills"), |
| 1882 | "workspace-skill", |
| 1883 | "workspace skill", |
| 1884 | ); |
| 1885 | write_test_skill(&configured_dir, "configured-skill", "configured skill"); |
| 1886 | |
| 1887 | let text = system_prompt_flat_text(&system_prompt_for_mode_with_context_and_skills( |
| 1888 | &workspace, |
| 1889 | None, |
| 1890 | Some(&configured_dir), |
| 1891 | None, |
| 1892 | None, |
| 1893 | )); |
| 1894 | |
| 1895 | assert!(text.contains("workspace-skill")); |
| 1896 | assert!(text.contains("configured-skill")); |
| 1897 | } |
| 1898 | |
| 1899 | struct ScopedHome { |
| 1900 | previous: Option<std::ffi::OsString>, |
| 1901 | } |
| 1902 | |
| 1903 | impl ScopedHome { |
| 1904 | fn set(path: std::path::PathBuf) -> Self { |
| 1905 | let previous = std::env::var_os("HOME"); |
| 1906 | // Safety: this test serializes environment access with |
| 1907 | // lock_test_env and restores HOME in Drop. |
| 1908 | unsafe { |
| 1909 | std::env::set_var("HOME", path); |
| 1910 | } |
| 1911 | Self { previous } |
| 1912 | } |
| 1913 | } |
| 1914 | |
| 1915 | impl Drop for ScopedHome { |
| 1916 | fn drop(&mut self) { |
| 1917 | // Safety: this test serializes environment access with |
| 1918 | // lock_test_env and restores HOME in Drop. |
| 1919 | unsafe { |
| 1920 | if let Some(previous) = self.previous.take() { |
| 1921 | std::env::set_var("HOME", previous); |
| 1922 | } else { |
| 1923 | std::env::remove_var("HOME"); |
| 1924 | } |
| 1925 | } |
| 1926 | } |
| 1927 | } |
| 1928 | |
| 1929 | fn write_test_skill(root: &std::path::Path, name: &str, description: &str) { |
| 1930 | let dir = root.join(name); |
| 1931 | std::fs::create_dir_all(&dir).expect("skill dir"); |
| 1932 | std::fs::write( |
| 1933 | dir.join("SKILL.md"), |
| 1934 | format!("---\nname: {name}\ndescription: {description}\n---\n\n# {name}\n"), |
| 1935 | ) |
| 1936 | .expect("skill file"); |
| 1937 | } |
| 1938 | |
| 1939 | #[test] |
| 1940 | fn constitution_has_no_separate_personality_tier() { |
| 1941 | // 0.9.0 has no personality tier. Voice and tone live in the |
| 1942 | // compact constitution rather than a separate section, so |
| 1943 | // personality remains folded in by omission. |
| 1944 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 1945 | assert!( |
| 1946 | !prompt.contains("Personality: Calm — Tier 8"), |
| 1947 | "Personality tier should not appear as a separate section" |
| 1948 | ); |
| 1949 | assert!( |
| 1950 | prompt.contains("Take the work seriously. Don't take"), |
| 1951 | "Preamble should carry tone guidance (take the work, not yourself, seriously)" |
| 1952 | ); |
| 1953 | // Verify the preamble still carries the Codewhale identity. |
| 1954 | assert!(prompt.contains("You are Codewhale")); |
| 1955 | assert!(prompt.contains("Let the work speak")); |
| 1956 | } |
| 1957 | |
| 1958 | #[test] |
| 1959 | fn execution_discipline_lives_in_agent_mode_after_core_constitution() { |
| 1960 | assert!(AGENT_MODE.contains("Execute the user's task autonomously")); |
| 1961 | assert!(AGENT_MODE.contains("verify load-bearing child")); |
| 1962 | assert!( |
| 1963 | !BASE_PROMPT.contains("Execution Discipline") |
| 1964 | && !BASE_PROMPT.contains("<tool_persistence>"), |
| 1965 | "base constitution should stay reduced; execution discipline belongs to Agent mode" |
| 1966 | ); |
| 1967 | } |
| 1968 | |
| 1969 | #[test] |
| 1970 | fn plan_mode_prompt_uses_one_progress_surface() { |
| 1971 | assert!( |
| 1972 | PLAN_MODE.contains("When `work_update` is present") |
| 1973 | && PLAN_MODE.contains("canonical list there") |
| 1974 | && PLAN_MODE.contains("otherwise keep progress in your response"), |
| 1975 | "Plan mode must condition progress guidance on the live catalog" |
| 1976 | ); |
| 1977 | assert!(!PLAN_MODE.contains("call `update_plan`")); |
| 1978 | assert!( |
| 1979 | PLAN_MODE.contains("switch to Act (`/mode act`)"), |
| 1980 | "Plan mode must use a normal conversational handoff" |
| 1981 | ); |
| 1982 | } |
| 1983 | |
| 1984 | #[test] |
| 1985 | fn render_environment_block_keeps_actionable_host_facts_without_version() { |
| 1986 | let tmp = tempdir().expect("tempdir"); |
| 1987 | let block = render_environment_block(tmp.path(), "zh-Hans"); |
| 1988 | assert!(block.starts_with("## Environment")); |
| 1989 | assert!(block.contains("- lang: zh-Hans")); |
| 1990 | // The workspace remains per-turn and the release version is telemetry; |
| 1991 | // platform and shell still steer valid command syntax. |
| 1992 | assert!(!block.contains("- pwd:")); |
| 1993 | assert!(!block.contains("- codewhale_version:")); |
| 1994 | assert!(block.contains("- platform:")); |
| 1995 | assert!(block.contains("- shell:")); |
| 1996 | } |
| 1997 | |
| 1998 | #[test] |
| 1999 | fn locale_reinforcement_preamble_returns_native_script_for_supported_locales() { |
| 2000 | // English (and unknown locales) get None — the existing English |
| 2001 | // directive in `constitution.md` is sufficient. |
| 2002 | assert!(locale_reinforcement_preamble("en").is_none()); |
| 2003 | assert!(locale_reinforcement_preamble("en-US").is_none()); |
| 2004 | assert!(locale_reinforcement_preamble("fr-FR").is_none()); |
| 2005 | assert!(locale_reinforcement_preamble("").is_none()); |
| 2006 | |
| 2007 | // zh-Hans (and the de-facto equivalents the TUI accepts) get a |
| 2008 | // native-script preamble. The text must explicitly mention |
| 2009 | // `reasoning_content` (the V4 knob this is meant to steer) and |
| 2010 | // preserve tool-name immutability — those are the load-bearing |
| 2011 | // claims behind the #1118 fix that someone could quietly |
| 2012 | // delete in a future translation pass. |
| 2013 | for tag in ["zh-Hans", "zh-CN", "zh"] { |
| 2014 | let preamble = |
| 2015 | locale_reinforcement_preamble(tag).expect("zh-Hans preamble should exist"); |
| 2016 | assert!( |
| 2017 | preamble.contains("简体中文"), |
| 2018 | "zh preamble must be in Simplified Chinese: {preamble:?}" |
| 2019 | ); |
| 2020 | assert!( |
| 2021 | preamble.contains("reasoning_content"), |
| 2022 | "zh preamble must steer reasoning_content: {preamble:?}" |
| 2023 | ); |
| 2024 | assert!( |
| 2025 | preamble.contains("`File`"), |
| 2026 | "zh preamble must call out tool-name immutability with a LIVE tool \ |
| 2027 | name; `read_file` is retired (registry.rs:2067): {preamble:?}" |
| 2028 | ); |
| 2029 | assert!( |
| 2030 | !preamble.contains("read_file") && !preamble.contains("exec_shell"), |
| 2031 | "zh preamble must never teach a retired tool name: {preamble:?}" |
| 2032 | ); |
| 2033 | } |
| 2034 | |
| 2035 | let ja = locale_reinforcement_preamble("ja").expect("ja preamble"); |
| 2036 | assert!(ja.contains("日本語"), "ja preamble must be in Japanese"); |
| 2037 | assert!(ja.contains("reasoning_content")); |
| 2038 | |
| 2039 | let pt = locale_reinforcement_preamble("pt-BR").expect("pt-BR preamble"); |
| 2040 | assert!( |
| 2041 | pt.contains("português do Brasil"), |
| 2042 | "pt preamble must call out pt-BR explicitly" |
| 2043 | ); |
| 2044 | assert!(pt.contains("reasoning_content")); |
| 2045 | } |
| 2046 | |
| 2047 | #[test] |
| 2048 | fn system_prompt_prepends_locale_preamble_for_zh_hans() { |
| 2049 | // Build the full system prompt with locale=zh-Hans and assert |
| 2050 | // the native-script preamble shows up *before* the English |
| 2051 | // base-prompt body. Cache stability and attention precedence |
| 2052 | // both depend on this ordering. |
| 2053 | let tmp = tempdir().expect("tempdir"); |
| 2054 | let text = system_prompt_flat_text( |
| 2055 | &system_prompt_for_mode_with_context_skills_session_and_approval( |
| 2056 | tmp.path(), |
| 2057 | None, |
| 2058 | None, |
| 2059 | None, |
| 2060 | PromptSessionContext { |
| 2061 | user_memory_block: None, |
| 2062 | goal_objective: None, |
| 2063 | project_context_pack_enabled: false, |
| 2064 | locale_tag: "zh-Hans", |
| 2065 | translation_enabled: false, |
| 2066 | model_id: "codewhale", |
| 2067 | context_window_override: None, |
| 2068 | verbosity: None, |
| 2069 | skills_scan_codewhale_only: false, |
| 2070 | plugin_registry: None, |
| 2071 | mode: crate::tui::app::AppMode::Agent, |
| 2072 | }, |
| 2073 | ), |
| 2074 | ); |
| 2075 | let preamble_marker = "## 语言要求"; |
| 2076 | let base_marker = "You are Codewhale"; |
| 2077 | let preamble_pos = text |
| 2078 | .find(preamble_marker) |
| 2079 | .expect("zh-Hans preamble should be present"); |
| 2080 | let base_pos = text |
| 2081 | .find(base_marker) |
| 2082 | .expect("base prompt should be present"); |
| 2083 | assert!( |
| 2084 | preamble_pos < base_pos, |
| 2085 | "locale preamble must precede the English base prompt (preamble={preamble_pos}, base={base_pos})", |
| 2086 | ); |
| 2087 | } |
| 2088 | |
| 2089 | #[test] |
| 2090 | fn locale_reinforcement_closer_returns_native_script_for_supported_locales() { |
| 2091 | // English (and unknown locales) get None. |
| 2092 | assert!(locale_reinforcement_closer("en").is_none()); |
| 2093 | assert!(locale_reinforcement_closer("fr-FR").is_none()); |
| 2094 | assert!(locale_reinforcement_closer("").is_none()); |
| 2095 | |
| 2096 | // Each supported locale gets a closer in its own script that |
| 2097 | // explicitly tells the model "don't drift to English even as |
| 2098 | // English context accumulates" — that's the load-bearing claim |
| 2099 | // behind the bookend pattern. |
| 2100 | let zh = locale_reinforcement_closer("zh-Hans").expect("zh closer"); |
| 2101 | assert!( |
| 2102 | zh.contains("简体中文"), |
| 2103 | "zh closer must be in Simplified Chinese" |
| 2104 | ); |
| 2105 | assert!( |
| 2106 | zh.contains("reasoning_content"), |
| 2107 | "zh closer must steer reasoning_content" |
| 2108 | ); |
| 2109 | let ja = locale_reinforcement_closer("ja").expect("ja closer"); |
| 2110 | assert!(ja.contains("日本語"), "ja closer must be in Japanese"); |
| 2111 | assert!(ja.contains("reasoning_content")); |
| 2112 | let pt = locale_reinforcement_closer("pt-BR").expect("pt-BR closer"); |
| 2113 | assert!(pt.contains("português do Brasil")); |
| 2114 | assert!(pt.contains("reasoning_content")); |
| 2115 | } |
| 2116 | |
| 2117 | #[test] |
| 2118 | fn v092_locales_add_no_prompt_bookends_so_prompt_bytes_stay_stable() { |
| 2119 | // Cache-stability contract: adding the v0.9.2 UI locales |
| 2120 | // (ca, de, fr, id, hi, ru, uk) — and the already-shipped UI packs |
| 2121 | // that never had bookends (ko, es-419, zh-Hant) — must not change |
| 2122 | // the model-visible system prompt for an identical route/session |
| 2123 | // when translation is not explicitly enabled. The bookend list |
| 2124 | // stays intentionally short (zh-Hans, ja, pt-BR, vi); every other |
| 2125 | // shipped locale resolves to None and therefore renders the exact |
| 2126 | // same prompt bytes as English. |
| 2127 | for tag in [ |
| 2128 | "zh-Hant", "ko", "es-419", "ca", "de", "fr", "id", "hi", "ru", "uk", |
| 2129 | ] { |
| 2130 | assert!( |
| 2131 | locale_reinforcement_preamble(tag).is_none(), |
| 2132 | "{tag} must not gain a locale preamble" |
| 2133 | ); |
| 2134 | assert!( |
| 2135 | locale_reinforcement_closer(tag).is_none(), |
| 2136 | "{tag} must not gain a locale closer" |
| 2137 | ); |
| 2138 | } |
| 2139 | // The bookend set is exactly the original four locales — growing it |
| 2140 | // is a deliberate, reviewable prompt change, not a side effect of |
| 2141 | // adding a UI pack. |
| 2142 | for tag in ["zh-Hans", "ja", "pt-BR", "vi"] { |
| 2143 | assert!( |
| 2144 | locale_reinforcement_preamble(tag).is_some(), |
| 2145 | "{tag} lost its locale preamble" |
| 2146 | ); |
| 2147 | assert!( |
| 2148 | locale_reinforcement_closer(tag).is_some(), |
| 2149 | "{tag} lost its locale closer" |
| 2150 | ); |
| 2151 | } |
| 2152 | } |
| 2153 | |
| 2154 | #[test] |
| 2155 | fn translation_seam_names_every_shipped_locale_canonically() { |
| 2156 | // The translation output instruction is the declared model-facing |
| 2157 | // seam: it only enters the prompt when `translation_enabled` is |
| 2158 | // true. When it does, every shipped locale must be named |
| 2159 | // canonically (English name + endonym) — never silently "English". |
| 2160 | for locale in crate::localization::Locale::shipped() { |
| 2161 | assert_eq!( |
| 2162 | translation_target_language_for_tag(locale.tag()), |
| 2163 | locale.translation_target_name(), |
| 2164 | "{} translation seam drifted from the canonical locale name", |
| 2165 | locale.tag() |
| 2166 | ); |
| 2167 | } |
| 2168 | } |
| 2169 | |
| 2170 | #[test] |
| 2171 | fn system_prompt_bookends_zh_hans_with_preamble_and_closer() { |
| 2172 | // The full system prompt for zh-Hans must contain BOTH the |
| 2173 | // opening preamble (`## 语言要求`) and the closing reinforcement |
| 2174 | // (`## 语言再次提醒`), with the closer appearing AFTER the |
| 2175 | // preamble — i.e. the prompt is "bookended" in native script, |
| 2176 | // matching the empirical finding from the WeChat thread that |
| 2177 | // motivated the closer. |
| 2178 | let tmp = tempdir().expect("tempdir"); |
| 2179 | let text = system_prompt_flat_text( |
| 2180 | &system_prompt_for_mode_with_context_skills_session_and_approval( |
| 2181 | tmp.path(), |
| 2182 | None, |
| 2183 | None, |
| 2184 | None, |
| 2185 | PromptSessionContext { |
| 2186 | user_memory_block: None, |
| 2187 | goal_objective: None, |
| 2188 | project_context_pack_enabled: false, |
| 2189 | locale_tag: "zh-Hans", |
| 2190 | translation_enabled: false, |
| 2191 | model_id: "codewhale", |
| 2192 | context_window_override: None, |
| 2193 | verbosity: None, |
| 2194 | skills_scan_codewhale_only: false, |
| 2195 | plugin_registry: None, |
| 2196 | mode: crate::tui::app::AppMode::Agent, |
| 2197 | }, |
| 2198 | ), |
| 2199 | ); |
| 2200 | let preamble_pos = text |
| 2201 | .find("## 语言要求") |
| 2202 | .expect("zh-Hans preamble must be in prompt"); |
| 2203 | let closer_pos = text |
| 2204 | .find("## 语言再次提醒") |
| 2205 | .expect("zh-Hans closer must be in prompt"); |
| 2206 | assert!( |
| 2207 | preamble_pos < closer_pos, |
| 2208 | "closer must come after preamble (preamble={preamble_pos}, closer={closer_pos})", |
| 2209 | ); |
| 2210 | // The closer must be the very last block — anything else after |
| 2211 | // it defeats the recency-bias purpose. Skip the closer's own |
| 2212 | // `## ` header before scanning. |
| 2213 | let closer_header_end = closer_pos + "## 语言再次提醒".len(); |
| 2214 | let after_closer_body = &text[closer_header_end..]; |
| 2215 | assert!( |
| 2216 | !after_closer_body.contains("\n## "), |
| 2217 | "no other top-level section should follow the closer; got: {after_closer_body:?}", |
| 2218 | ); |
| 2219 | } |
| 2220 | |
| 2221 | #[test] |
| 2222 | fn system_prompt_skips_locale_preamble_for_english() { |
| 2223 | // English locale → no preamble injected. Asserts the |
| 2224 | // "preamble is opt-in for non-English" invariant. |
| 2225 | let tmp = tempdir().expect("tempdir"); |
| 2226 | let text = system_prompt_flat_text( |
| 2227 | &system_prompt_for_mode_with_context_skills_session_and_approval( |
| 2228 | tmp.path(), |
| 2229 | None, |
| 2230 | None, |
| 2231 | None, |
| 2232 | PromptSessionContext { |
| 2233 | user_memory_block: None, |
| 2234 | goal_objective: None, |
| 2235 | project_context_pack_enabled: false, |
| 2236 | locale_tag: "en", |
| 2237 | translation_enabled: false, |
| 2238 | model_id: "codewhale", |
| 2239 | context_window_override: None, |
| 2240 | verbosity: None, |
| 2241 | skills_scan_codewhale_only: false, |
| 2242 | plugin_registry: None, |
| 2243 | mode: crate::tui::app::AppMode::Agent, |
| 2244 | }, |
| 2245 | ), |
| 2246 | ); |
| 2247 | assert!( |
| 2248 | !text.contains("语言要求"), |
| 2249 | "English locale must not get a zh preamble: {text:?}" |
| 2250 | ); |
| 2251 | assert!( |
| 2252 | !text.contains("言語要件"), |
| 2253 | "English locale must not get a ja preamble: {text:?}" |
| 2254 | ); |
| 2255 | assert!( |
| 2256 | !text.contains("Requisito de Idioma"), |
| 2257 | "English locale must not get a pt-BR preamble: {text:?}" |
| 2258 | ); |
| 2259 | // Closer too — same bookend rule. |
| 2260 | assert!( |
| 2261 | !text.contains("语言再次提醒"), |
| 2262 | "English locale must not get a zh closer: {text:?}" |
| 2263 | ); |
| 2264 | assert!( |
| 2265 | !text.contains("言語再確認"), |
| 2266 | "English locale must not get a ja closer: {text:?}" |
| 2267 | ); |
| 2268 | assert!( |
| 2269 | !text.contains("Reforço de Idioma"), |
| 2270 | "English locale must not get a pt-BR closer: {text:?}" |
| 2271 | ); |
| 2272 | assert!( |
| 2273 | !contains_cjk(BASE_PROMPT), |
| 2274 | "base prompt must not contain static CJK priming tokens" |
| 2275 | ); |
| 2276 | // Do not assert on arbitrary CJK in the full system prompt: project |
| 2277 | // context may legitimately contain localized file names, README text, |
| 2278 | // or user-authored instructions. The locale bookend markers above are |
| 2279 | // the priming tokens this test is meant to guard. |
| 2280 | } |
| 2281 | |
| 2282 | #[test] |
| 2283 | fn locale_bookends_carry_reasoning_content_directives_for_1118() { |
| 2284 | // #1118 ("Language has been configured to Chinese, but thinking |
| 2285 | // outputs are still in English"): after the 0.9.0 constitution |
| 2286 | // reduction, locale-native bookends carry the runtime language |
| 2287 | // reinforcement instead of the base constitution. |
| 2288 | let lang = LOCALE_PREAMBLE_ZH_HANS; |
| 2289 | assert!( |
| 2290 | lang.contains("reasoning_content"), |
| 2291 | "locale preamble must explicitly call out reasoning_content" |
| 2292 | ); |
| 2293 | assert!( |
| 2294 | lang.contains("最终回复"), |
| 2295 | "locale preamble must explicitly cover the final reply" |
| 2296 | ); |
| 2297 | assert!( |
| 2298 | lang.contains("代码") && lang.contains("工具名称"), |
| 2299 | "code and tool names must be named as non-language signals" |
| 2300 | ); |
| 2301 | assert!( |
| 2302 | LOCALE_CLOSER_ZH_HANS.contains("reasoning_content") |
| 2303 | && LOCALE_CLOSER_ZH_HANS.contains("继续用简体中文思考和回答"), |
| 2304 | "closing bookend must preserve recency-positioned language reinforcement" |
| 2305 | ); |
| 2306 | // Explicit-user-override clause keeps the prompt useful for the |
| 2307 | // opposite preference (#1118 commenters who want English |
| 2308 | // thinking for token-cost reasons). |
| 2309 | let phrase = "think in English"; |
| 2310 | assert!( |
| 2311 | lang.contains(phrase) && LOCALE_CLOSER_ZH_HANS.contains(phrase), |
| 2312 | "expected the user-override example `{phrase}`" |
| 2313 | ); |
| 2314 | } |
| 2315 | |
| 2316 | #[test] |
| 2317 | fn environment_block_is_inserted_into_system_prompt() { |
| 2318 | let tmp = tempdir().expect("tempdir"); |
| 2319 | let prompt = |
| 2320 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2321 | tmp.path(), |
| 2322 | None, |
| 2323 | None, |
| 2324 | None, |
| 2325 | PromptSessionContext { |
| 2326 | user_memory_block: None, |
| 2327 | goal_objective: None, |
| 2328 | project_context_pack_enabled: false, |
| 2329 | locale_tag: "ja", |
| 2330 | translation_enabled: false, |
| 2331 | model_id: "codewhale", |
| 2332 | context_window_override: None, |
| 2333 | verbosity: None, |
| 2334 | skills_scan_codewhale_only: false, |
| 2335 | plugin_registry: None, |
| 2336 | mode: crate::tui::app::AppMode::Agent, |
| 2337 | }, |
| 2338 | )); |
| 2339 | assert!(prompt.contains("## Environment")); |
| 2340 | assert!(prompt.contains("- lang: ja")); |
| 2341 | assert!(!prompt.contains("- codewhale_version:")); |
| 2342 | assert!(prompt.contains("- platform:")); |
| 2343 | assert!(prompt.contains("- shell:")); |
| 2344 | } |
| 2345 | |
| 2346 | #[test] |
| 2347 | fn user_global_constitution_block_is_injected_separately() { |
| 2348 | let _env_guard = crate::test_support::lock_test_env(); |
| 2349 | let tmp = tempdir().expect("tempdir"); |
| 2350 | let workspace = tmp.path().join("workspace"); |
| 2351 | std::fs::create_dir_all(&workspace).expect("workspace dir"); |
| 2352 | let codewhale_home = tmp.path().join("codewhale-home"); |
| 2353 | std::fs::create_dir_all(&codewhale_home).expect("codewhale home"); |
| 2354 | let _codewhale_home = |
| 2355 | crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", codewhale_home.as_os_str()); |
| 2356 | |
| 2357 | let constitution = codewhale_config::UserConstitution { |
| 2358 | about: Some("Maintains Codewhale release lanes.".to_string()), |
| 2359 | working_style: vec!["Prefer live verification before claims.".to_string()], |
| 2360 | priorities: vec!["Keep release gates green.".to_string()], |
| 2361 | autonomy_preference: codewhale_config::AutonomyPreference::Balanced, |
| 2362 | ..codewhale_config::UserConstitution::default() |
| 2363 | }; |
| 2364 | constitution |
| 2365 | .save_to( |
| 2366 | &codewhale_home |
| 2367 | .join(codewhale_config::user_constitution::USER_CONSTITUTION_FILE_NAME), |
| 2368 | ) |
| 2369 | .expect("save user constitution"); |
| 2370 | |
| 2371 | let prompt = |
| 2372 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2373 | &workspace, |
| 2374 | None, |
| 2375 | None, |
| 2376 | None, |
| 2377 | PromptSessionContext { |
| 2378 | project_context_pack_enabled: false, |
| 2379 | ..PromptSessionContext::default() |
| 2380 | }, |
| 2381 | )); |
| 2382 | |
| 2383 | let base_at = prompt.find("### Whose word wins").expect("base prompt"); |
| 2384 | let user_block_at = prompt |
| 2385 | .find("<codewhale_user_constitution") |
| 2386 | .expect("user constitution block"); |
| 2387 | let env_at = prompt.find("- lang:").expect("rendered environment block"); |
| 2388 | assert!( |
| 2389 | base_at < user_block_at && user_block_at < env_at, |
| 2390 | "user constitution should be its own layer after the base/project context and before volatile environment data" |
| 2391 | ); |
| 2392 | assert!(prompt.contains("source=\"user-global\"")); |
| 2393 | assert!(prompt.contains("Maintains Codewhale release lanes.")); |
| 2394 | assert!(prompt.contains("Prefer live verification before claims.")); |
| 2395 | assert!( |
| 2396 | !prompt.contains(&codewhale_home.display().to_string()), |
| 2397 | "prompt should use the stable user-global source label, not a device-specific home path" |
| 2398 | ); |
| 2399 | } |
| 2400 | |
| 2401 | #[test] |
| 2402 | fn bundled_choice_disables_user_global_constitution_block() { |
| 2403 | let _env_guard = crate::test_support::lock_test_env(); |
| 2404 | let tmp = tempdir().expect("tempdir"); |
| 2405 | let workspace = tmp.path().join("workspace"); |
| 2406 | std::fs::create_dir_all(&workspace).expect("workspace dir"); |
| 2407 | let codewhale_home = tmp.path().join("codewhale-home"); |
| 2408 | std::fs::create_dir_all(&codewhale_home).expect("codewhale home"); |
| 2409 | let _codewhale_home = |
| 2410 | crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", codewhale_home.as_os_str()); |
| 2411 | |
| 2412 | let constitution = codewhale_config::UserConstitution { |
| 2413 | about: Some("This file should stay inactive.".to_string()), |
| 2414 | ..codewhale_config::UserConstitution::default() |
| 2415 | }; |
| 2416 | constitution |
| 2417 | .save_to( |
| 2418 | &codewhale_home |
| 2419 | .join(codewhale_config::user_constitution::USER_CONSTITUTION_FILE_NAME), |
| 2420 | ) |
| 2421 | .expect("save user constitution"); |
| 2422 | |
| 2423 | let mut state = codewhale_config::SetupState::default(); |
| 2424 | state.complete_constitution_checkpoint( |
| 2425 | crate::tui::setup::CONSTITUTION_CHECKPOINT_VERSION, |
| 2426 | codewhale_config::ConstitutionChoice::Bundled, |
| 2427 | ); |
| 2428 | state |
| 2429 | .save_to(&codewhale_home.join(codewhale_config::setup_state::SETUP_STATE_FILE_NAME)) |
| 2430 | .expect("save setup state"); |
| 2431 | |
| 2432 | let prompt = |
| 2433 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2434 | &workspace, |
| 2435 | None, |
| 2436 | None, |
| 2437 | None, |
| 2438 | PromptSessionContext { |
| 2439 | project_context_pack_enabled: false, |
| 2440 | ..PromptSessionContext::default() |
| 2441 | }, |
| 2442 | )); |
| 2443 | |
| 2444 | assert!(!prompt.contains("<codewhale_user_constitution")); |
| 2445 | assert!(!prompt.contains("This file should stay inactive.")); |
| 2446 | } |
| 2447 | |
| 2448 | #[test] |
| 2449 | fn invalid_user_global_constitution_is_skipped() { |
| 2450 | let _env_guard = crate::test_support::lock_test_env(); |
| 2451 | let tmp = tempdir().expect("tempdir"); |
| 2452 | let workspace = tmp.path().join("workspace"); |
| 2453 | std::fs::create_dir_all(&workspace).expect("workspace dir"); |
| 2454 | let codewhale_home = tmp.path().join("codewhale-home"); |
| 2455 | std::fs::create_dir_all(&codewhale_home).expect("codewhale home"); |
| 2456 | std::fs::write( |
| 2457 | codewhale_home.join(codewhale_config::user_constitution::USER_CONSTITUTION_FILE_NAME), |
| 2458 | "{ not valid json", |
| 2459 | ) |
| 2460 | .expect("write invalid user constitution"); |
| 2461 | let _codewhale_home = |
| 2462 | crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", codewhale_home.as_os_str()); |
| 2463 | |
| 2464 | let prompt = |
| 2465 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2466 | &workspace, |
| 2467 | None, |
| 2468 | None, |
| 2469 | None, |
| 2470 | PromptSessionContext { |
| 2471 | project_context_pack_enabled: false, |
| 2472 | ..PromptSessionContext::default() |
| 2473 | }, |
| 2474 | )); |
| 2475 | |
| 2476 | assert!(!prompt.contains("<codewhale_user_constitution")); |
| 2477 | } |
| 2478 | |
| 2479 | #[test] |
| 2480 | fn memory_guidance_carries_paired_examples() { |
| 2481 | // The fragment is the contract — verify the verbatim ✓ / ✗ |
| 2482 | // pair is present so V4 has both shapes to imitate. |
| 2483 | assert!(MEMORY_GUIDANCE.contains("declarative facts")); |
| 2484 | assert!(MEMORY_GUIDANCE.contains(" ✓")); |
| 2485 | assert!(MEMORY_GUIDANCE.contains(" ✗")); |
| 2486 | assert!(MEMORY_GUIDANCE.contains("Imperative")); |
| 2487 | } |
| 2488 | |
| 2489 | #[test] |
| 2490 | fn memory_guidance_does_not_reference_scrapped_moraine() { |
| 2491 | // Moraine was scrapped for v0.9.4 (no in-repo server ever existed); |
| 2492 | // the native Markdown + SQLite FTS5 memory is the surviving system. |
| 2493 | assert!(!MEMORY_GUIDANCE.contains("Moraine")); |
| 2494 | assert!(!MEMORY_GUIDANCE.contains("moraine")); |
| 2495 | } |
| 2496 | |
| 2497 | #[test] |
| 2498 | fn memory_guidance_absent_when_no_memory_block() { |
| 2499 | let tmp = tempdir().expect("tempdir"); |
| 2500 | let prompt = |
| 2501 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2502 | tmp.path(), |
| 2503 | None, |
| 2504 | None, |
| 2505 | None, |
| 2506 | PromptSessionContext { |
| 2507 | user_memory_block: None, |
| 2508 | goal_objective: None, |
| 2509 | project_context_pack_enabled: false, |
| 2510 | locale_tag: "en", |
| 2511 | translation_enabled: false, |
| 2512 | model_id: "codewhale", |
| 2513 | context_window_override: None, |
| 2514 | verbosity: None, |
| 2515 | skills_scan_codewhale_only: false, |
| 2516 | plugin_registry: None, |
| 2517 | mode: crate::tui::app::AppMode::Agent, |
| 2518 | }, |
| 2519 | )); |
| 2520 | assert!( |
| 2521 | !prompt.contains("Memory Hygiene"), |
| 2522 | "memory guidance must not leak into sessions without a memory block" |
| 2523 | ); |
| 2524 | } |
| 2525 | |
| 2526 | #[test] |
| 2527 | fn memory_guidance_appended_after_memory_block() { |
| 2528 | let tmp = tempdir().expect("tempdir"); |
| 2529 | let block = "## User Memory\n\n- prefers Rust\n"; |
| 2530 | let prompt = |
| 2531 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2532 | tmp.path(), |
| 2533 | None, |
| 2534 | None, |
| 2535 | None, |
| 2536 | PromptSessionContext { |
| 2537 | user_memory_block: Some(block), |
| 2538 | goal_objective: None, |
| 2539 | project_context_pack_enabled: false, |
| 2540 | locale_tag: "en", |
| 2541 | translation_enabled: false, |
| 2542 | model_id: "codewhale", |
| 2543 | context_window_override: None, |
| 2544 | verbosity: None, |
| 2545 | skills_scan_codewhale_only: false, |
| 2546 | plugin_registry: None, |
| 2547 | mode: crate::tui::app::AppMode::Agent, |
| 2548 | }, |
| 2549 | )); |
| 2550 | let mem_at = prompt.find("User Memory").expect("user memory present"); |
| 2551 | let guide_at = prompt.find("Memory Hygiene").expect("guidance present"); |
| 2552 | assert!( |
| 2553 | mem_at < guide_at, |
| 2554 | "guidance must come after the user memory block" |
| 2555 | ); |
| 2556 | } |
| 2557 | |
| 2558 | #[test] |
| 2559 | fn continual_harness_is_injected_as_untrusted_world_state() { |
| 2560 | let tmp = tempdir().expect("tempdir"); |
| 2561 | crate::continual_harness::refine( |
| 2562 | tmp.path(), |
| 2563 | crate::continual_harness::HarnessRefinement { |
| 2564 | kind: crate::continual_harness::HarnessEntryKind::PromptNote, |
| 2565 | title: "Verify release claims from direct evidence".to_string(), |
| 2566 | content: "Retain exact current command output for each release gate.".to_string(), |
| 2567 | evidence: |
| 2568 | "A prior release report mixed stale hosted CI with newer local test output." |
| 2569 | .to_string(), |
| 2570 | }, |
| 2571 | ) |
| 2572 | .expect("persist harness state"); |
| 2573 | |
| 2574 | let prompt = |
| 2575 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2576 | tmp.path(), |
| 2577 | None, |
| 2578 | None, |
| 2579 | None, |
| 2580 | PromptSessionContext { |
| 2581 | user_memory_block: None, |
| 2582 | goal_objective: None, |
| 2583 | project_context_pack_enabled: false, |
| 2584 | locale_tag: "en", |
| 2585 | translation_enabled: false, |
| 2586 | model_id: "codewhale", |
| 2587 | context_window_override: None, |
| 2588 | verbosity: None, |
| 2589 | skills_scan_codewhale_only: false, |
| 2590 | plugin_registry: None, |
| 2591 | mode: crate::tui::app::AppMode::Agent, |
| 2592 | }, |
| 2593 | )); |
| 2594 | assert!(prompt.contains("<continual_harness trust=\"untrusted\">")); |
| 2595 | assert!(prompt.contains("supplemental working guidance")); |
| 2596 | assert!(prompt.contains("Verify release claims from direct evidence")); |
| 2597 | } |
| 2598 | |
| 2599 | #[test] |
| 2600 | fn memory_guidance_does_not_state_precedence() { |
| 2601 | // #4777: only BASE_PROMPT § Whose word wins states ranks. Memory |
| 2602 | // hygiene keeps the imperative→preference rule and drops the |
| 2603 | // inverted Tier list that used to put Constitution above the user. |
| 2604 | let guidance = MEMORY_GUIDANCE.to_ascii_lowercase(); |
| 2605 | for forbidden in [ |
| 2606 | "tier 1", |
| 2607 | "tier 2", |
| 2608 | "tier 7", |
| 2609 | "statute", |
| 2610 | "regulation", |
| 2611 | "local law", |
| 2612 | "constitutional hierarchy", |
| 2613 | ] { |
| 2614 | assert!( |
| 2615 | !guidance.contains(forbidden), |
| 2616 | "MEMORY_GUIDANCE must not restate ranks (found {forbidden:?})" |
| 2617 | ); |
| 2618 | } |
| 2619 | assert!( |
| 2620 | MEMORY_GUIDANCE.contains("treated as a preference") |
| 2621 | && MEMORY_GUIDANCE.contains("not a command"), |
| 2622 | "keep the imperative-as-preference rule" |
| 2623 | ); |
| 2624 | } |
| 2625 | |
| 2626 | #[test] |
| 2627 | fn only_the_constitution_states_precedence() { |
| 2628 | // Composed overlays must describe behavior, never their own rank. |
| 2629 | let overlays = [ |
| 2630 | ("CALM_PERSONALITY", CALM_PERSONALITY), |
| 2631 | ("AGENT_MODE", AGENT_MODE), |
| 2632 | ("PLAN_MODE", PLAN_MODE), |
| 2633 | ("OPERATE_MODE", OPERATE_MODE), |
| 2634 | ("COMPACT_TEMPLATE", COMPACT_TEMPLATE), |
| 2635 | ("MEMORY_GUIDANCE", MEMORY_GUIDANCE), |
| 2636 | ("LANGUAGE_PROMPT", LANGUAGE_PROMPT), |
| 2637 | ("OUTPUT_PROMPT", OUTPUT_PROMPT), |
| 2638 | ("AUTHORITY_RECAP", AUTHORITY_RECAP), |
| 2639 | ]; |
| 2640 | let rank_markers = [ |
| 2641 | "Tier 1", |
| 2642 | "Tier 2", |
| 2643 | "Tier 3", |
| 2644 | "Tier 4", |
| 2645 | "Tier 5", |
| 2646 | "Tier 6", |
| 2647 | "Tier 7", |
| 2648 | "Tier 8", |
| 2649 | "Tier 9", |
| 2650 | "Statute", |
| 2651 | "Article IV", |
| 2652 | "Article V", |
| 2653 | "Article VII", |
| 2654 | "Local Law", |
| 2655 | "Regulation (Tier", |
| 2656 | ]; |
| 2657 | for (name, text) in overlays { |
| 2658 | for marker in rank_markers { |
| 2659 | assert!( |
| 2660 | !text.contains(marker), |
| 2661 | "{name} must not carry rank vocabulary {marker:?}" |
| 2662 | ); |
| 2663 | } |
| 2664 | } |
| 2665 | assert!( |
| 2666 | BASE_PROMPT.contains("### Whose word wins"), |
| 2667 | "canonical precedence section must remain in BASE_PROMPT" |
| 2668 | ); |
| 2669 | assert!( |
| 2670 | BASE_PROMPT.contains("This ordering is stated here and nowhere else"), |
| 2671 | "BASE_PROMPT must assert single-source precedence" |
| 2672 | ); |
| 2673 | } |
| 2674 | |
| 2675 | #[test] |
| 2676 | fn project_context_pack_can_be_disabled() { |
| 2677 | let tmp = tempdir().expect("tempdir"); |
| 2678 | std::fs::write(tmp.path().join("README.md"), "# Pack test").expect("write readme"); |
| 2679 | let prompt = |
| 2680 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2681 | tmp.path(), |
| 2682 | None, |
| 2683 | None, |
| 2684 | None, |
| 2685 | PromptSessionContext { |
| 2686 | user_memory_block: None, |
| 2687 | goal_objective: None, |
| 2688 | project_context_pack_enabled: false, |
| 2689 | locale_tag: "en", |
| 2690 | translation_enabled: false, |
| 2691 | model_id: "codewhale", |
| 2692 | context_window_override: None, |
| 2693 | verbosity: None, |
| 2694 | skills_scan_codewhale_only: false, |
| 2695 | plugin_registry: None, |
| 2696 | mode: crate::tui::app::AppMode::Agent, |
| 2697 | }, |
| 2698 | )); |
| 2699 | assert!(!prompt.contains("<project_context_pack>")); |
| 2700 | } |
| 2701 | |
| 2702 | #[test] |
| 2703 | fn project_context_pack_is_before_dynamic_tail() { |
| 2704 | let tmp = tempdir().expect("tempdir"); |
| 2705 | std::fs::write(tmp.path().join("README.md"), "# Pack test").expect("write readme"); |
| 2706 | std::fs::create_dir_all(tmp.path().join(".deepseek")).expect("mkdir"); |
| 2707 | std::fs::write(tmp.path().join(".deepseek").join("handoff.md"), "handoff") |
| 2708 | .expect("handoff"); |
| 2709 | let prompt = |
| 2710 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 2711 | tmp.path(), |
| 2712 | None, |
| 2713 | None, |
| 2714 | None, |
| 2715 | PromptSessionContext { |
| 2716 | user_memory_block: None, |
| 2717 | goal_objective: None, |
| 2718 | // Explicit opt-in — pack is off by default (#4781). |
| 2719 | project_context_pack_enabled: true, |
| 2720 | locale_tag: "en", |
| 2721 | translation_enabled: false, |
| 2722 | model_id: "codewhale", |
| 2723 | context_window_override: None, |
| 2724 | verbosity: None, |
| 2725 | skills_scan_codewhale_only: false, |
| 2726 | plugin_registry: None, |
| 2727 | mode: crate::tui::app::AppMode::Agent, |
| 2728 | }, |
| 2729 | )); |
| 2730 | assert!(prompt.contains("<project_context_pack>")); |
| 2731 | assert!( |
| 2732 | prompt.find("<project_context_pack>").expect("pack") |
| 2733 | < prompt.find("## Previous Session Relay").expect("relay") |
| 2734 | ); |
| 2735 | } |
| 2736 | |
| 2737 | #[test] |
| 2738 | fn handoff_artifact_is_prepended_to_system_prompt_when_present() { |
| 2739 | let tmp = tempdir().expect("tempdir"); |
| 2740 | let workspace = tmp.path(); |
| 2741 | let handoff_dir = workspace.join(".deepseek"); |
| 2742 | std::fs::create_dir_all(&handoff_dir).unwrap(); |
| 2743 | std::fs::write( |
| 2744 | handoff_dir.join("handoff.md"), |
| 2745 | "# Session relay — prior\n\n## Active task\nFinish #32.\n\n## Open blockers\n- [ ] write the basic version\n", |
| 2746 | ) |
| 2747 | .unwrap(); |
| 2748 | |
| 2749 | let prompt = system_prompt_flat_text(&system_prompt_for_mode_with_context(workspace, None)); |
| 2750 | |
| 2751 | assert!(prompt.contains(HANDOFF_BLOCK_MARKER)); |
| 2752 | assert!(prompt.contains("Finish #32.")); |
| 2753 | assert!(prompt.contains("write the basic version")); |
| 2754 | } |
| 2755 | |
| 2756 | #[test] |
| 2757 | fn missing_handoff_does_not_inject_block() { |
| 2758 | let tmp = tempdir().expect("tempdir"); |
| 2759 | let prompt = |
| 2760 | system_prompt_flat_text(&system_prompt_for_mode_with_context(tmp.path(), None)); |
| 2761 | assert!(!prompt.contains(HANDOFF_BLOCK_MARKER)); |
| 2762 | } |
| 2763 | |
| 2764 | #[test] |
| 2765 | fn empty_handoff_file_does_not_inject_block() { |
| 2766 | let tmp = tempdir().expect("tempdir"); |
| 2767 | let dir = tmp.path().join(".deepseek"); |
| 2768 | std::fs::create_dir_all(&dir).unwrap(); |
| 2769 | std::fs::write(dir.join("handoff.md"), " \n\n ").unwrap(); |
| 2770 | let prompt = |
| 2771 | system_prompt_flat_text(&system_prompt_for_mode_with_context(tmp.path(), None)); |
| 2772 | assert!(!prompt.contains(HANDOFF_BLOCK_MARKER)); |
| 2773 | } |
| 2774 | |
| 2775 | #[test] |
| 2776 | fn compose_prompt_includes_all_layers() { |
| 2777 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 2778 | // Base layer — balanced Constitution; procedural recipes stay out. |
| 2779 | assert!(prompt.contains("## Codewhale")); |
| 2780 | assert!(prompt.contains("### Whose word wins")); |
| 2781 | assert!(!prompt.contains("## STATUTES (Tier 2)")); |
| 2782 | assert!(!prompt.contains("## EVIDENCE (Tier 6)")); |
| 2783 | // Mode and approval are not inlined — they travel as |
| 2784 | // request-time runtime metadata. |
| 2785 | assert!(!prompt.contains("Mode: Agent")); |
| 2786 | assert!(!prompt.contains("Approval Policy:")); |
| 2787 | } |
| 2788 | |
| 2789 | /// `constitution.md` is the single hand-maintained source of the balanced |
| 2790 | /// constitutional core. This replaces the old 600-line policy tail: a |
| 2791 | /// hand-edit that drops a core section or reorders the skeleton fails the |
| 2792 | /// build instead of silently shipping a malformed prompt. |
| 2793 | #[test] |
| 2794 | fn constitution_md_carries_required_structure() { |
| 2795 | let md = BASE_PROMPT; |
| 2796 | assert!(md.contains("## Codewhale"), "missing title"); |
| 2797 | let mut cursor = 0usize; |
| 2798 | for needle in [ |
| 2799 | "## Codewhale", |
| 2800 | "### Ground truth", |
| 2801 | "### User intent and scope", |
| 2802 | "### Truthful completion", |
| 2803 | "### Put guarantees in mechanism", |
| 2804 | "### Whose word wins", |
| 2805 | ] { |
| 2806 | let pos = md |
| 2807 | .find(needle) |
| 2808 | .unwrap_or_else(|| panic!("ordering check: {needle:?} not found")); |
| 2809 | assert!( |
| 2810 | pos >= cursor, |
| 2811 | "cache-stable ordering broken: {needle:?} at {pos} precedes a previous section at {cursor}" |
| 2812 | ); |
| 2813 | cursor = pos + needle.len(); |
| 2814 | } |
| 2815 | } |
| 2816 | |
| 2817 | /// Gate against shipping a release with a missing CHANGELOG entry — which |
| 2818 | /// is exactly what happened with v0.8.21 / v0.8.22 (entries had to be |
| 2819 | /// backfilled in v0.8.23). Asserts the top-of-file CHANGELOG contains a |
| 2820 | /// `## [X.Y.Z]` heading matching the current `CARGO_PKG_VERSION`. No |
| 2821 | /// hardcoded version string — the test self-updates with the workspace |
| 2822 | /// version bump and only fires when the CHANGELOG is the missing piece. |
| 2823 | /// |
| 2824 | /// Walks up from `CARGO_MANIFEST_DIR` to find `CHANGELOG.md` instead of |
| 2825 | /// assuming a fixed `../../CHANGELOG.md` layout. The workspace root is |
| 2826 | /// the common case, but the walk also tolerates deeper crate layouts and |
| 2827 | /// the packaged-crate case (where the workspace root has been stripped |
| 2828 | /// out): if no `CHANGELOG.md` is reachable, the gate quietly skips |
| 2829 | /// rather than panicking, so consumers running the suite outside the |
| 2830 | /// workspace checkout don't see a spurious failure. |
| 2831 | #[test] |
| 2832 | fn changelog_entry_exists_for_current_package_version() { |
| 2833 | let version = env!("CARGO_PKG_VERSION"); |
| 2834 | let manifest_dir = std::path::Path::new(env!("CARGO_MANIFEST_DIR")); |
| 2835 | let Some(changelog_path) = manifest_dir |
| 2836 | .ancestors() |
| 2837 | .map(|dir| dir.join("CHANGELOG.md")) |
| 2838 | .find(|candidate| candidate.is_file()) |
| 2839 | else { |
| 2840 | eprintln!( |
| 2841 | "changelog_entry_exists_for_current_package_version: no \ |
| 2842 | CHANGELOG.md found above {} — skipping (this gate only \ |
| 2843 | fires inside a workspace checkout).", |
| 2844 | manifest_dir.display() |
| 2845 | ); |
| 2846 | return; |
| 2847 | }; |
| 2848 | |
| 2849 | let contents = std::fs::read_to_string(&changelog_path).unwrap_or_else(|err| { |
| 2850 | panic!( |
| 2851 | "failed to read CHANGELOG.md at {}: {err}", |
| 2852 | changelog_path.display() |
| 2853 | ) |
| 2854 | }); |
| 2855 | let header = format!("## [{version}]"); |
| 2856 | assert!( |
| 2857 | contents.contains(&header), |
| 2858 | "CHANGELOG.md is missing a `{header}` entry for the current package \ |
| 2859 | version. Add a release section at the top before tagging — see \ |
| 2860 | docs/RELEASE_CHECKLIST.md." |
| 2861 | ); |
| 2862 | } |
| 2863 | |
| 2864 | #[test] |
| 2865 | fn compose_prompt_deterministic_order() { |
| 2866 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 2867 | let base_pos = prompt.find("## Codewhale").unwrap(); |
| 2868 | let article_pos = prompt.find("### Ground truth").unwrap(); |
| 2869 | |
| 2870 | assert!(base_pos < article_pos); |
| 2871 | } |
| 2872 | |
| 2873 | #[test] |
| 2874 | fn base_prompt_is_mode_agnostic() { |
| 2875 | // Mode and approval text are no longer inlined into compose_prompt — |
| 2876 | // they travel as request-time runtime metadata. |
| 2877 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 2878 | assert!(!prompt.contains("Mode: Agent")); |
| 2879 | assert!(!prompt.contains("Mode: YOLO")); |
| 2880 | assert!(!prompt.contains("Mode: Plan")); |
| 2881 | assert!(!prompt.contains("Approval Policy:")); |
| 2882 | // Base prompt carries the 0.9.0 compact Constitution. |
| 2883 | assert!(prompt.contains("You are Codewhale")); |
| 2884 | assert!(prompt.contains("Take the work seriously. Don't take")); |
| 2885 | } |
| 2886 | |
| 2887 | #[test] |
| 2888 | fn agent_mode_prompt_keeps_safety_invariants_after_compression() { |
| 2889 | let prompt = AGENT_MODE.replace("\r\n", "\n").replace('\r', "\n"); |
| 2890 | for must in [ |
| 2891 | "autonomously", |
| 2892 | "tools in the current catalog", |
| 2893 | "work_update", |
| 2894 | "current catalog includes delegation", |
| 2895 | "Do not announce the mode", |
| 2896 | ] { |
| 2897 | assert!( |
| 2898 | prompt.contains(must), |
| 2899 | "compressed agent mode missing invariant {must:?}" |
| 2900 | ); |
| 2901 | } |
| 2902 | for unavailable_claim in ["`File`", "`Git`", "`Run`", "`Bash`"] { |
| 2903 | assert!(!prompt.contains(unavailable_claim)); |
| 2904 | } |
| 2905 | // Procedural PowerShell manuals must not live in the mode delta. |
| 2906 | for forbidden in ["Invoke-Expression", "pwsh.exe -NoLogo", "ProcessStartInfo"] { |
| 2907 | assert!( |
| 2908 | !prompt.contains(forbidden), |
| 2909 | "agent mode must not absorb PowerShell manuals: {forbidden}" |
| 2910 | ); |
| 2911 | } |
| 2912 | } |
| 2913 | |
| 2914 | #[test] |
| 2915 | fn mode_prompts_remain_small_deltas_not_base_policy_copies() { |
| 2916 | assert!( |
| 2917 | PLAN_MODE.contains("All writes, patches, shell commands"), |
| 2918 | "Plan may summarize the user-facing mode delta" |
| 2919 | ); |
| 2920 | for (name, prompt) in [("agent", AGENT_MODE), ("plan", PLAN_MODE)] { |
| 2921 | // Measure semantic size on LF so Windows autocrlf checkouts do not |
| 2922 | // inflate char/3 token estimates via extra `\r` bytes. |
| 2923 | let normalized = prompt.replace("\r\n", "\n").replace('\r', "\n"); |
| 2924 | let word_count = normalized.split_whitespace().count(); |
| 2925 | let estimated_tokens = |
| 2926 | crate::compaction::estimate_text_tokens_conservative(&normalized); |
| 2927 | // 2026-07-21: mode deltas contain permissions and durable behavior |
| 2928 | // only. Action recipes belong to the canonical tool schemas. |
| 2929 | let max_words = 120; |
| 2930 | let max_tokens = 320; |
| 2931 | |
| 2932 | assert!( |
| 2933 | word_count <= max_words, |
| 2934 | "{name} mode prompt should remain a delta, got {word_count} words" |
| 2935 | ); |
| 2936 | assert!( |
| 2937 | estimated_tokens <= max_tokens, |
| 2938 | "{name} mode prompt should remain compact, got {estimated_tokens} estimated tokens" |
| 2939 | ); |
| 2940 | for forbidden in [ |
| 2941 | "## Codewhale", |
| 2942 | "## STATUTES (Tier 2)", |
| 2943 | "## REGULATIONS (Tier 3)", |
| 2944 | "## EVIDENCE (Tier 6)", |
| 2945 | "## Context Management", |
| 2946 | "## Runtime Policy Reference", |
| 2947 | ] { |
| 2948 | assert!( |
| 2949 | !normalized.contains(forbidden), |
| 2950 | "{name} mode prompt duplicated shared base section {forbidden:?}" |
| 2951 | ); |
| 2952 | } |
| 2953 | } |
| 2954 | } |
| 2955 | |
| 2956 | #[test] |
| 2957 | fn approval_policy_no_longer_inlined_in_base_prompt() { |
| 2958 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 2959 | assert!(!prompt.contains("Mode: Agent")); |
| 2960 | assert!(!prompt.contains("Approval Policy:")); |
| 2961 | // The compact Constitutional preamble is still present. |
| 2962 | assert!(prompt.contains("You are Codewhale")); |
| 2963 | } |
| 2964 | |
| 2965 | #[test] |
| 2966 | fn execution_contract_states_proposal_is_not_execution() { |
| 2967 | // #5146: the live execution layer must make the propose-vs-execute |
| 2968 | // contract explicit after the legacy approval overlay was removed. |
| 2969 | assert!( |
| 2970 | CORE_EXECUTION_PROFILE_PROMPT.contains("is the proposal, not the execution"), |
| 2971 | "Execution profile must state the propose-vs-execute contract" |
| 2972 | ); |
| 2973 | assert!( |
| 2974 | CORE_EXECUTION_PROFILE_PROMPT.contains("present the change in your plan"), |
| 2975 | "Execution profile must name the correct behavior on rejection" |
| 2976 | ); |
| 2977 | } |
| 2978 | |
| 2979 | #[test] |
| 2980 | fn personality_is_folded_into_constitution() { |
| 2981 | // v4 has no separate personality tier. Voice and tone live in |
| 2982 | // the preamble, so composition appends no personality overlay. |
| 2983 | let calm = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 2984 | assert!(!calm.contains("## Personality:")); |
| 2985 | assert!(calm.contains("Take the work seriously. Don't take")); |
| 2986 | assert!(calm.contains("You are Codewhale")); |
| 2987 | } |
| 2988 | |
| 2989 | #[test] |
| 2990 | fn compact_template_is_lazy_in_fresh_prompt() { |
| 2991 | let tmp = tempdir().expect("tempdir"); |
| 2992 | let prompt = |
| 2993 | system_prompt_flat_text(&system_prompt_for_mode_with_context(tmp.path(), None)); |
| 2994 | assert!(!prompt.contains("# Session relay")); |
| 2995 | assert!(!prompt.contains("## Verification")); |
| 2996 | } |
| 2997 | |
| 2998 | #[test] |
| 2999 | fn session_goal_stays_volatile_while_compact_template_is_lazy() { |
| 3000 | let tmp = tempdir().expect("tempdir"); |
| 3001 | let prompt = |
| 3002 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 3003 | tmp.path(), |
| 3004 | Some("## Repo Working Set\nsrc/lib.rs"), |
| 3005 | None, |
| 3006 | None, |
| 3007 | PromptSessionContext { |
| 3008 | user_memory_block: None, |
| 3009 | goal_objective: Some("Fix transcript corruption"), |
| 3010 | project_context_pack_enabled: false, |
| 3011 | locale_tag: "en", |
| 3012 | translation_enabled: false, |
| 3013 | model_id: "codewhale", |
| 3014 | context_window_override: None, |
| 3015 | verbosity: None, |
| 3016 | skills_scan_codewhale_only: false, |
| 3017 | plugin_registry: None, |
| 3018 | mode: crate::tui::app::AppMode::Agent, |
| 3019 | }, |
| 3020 | )); |
| 3021 | |
| 3022 | let goal_pos = prompt.find("<session_goal>").expect("goal block"); |
| 3023 | assert!(prompt.contains("Fix transcript corruption")); |
| 3024 | // Session goal remains volatile content below the stable static |
| 3025 | // layers. The relay template is injected only when relay/compaction |
| 3026 | // actually needs it. |
| 3027 | assert!(goal_pos > 0); |
| 3028 | assert!(!prompt.contains("# Session relay")); |
| 3029 | assert!(!prompt.contains("src/lib.rs")); |
| 3030 | } |
| 3031 | |
| 3032 | #[test] |
| 3033 | fn empty_session_goal_is_not_injected() { |
| 3034 | let tmp = tempdir().expect("tempdir"); |
| 3035 | let prompt = |
| 3036 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 3037 | tmp.path(), |
| 3038 | None, |
| 3039 | None, |
| 3040 | None, |
| 3041 | PromptSessionContext { |
| 3042 | user_memory_block: None, |
| 3043 | goal_objective: Some(" "), |
| 3044 | project_context_pack_enabled: false, |
| 3045 | locale_tag: "en", |
| 3046 | translation_enabled: false, |
| 3047 | model_id: "codewhale", |
| 3048 | context_window_override: None, |
| 3049 | verbosity: None, |
| 3050 | skills_scan_codewhale_only: false, |
| 3051 | plugin_registry: None, |
| 3052 | mode: crate::tui::app::AppMode::Agent, |
| 3053 | }, |
| 3054 | )); |
| 3055 | |
| 3056 | assert!(!prompt.contains("<session_goal>")); |
| 3057 | assert!(!prompt.contains("## Current Goal")); |
| 3058 | } |
| 3059 | |
| 3060 | #[test] |
| 3061 | fn agent_mode_tool_guidance_avoids_defensive_tool_suppression() { |
| 3062 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3063 | assert!(!prompt.contains("Tool Selection Guide")); |
| 3064 | for tool in ["`File`", "`Git`", "`Run`", "`Bash`"] { |
| 3065 | assert!(!AGENT_MODE.contains(tool)); |
| 3066 | } |
| 3067 | assert!(AGENT_MODE.contains("tools in the current catalog")); |
| 3068 | for legacy in ["read_file", "git_status", "run_tests", "exec_shell"] { |
| 3069 | assert!(!AGENT_MODE.contains(legacy)); |
| 3070 | } |
| 3071 | assert!( |
| 3072 | !AGENT_MODE.contains("When NOT to use certain tools"), |
| 3073 | "agent mode should steer tool choice without training the model to avoid available tools" |
| 3074 | ); |
| 3075 | assert!( |
| 3076 | !AGENT_MODE.contains("Don't reach for"), |
| 3077 | "avoid defensive anti-tool wording in mode guidance" |
| 3078 | ); |
| 3079 | } |
| 3080 | |
| 3081 | /// #588: after the 0.9.0 constitution reduction, language-mirroring |
| 3082 | /// reinforcement lives in its own static segment plus locale bookends. |
| 3083 | #[test] |
| 3084 | fn language_segment_present_outside_reduced_constitution() { |
| 3085 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3086 | assert!( |
| 3087 | !BASE_PROMPT.contains("## Language"), |
| 3088 | "0.9.0 constitution.md should stay reduced; language belongs in its own segment" |
| 3089 | ); |
| 3090 | assert!( |
| 3091 | LANGUAGE_PROMPT.contains("## Language") && prompt.contains("## Language"), |
| 3092 | "default static prompt must still include the language segment" |
| 3093 | ); |
| 3094 | assert!( |
| 3095 | LANGUAGE_PROMPT.contains("latest user message") |
| 3096 | && LANGUAGE_PROMPT.contains("fallback, not an override") |
| 3097 | && LANGUAGE_PROMPT.contains("localized READMEs") |
| 3098 | && LANGUAGE_PROMPT.contains("Use the `lang` field only when") |
| 3099 | && LANGUAGE_PROMPT.contains("constitution and other system law stay English"), |
| 3100 | "language segment must keep the mirror contract while staying short (#4784)" |
| 3101 | ); |
| 3102 | assert!( |
| 3103 | LANGUAGE_PROMPT.contains("reasoning_content") |
| 3104 | && prompt.contains("reasoning_content") |
| 3105 | && LOCALE_PREAMBLE_ZH_HANS.contains("reasoning_content") |
| 3106 | && LOCALE_CLOSER_ZH_HANS.contains("reasoning_content"), |
| 3107 | "language segment and locale bookends must keep the reasoning_content anchor" |
| 3108 | ); |
| 3109 | } |
| 3110 | |
| 3111 | #[test] |
| 3112 | fn output_formatting_segment_present_outside_reduced_constitution() { |
| 3113 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3114 | assert!( |
| 3115 | !BASE_PROMPT.contains("## Output Formatting"), |
| 3116 | "0.9.0 constitution.md should stay reduced; output formatting belongs in its own segment" |
| 3117 | ); |
| 3118 | assert!(OUTPUT_PROMPT.contains("## Output Formatting")); |
| 3119 | assert!(prompt.contains("## Output Formatting")); |
| 3120 | assert!(prompt.contains("terminal, not a browser")); |
| 3121 | assert!(prompt.contains("Markdown tables almost never render correctly")); |
| 3122 | } |
| 3123 | |
| 3124 | #[test] |
| 3125 | fn runtime_prompt_assembly_preserves_split_static_layers() { |
| 3126 | let tmp = tempdir().expect("tempdir"); |
| 3127 | let prompt = |
| 3128 | system_prompt_flat_text(&system_prompt_for_mode_with_context_skills_and_session( |
| 3129 | tmp.path(), |
| 3130 | None, |
| 3131 | None, |
| 3132 | None, |
| 3133 | PromptSessionContext { |
| 3134 | user_memory_block: None, |
| 3135 | goal_objective: None, |
| 3136 | project_context_pack_enabled: false, |
| 3137 | locale_tag: "en", |
| 3138 | translation_enabled: false, |
| 3139 | model_id: "glm-5.2", |
| 3140 | context_window_override: Some(1_000_000), |
| 3141 | verbosity: None, |
| 3142 | skills_scan_codewhale_only: false, |
| 3143 | plugin_registry: None, |
| 3144 | mode: crate::tui::app::AppMode::Agent, |
| 3145 | }, |
| 3146 | )); |
| 3147 | |
| 3148 | assert!(prompt.contains("## Codewhale")); |
| 3149 | assert!(prompt.contains("## Language")); |
| 3150 | assert!(prompt.contains("## Output Formatting")); |
| 3151 | assert!(prompt.contains("Use the `lang` field only when")); |
| 3152 | } |
| 3153 | |
| 3154 | #[test] |
| 3155 | fn locale_bookends_resist_english_context_drift() { |
| 3156 | assert!( |
| 3157 | LOCALE_PREAMBLE_ZH_HANS.contains("reasoning_content") |
| 3158 | && LOCALE_CLOSER_ZH_HANS.contains("reasoning_content"), |
| 3159 | "locale bookends must keep the reasoning_content anchor" |
| 3160 | ); |
| 3161 | assert!( |
| 3162 | LOCALE_CLOSER_ZH_HANS.contains("英文代码") |
| 3163 | && LOCALE_CLOSER_ZH_HANS.contains("用户的语言决定"), |
| 3164 | "closing locale bookend must explicitly resist English-context drift" |
| 3165 | ); |
| 3166 | assert!( |
| 3167 | LOCALE_PREAMBLE_ZH_HANS.contains("代码、文件路径、工具名称"), |
| 3168 | "opening locale bookend must keep code/tool tokens untranslated" |
| 3169 | ); |
| 3170 | } |
| 3171 | |
| 3172 | #[test] |
| 3173 | fn english_base_prompt_avoids_native_script_language_priming() { |
| 3174 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3175 | assert!( |
| 3176 | !contains_cjk(&prompt), |
| 3177 | "English base prompt should keep native-script reinforcement in locale bookends only" |
| 3178 | ); |
| 3179 | assert!( |
| 3180 | !prompt.contains("multilingual coding agent"), |
| 3181 | "identity should not prime language switching; language belongs in runtime bookends" |
| 3182 | ); |
| 3183 | } |
| 3184 | |
| 3185 | #[test] |
| 3186 | fn legacy_rlm_compatibility_descriptions_remain_available() { |
| 3187 | assert!(!AGENT_MODE.contains("Large Context Tools")); |
| 3188 | |
| 3189 | let descriptions = [ |
| 3190 | RlmTool::alias("rlm_open", "open", None) |
| 3191 | .description() |
| 3192 | .to_string(), |
| 3193 | RlmTool::alias("rlm_eval", "eval", None) |
| 3194 | .description() |
| 3195 | .to_string(), |
| 3196 | RlmTool::alias("rlm_configure", "configure", None) |
| 3197 | .description() |
| 3198 | .to_string(), |
| 3199 | RlmTool::alias("rlm_close", "close", None) |
| 3200 | .description() |
| 3201 | .to_string(), |
| 3202 | HandleReadTool.description().to_string(), |
| 3203 | ] |
| 3204 | .join("\n"); |
| 3205 | let rlm_count = descriptions.to_lowercase().matches("rlm").count(); |
| 3206 | assert!( |
| 3207 | rlm_count >= 5, |
| 3208 | "RLM tool descriptions present: expected >= 5 mentions of 'rlm', got {rlm_count}" |
| 3209 | ); |
| 3210 | assert!( |
| 3211 | !AGENT_MODE.contains("`rlm`"), |
| 3212 | "the normal Agent prompt should not teach the retired RLM control surface" |
| 3213 | ); |
| 3214 | } |
| 3215 | |
| 3216 | /// Project instructions rank above memory, with the nearest scope winning |
| 3217 | /// over the broader. The embedder-injected-instructions case is covered |
| 3218 | /// by project law/instructions sitting above memory/handoffs. |
| 3219 | #[test] |
| 3220 | fn project_instructions_outrank_memory_in_whose_word_wins() { |
| 3221 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3222 | let project_at = prompt |
| 3223 | .find("3. Project law and instructions") |
| 3224 | .expect("Whose word wins must rank project instructions"); |
| 3225 | let memory_at = prompt |
| 3226 | .find("5. Memory and previous-session handoffs.") |
| 3227 | .expect("Whose word wins must rank memory below project instructions"); |
| 3228 | assert!( |
| 3229 | project_at < memory_at, |
| 3230 | "project instructions must outrank memory so embedder-injected \ |
| 3231 | instructions are not treated as mere memory preferences" |
| 3232 | ); |
| 3233 | } |
| 3234 | |
| 3235 | #[test] |
| 3236 | fn workspace_orientation_guidance_present() { |
| 3237 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3238 | assert!(prompt.contains("Project law and instructions")); |
| 3239 | assert!( |
| 3240 | prompt.contains("the nearest in\nscope winning over the broader") |
| 3241 | || prompt.contains("the nearest in scope winning over the broader"), |
| 3242 | "Whose word wins must keep the nearest-scope-wins rule for project instructions" |
| 3243 | ); |
| 3244 | } |
| 3245 | |
| 3246 | #[test] |
| 3247 | fn prompt_uses_the_single_agent_kernel_surface() { |
| 3248 | for tool in ["rlm_open", "rlm_eval", "rlm_configure", "rlm_close"] { |
| 3249 | assert!(!AGENT_MODE.contains(tool)); |
| 3250 | } |
| 3251 | assert!(AGENT_MODE.contains("sub-agent")); |
| 3252 | assert!(AGENT_MODE.contains("session-persistent `repl` blocks")); |
| 3253 | } |
| 3254 | |
| 3255 | #[test] |
| 3256 | fn prompt_documents_fork_context_prefix_cache_contract() { |
| 3257 | let source = include_str!("tools/subagent/mod.rs"); |
| 3258 | assert!(source.contains("fork_context")); |
| 3259 | assert!(!AGENT_MODE.contains("fork_context")); |
| 3260 | } |
| 3261 | |
| 3262 | #[test] |
| 3263 | fn prompt_documents_explicit_subagent_model_strength() { |
| 3264 | let source = include_str!("tools/subagent/mod.rs"); |
| 3265 | assert!(source.contains("model_strength")); |
| 3266 | assert!(!AGENT_MODE.contains("model_strength")); |
| 3267 | } |
| 3268 | |
| 3269 | #[test] |
| 3270 | fn prompt_documents_structured_subagent_briefs() { |
| 3271 | assert!(!AGENT_MODE.contains("Subagent Brief")); |
| 3272 | for heading in [ |
| 3273 | "### SUMMARY", |
| 3274 | "### EVIDENCE", |
| 3275 | "### CHANGES", |
| 3276 | "### RISKS", |
| 3277 | "### BLOCKERS", |
| 3278 | ] { |
| 3279 | assert!(text::SUBAGENT_OUTPUT_FORMAT.contains(heading)); |
| 3280 | } |
| 3281 | } |
| 3282 | |
| 3283 | #[test] |
| 3284 | fn prompt_bounds_explore_without_tiny_cap_for_implementers() { |
| 3285 | assert!(AGENT_MODE.contains("current catalog includes delegation")); |
| 3286 | assert!(!AGENT_MODE.contains("3-5 tool calls")); |
| 3287 | assert!(!AGENT_MODE.contains("No fan-out without a fan-in owner")); |
| 3288 | } |
| 3289 | |
| 3290 | #[test] |
| 3291 | fn agent_mode_prompt_teaches_automatic_workflow_use() { |
| 3292 | for recipe in [ |
| 3293 | "Workflow", |
| 3294 | "responseSchema", |
| 3295 | "request_user_input", |
| 3296 | ".workflow.js", |
| 3297 | ] { |
| 3298 | assert!(!AGENT_MODE.contains(recipe)); |
| 3299 | } |
| 3300 | } |
| 3301 | |
| 3302 | #[test] |
| 3303 | fn operate_mode_prompt_keeps_multitask_simple_and_async() { |
| 3304 | for phrase in [ |
| 3305 | "dispatch, join,", |
| 3306 | "capabilities present in the current catalog", |
| 3307 | "When worker dispatch is available", |
| 3308 | "Fan out, block on one wait until the batch lands", |
| 3309 | "queued user messages as new tasks", |
| 3310 | "Preserve approval", |
| 3311 | "settled is not", |
| 3312 | "internal control-plane mechanics", |
| 3313 | "When goal control is available", |
| 3314 | "Dispatch is not completion", |
| 3315 | "verification capabilities", |
| 3316 | "When an ordered Workflow capability is present", |
| 3317 | ] { |
| 3318 | assert!( |
| 3319 | OPERATE_MODE.contains(phrase), |
| 3320 | "OPERATE_MODE missing multitask phrase {phrase:?}" |
| 3321 | ); |
| 3322 | } |
| 3323 | for implementation_detail in [ |
| 3324 | "risk` is exactly", |
| 3325 | "parallel([() =>", |
| 3326 | "terminal Workflow receipt", |
| 3327 | "/multitask", |
| 3328 | ] { |
| 3329 | assert!( |
| 3330 | !OPERATE_MODE.contains(implementation_detail), |
| 3331 | "OPERATE_MODE leaks implementation detail {implementation_detail:?}" |
| 3332 | ); |
| 3333 | } |
| 3334 | } |
| 3335 | |
| 3336 | /// The owner watched a model reason that it had to stay available for its |
| 3337 | /// children instead of simply joining them. Operate must not frame a |
| 3338 | /// blocking join as a lapse: the anti-pattern is the poll loop, not the |
| 3339 | /// single wait, and returning control mid-flight is the exception. |
| 3340 | #[test] |
| 3341 | fn operate_mode_endorses_one_blocking_join_over_staying_responsive() { |
| 3342 | for obligation in [ |
| 3343 | "keep the parent responsive", |
| 3344 | "parent responsive", |
| 3345 | "busy-waiting", |
| 3346 | "return control instead", |
| 3347 | ] { |
| 3348 | assert!( |
| 3349 | !OPERATE_MODE.contains(obligation), |
| 3350 | "OPERATE_MODE still frames staying responsive as an obligation: {obligation:?}" |
| 3351 | ); |
| 3352 | } |
| 3353 | for endorsement in [ |
| 3354 | "Fan out, block on one wait until the batch lands, then synthesize", |
| 3355 | "endorsed default", |
| 3356 | "Polling in a loop is the anti-pattern; one blocking wait", |
| 3357 | "Returning control mid-flight is the exception", |
| 3358 | ] { |
| 3359 | assert!( |
| 3360 | OPERATE_MODE.contains(endorsement), |
| 3361 | "OPERATE_MODE missing fan-out/join endorsement {endorsement:?}" |
| 3362 | ); |
| 3363 | } |
| 3364 | } |
| 3365 | |
| 3366 | #[test] |
| 3367 | fn subagent_done_sentinel_section_present() { |
| 3368 | assert!(AGENT_MODE.contains("completion events as internal evidence")); |
| 3369 | assert!(AGENT_MODE.contains("verify load-bearing child")); |
| 3370 | assert!(AGENT_MODE.contains("never manufacture completion sentinels")); |
| 3371 | assert!(!AGENT_MODE.contains("<codewhale:subagent.done>")); |
| 3372 | } |
| 3373 | |
| 3374 | #[test] |
| 3375 | fn preamble_carries_tone_and_ownership_guidance() { |
| 3376 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3377 | assert!(prompt.contains("The A is already yours")); |
| 3378 | assert!(prompt.contains("Your competence is a settled fact")); |
| 3379 | assert!(prompt.contains("Take the work seriously. Don't take")); |
| 3380 | assert!(prompt.contains("Let the work speak")); |
| 3381 | } |
| 3382 | |
| 3383 | // ── Cache-prefix stability harness (#263 step 2) ─────────────────────── |
| 3384 | // |
| 3385 | // These tests pin the byte-stability invariant required for DeepSeek's |
| 3386 | // KV prefix cache to hit: any prompt-construction surface that ends up |
| 3387 | // in the cached prefix must produce identical bytes given identical |
| 3388 | // inputs across calls. |
| 3389 | |
| 3390 | use crate::test_support::{EnvVarGuard, assert_byte_identical}; |
| 3391 | |
| 3392 | #[test] |
| 3393 | fn compose_prompt_is_byte_stable_across_calls() { |
| 3394 | // Suspect #4 from #263: mode prompt churn within a single mode. |
| 3395 | // Two calls with identical (mode, personality) inputs must produce |
| 3396 | // identical bytes — anything else is a cache buster. |
| 3397 | let a = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3398 | let b = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3399 | assert_byte_identical("compose_prompt(Personality::Calm)", &a, &b); |
| 3400 | } |
| 3401 | |
| 3402 | #[test] |
| 3403 | fn system_prompt_for_mode_with_context_is_byte_stable_for_unchanged_workspace() { |
| 3404 | // Same workspace, no working_set / skills churn between calls → |
| 3405 | // identical bytes. This pins the most representative production |
| 3406 | // surface (engine.rs builds the system prompt via this fn or |
| 3407 | // its sibling _and_skills variant on every turn). |
| 3408 | let _env_guard = crate::test_support::lock_test_env(); |
| 3409 | let workspace_tmp = tempdir().expect("workspace tempdir"); |
| 3410 | let home_tmp = tempdir().expect("home tempdir"); |
| 3411 | let _home = EnvVarGuard::set("HOME", home_tmp.path().as_os_str()); |
| 3412 | let _userprofile = EnvVarGuard::set("USERPROFILE", home_tmp.path().as_os_str()); |
| 3413 | let _skills_dir = EnvVarGuard::remove("DEEPSEEK_SKILLS_DIR"); |
| 3414 | let workspace = workspace_tmp.path(); |
| 3415 | |
| 3416 | let a = system_prompt_flat_text(&system_prompt_for_mode_with_context(workspace, None)); |
| 3417 | let b = system_prompt_flat_text(&system_prompt_for_mode_with_context(workspace, None)); |
| 3418 | assert_byte_identical( |
| 3419 | "system_prompt_for_mode_with_context() on empty workspace", |
| 3420 | &a, |
| 3421 | &b, |
| 3422 | ); |
| 3423 | } |
| 3424 | |
| 3425 | #[test] |
| 3426 | fn system_prompt_ignores_working_set_summary_argument() { |
| 3427 | // Working-set metadata is now injected into the latest user message |
| 3428 | // per turn. The legacy argument remains for call-site compatibility |
| 3429 | // but must not reintroduce volatile bytes into the system prompt. |
| 3430 | let _env_guard = crate::test_support::lock_test_env(); |
| 3431 | let tmp = tempdir().expect("tempdir"); |
| 3432 | let home_tmp = tempdir().expect("home tempdir"); |
| 3433 | let _home = EnvVarGuard::set("HOME", home_tmp.path().as_os_str()); |
| 3434 | let _userprofile = EnvVarGuard::set("USERPROFILE", home_tmp.path().as_os_str()); |
| 3435 | let _skills_dir = EnvVarGuard::remove("DEEPSEEK_SKILLS_DIR"); |
| 3436 | let workspace = tmp.path(); |
| 3437 | let summary = "## Repo Working Set\nWorkspace: /tmp/x\n"; |
| 3438 | |
| 3439 | let a = system_prompt_flat_text(&system_prompt_for_mode_with_context( |
| 3440 | workspace, |
| 3441 | Some(summary), |
| 3442 | )); |
| 3443 | let b = system_prompt_flat_text(&system_prompt_for_mode_with_context( |
| 3444 | workspace, |
| 3445 | Some(summary), |
| 3446 | )); |
| 3447 | assert_byte_identical( |
| 3448 | "system_prompt_for_mode_with_context with constant working_set summary", |
| 3449 | &a, |
| 3450 | &b, |
| 3451 | ); |
| 3452 | assert!( |
| 3453 | !a.contains(summary), |
| 3454 | "summary must not be embedded in system prompt" |
| 3455 | ); |
| 3456 | } |
| 3457 | |
| 3458 | #[test] |
| 3459 | fn system_prompt_with_handoff_file_is_byte_stable_when_file_is_unchanged() { |
| 3460 | // If `.deepseek/handoff.md` hasn't moved between two builds, the |
| 3461 | // rendered prompt must produce identical bytes. The relay block |
| 3462 | // lands below the static boundary in |
| 3463 | // `system_prompt_for_mode_with_context_and_skills`. |
| 3464 | let _env_guard = crate::test_support::lock_test_env(); |
| 3465 | let tmp = tempdir().expect("tempdir"); |
| 3466 | let home_tmp = tempdir().expect("home tempdir"); |
| 3467 | let _home = EnvVarGuard::set("HOME", home_tmp.path().as_os_str()); |
| 3468 | let _userprofile = EnvVarGuard::set("USERPROFILE", home_tmp.path().as_os_str()); |
| 3469 | let _skills_dir = EnvVarGuard::remove("DEEPSEEK_SKILLS_DIR"); |
| 3470 | let workspace = tmp.path(); |
| 3471 | let handoff_dir = workspace.join(".deepseek"); |
| 3472 | std::fs::create_dir_all(&handoff_dir).unwrap(); |
| 3473 | std::fs::write( |
| 3474 | handoff_dir.join("handoff.md"), |
| 3475 | "# Session relay\n\n## Active task\nFinish #280.\n\n## Open blockers\n- [ ] none\n", |
| 3476 | ) |
| 3477 | .unwrap(); |
| 3478 | |
| 3479 | let a = system_prompt_flat_text(&system_prompt_for_mode_with_context(workspace, None)); |
| 3480 | let b = system_prompt_flat_text(&system_prompt_for_mode_with_context(workspace, None)); |
| 3481 | assert_byte_identical( |
| 3482 | "system_prompt_for_mode_with_context with constant handoff file", |
| 3483 | &a, |
| 3484 | &b, |
| 3485 | ); |
| 3486 | assert!(a.contains(HANDOFF_BLOCK_MARKER), "relay must be embedded"); |
| 3487 | assert!(a.contains("Finish #280."), "relay body must be present"); |
| 3488 | } |
| 3489 | |
| 3490 | #[test] |
| 3491 | fn handoff_appears_after_static_blocks_without_working_set() { |
| 3492 | // Cache-prefix invariant: the relay artifact must come after static |
| 3493 | // `## Core Execution`. The relay template itself is now action-local, |
| 3494 | // not part of every system prompt. Working-set metadata is per-turn |
| 3495 | // user metadata, not a system-prompt tail block. |
| 3496 | let tmp = tempdir().expect("tempdir"); |
| 3497 | let workspace = tmp.path(); |
| 3498 | let handoff_dir = workspace.join(".deepseek"); |
| 3499 | std::fs::create_dir_all(&handoff_dir).unwrap(); |
| 3500 | std::fs::write(handoff_dir.join("handoff.md"), "# handoff body\n").unwrap(); |
| 3501 | |
| 3502 | let summary = "## Repo Working Set\nWorkspace: /tmp/x\n"; |
| 3503 | let prompt = system_prompt_flat_text(&system_prompt_for_mode_with_context( |
| 3504 | workspace, |
| 3505 | Some(summary), |
| 3506 | )); |
| 3507 | |
| 3508 | let execution_pos = prompt |
| 3509 | .find("## Core Execution") |
| 3510 | .expect("Core Execution section present in Agent mode"); |
| 3511 | let handoff_pos = prompt |
| 3512 | .find(HANDOFF_BLOCK_MARKER) |
| 3513 | .expect("relay block present when fixture file exists"); |
| 3514 | assert!( |
| 3515 | !prompt.contains("## Repo Working Set"), |
| 3516 | "working-set summary must stay out of the system prompt" |
| 3517 | ); |
| 3518 | |
| 3519 | assert!( |
| 3520 | execution_pos < handoff_pos, |
| 3521 | "## Core Execution must precede the relay block" |
| 3522 | ); |
| 3523 | assert!(!prompt.contains("# Session relay")); |
| 3524 | } |
| 3525 | |
| 3526 | #[test] |
| 3527 | fn render_instructions_block_returns_none_for_empty_input() { |
| 3528 | let empty: &[super::InstructionSource] = &[]; |
| 3529 | assert!(super::render_instructions_block(empty).is_none()); |
| 3530 | } |
| 3531 | |
| 3532 | /// #4632 — The system prompt prefix (the byte-stable part cached by |
| 3533 | /// inference servers) must never contain private content: absolute |
| 3534 | /// filesystem paths, API keys, or home-directory references. |
| 3535 | #[test] |
| 3536 | fn system_prompt_prefix_never_leaks_private_content() { |
| 3537 | let tmp = tempdir().expect("tempdir"); |
| 3538 | let workspace = tmp.path(); |
| 3539 | let prompt = match system_prompt_for_mode_with_context(workspace, None) { |
| 3540 | SystemPrompt::Text(text) => text, |
| 3541 | SystemPrompt::Blocks(blocks) => blocks |
| 3542 | .iter() |
| 3543 | .map(|block| block.text.as_str()) |
| 3544 | .collect::<Vec<_>>() |
| 3545 | .join("\n"), |
| 3546 | }; |
| 3547 | |
| 3548 | // No absolute paths (Unix or Windows). |
| 3549 | let offending: Vec<&str> = prompt |
| 3550 | .lines() |
| 3551 | .filter(|line| { |
| 3552 | line.contains("/Users/") || line.contains("/home/") || line.contains("C:\\") |
| 3553 | }) |
| 3554 | .collect(); |
| 3555 | assert!( |
| 3556 | offending.is_empty(), |
| 3557 | "system prompt must not contain absolute user paths, found: {offending:?}" |
| 3558 | ); |
| 3559 | // No API key patterns. |
| 3560 | assert!( |
| 3561 | !prompt.contains("sk-") && !prompt.contains("api_key") && !prompt.contains("API_KEY"), |
| 3562 | "system prompt must not contain API key material" |
| 3563 | ); |
| 3564 | // The workspace path itself must not appear. |
| 3565 | assert!( |
| 3566 | !prompt.contains(workspace.to_str().unwrap_or("/nonexistent")), |
| 3567 | "system prompt must not embed the workspace path" |
| 3568 | ); |
| 3569 | } |
| 3570 | |
| 3571 | #[test] |
| 3572 | fn render_instructions_block_skips_missing_files_with_warning() { |
| 3573 | let tmp = tempdir().expect("tempdir"); |
| 3574 | let real = tmp.path().join("real.md"); |
| 3575 | std::fs::write(&real, "real content here").unwrap(); |
| 3576 | let bogus = tmp.path().join("does-not-exist.md"); |
| 3577 | |
| 3578 | let block = super::render_instructions_block(&[bogus.clone().into(), real.clone().into()]) |
| 3579 | .expect("present file should produce a block"); |
| 3580 | assert!(block.contains("real content here")); |
| 3581 | assert!(block.contains(&real.display().to_string())); |
| 3582 | // Bogus path is skipped, not rendered. |
| 3583 | assert!(!block.contains(&bogus.display().to_string())); |
| 3584 | } |
| 3585 | |
| 3586 | #[test] |
| 3587 | fn render_instructions_block_concatenates_in_declared_order() { |
| 3588 | let tmp = tempdir().expect("tempdir"); |
| 3589 | let a = tmp.path().join("a.md"); |
| 3590 | let b = tmp.path().join("b.md"); |
| 3591 | std::fs::write(&a, "ALPHA_MARKER").unwrap(); |
| 3592 | std::fs::write(&b, "BRAVO_MARKER").unwrap(); |
| 3593 | |
| 3594 | let block = super::render_instructions_block(&[a.into(), b.into()]).expect("non-empty"); |
| 3595 | let alpha_pos = block.find("ALPHA_MARKER").expect("alpha rendered"); |
| 3596 | let bravo_pos = block.find("BRAVO_MARKER").expect("bravo rendered"); |
| 3597 | assert!( |
| 3598 | alpha_pos < bravo_pos, |
| 3599 | "instructions must concatenate in declared order" |
| 3600 | ); |
| 3601 | } |
| 3602 | |
| 3603 | #[test] |
| 3604 | fn render_instructions_block_skips_empty_files() { |
| 3605 | let tmp = tempdir().expect("tempdir"); |
| 3606 | let empty = tmp.path().join("empty.md"); |
| 3607 | let real = tmp.path().join("real.md"); |
| 3608 | std::fs::write(&empty, " \n \n").unwrap(); |
| 3609 | std::fs::write(&real, "real content").unwrap(); |
| 3610 | |
| 3611 | let block = |
| 3612 | super::render_instructions_block(&[empty.into(), real.into()]).expect("non-empty"); |
| 3613 | // Empty file produces no `<instructions>` section, only the real one. |
| 3614 | let count = block.matches("<instructions").count(); |
| 3615 | assert_eq!(count, 1, "only the non-empty file should produce a section"); |
| 3616 | } |
| 3617 | |
| 3618 | #[test] |
| 3619 | fn render_instructions_block_truncates_oversize_files() { |
| 3620 | let tmp = tempdir().expect("tempdir"); |
| 3621 | let big = tmp.path().join("big.md"); |
| 3622 | // 200 KiB of content — well above the 100 KiB cap. |
| 3623 | std::fs::write(&big, "X".repeat(200 * 1024)).unwrap(); |
| 3624 | |
| 3625 | let block = super::render_instructions_block(&[big.into()]).expect("non-empty"); |
| 3626 | assert!(block.contains("[…truncated:"), "truncation marker missing"); |
| 3627 | // Block should be much smaller than the original file. |
| 3628 | assert!( |
| 3629 | block.len() < 110 * 1024, |
| 3630 | "block should be capped near 100 KiB" |
| 3631 | ); |
| 3632 | } |
| 3633 | |
| 3634 | /// `InstructionSource::Inline` bypasses disk reads — the content is used |
| 3635 | /// directly and `name` becomes the `<instructions source="…">` attribute. |
| 3636 | /// Empty / oversize handling mirrors `File` variant. |
| 3637 | #[test] |
| 3638 | fn render_instructions_block_handles_inline_source() { |
| 3639 | let block = super::render_instructions_block(&[super::InstructionSource::Inline { |
| 3640 | name: "embedded:test/template".to_string(), |
| 3641 | content: "INLINE_MARKER_CONTENT".to_string(), |
| 3642 | }]) |
| 3643 | .expect("non-empty"); |
| 3644 | assert!(block.contains("INLINE_MARKER_CONTENT")); |
| 3645 | assert!(block.contains("source=\"embedded:test/template\"")); |
| 3646 | |
| 3647 | // Empty inline → skipped just like empty file. |
| 3648 | let empty_inline = super::InstructionSource::Inline { |
| 3649 | name: "empty".to_string(), |
| 3650 | content: " ".to_string(), |
| 3651 | }; |
| 3652 | assert!(super::render_instructions_block(&[empty_inline]).is_none()); |
| 3653 | |
| 3654 | // Oversize inline → truncated with elided marker. |
| 3655 | let big_inline = super::InstructionSource::Inline { |
| 3656 | name: "huge".to_string(), |
| 3657 | content: "Y".repeat(200 * 1024), |
| 3658 | }; |
| 3659 | let trimmed = super::render_instructions_block(&[big_inline]).expect("non-empty"); |
| 3660 | assert!(trimmed.contains("[…truncated:")); |
| 3661 | |
| 3662 | // File + Inline 混用,顺序保持。 |
| 3663 | let tmp = tempdir().expect("tempdir"); |
| 3664 | let file_path = tmp.path().join("file-first.md"); |
| 3665 | std::fs::write(&file_path, "FILE_MARKER").unwrap(); |
| 3666 | let mixed = super::render_instructions_block(&[ |
| 3667 | file_path.into(), |
| 3668 | super::InstructionSource::Inline { |
| 3669 | name: "inline-second".to_string(), |
| 3670 | content: "INLINE_MARKER".to_string(), |
| 3671 | }, |
| 3672 | ]) |
| 3673 | .expect("non-empty"); |
| 3674 | let file_pos = mixed.find("FILE_MARKER").expect("file rendered"); |
| 3675 | let inline_pos = mixed.find("INLINE_MARKER").expect("inline rendered"); |
| 3676 | assert!(file_pos < inline_pos, "声明顺序必须保留(File then Inline)"); |
| 3677 | } |
| 3678 | |
| 3679 | #[test] |
| 3680 | fn instructions_block_appears_in_system_prompt_when_configured() { |
| 3681 | let tmp = tempdir().expect("tempdir"); |
| 3682 | let workspace = tmp.path(); |
| 3683 | let extra = workspace.join("extra-instructions.md"); |
| 3684 | std::fs::write(&extra, "EXTRA_INSTRUCTIONS_MARKER_BODY").unwrap(); |
| 3685 | |
| 3686 | let extra_source: super::InstructionSource = extra.clone().into(); |
| 3687 | let prompt = |
| 3688 | system_prompt_flat_text(&super::system_prompt_for_mode_with_context_and_skills( |
| 3689 | workspace, |
| 3690 | None, |
| 3691 | None, |
| 3692 | Some(std::slice::from_ref(&extra_source)), |
| 3693 | None, |
| 3694 | )); |
| 3695 | |
| 3696 | assert!( |
| 3697 | prompt.contains("EXTRA_INSTRUCTIONS_MARKER_BODY"), |
| 3698 | "configured instructions file body must appear in the prompt" |
| 3699 | ); |
| 3700 | assert!( |
| 3701 | prompt.contains(&extra.display().to_string()), |
| 3702 | "instructions block must annotate its source path" |
| 3703 | ); |
| 3704 | } |
| 3705 | |
| 3706 | #[test] |
| 3707 | fn verbosity_concise_appends_discipline_block() { |
| 3708 | let tmp = tempdir().expect("tempdir"); |
| 3709 | let workspace = tmp.path(); |
| 3710 | let prompt = system_prompt_flat_text( |
| 3711 | &super::system_prompt_for_mode_with_context_skills_session_and_approval( |
| 3712 | workspace, |
| 3713 | None, |
| 3714 | None, |
| 3715 | None, |
| 3716 | PromptSessionContext { |
| 3717 | user_memory_block: None, |
| 3718 | goal_objective: None, |
| 3719 | project_context_pack_enabled: false, |
| 3720 | locale_tag: "en", |
| 3721 | translation_enabled: false, |
| 3722 | model_id: "codewhale", |
| 3723 | context_window_override: None, |
| 3724 | verbosity: Some(" Concise "), |
| 3725 | skills_scan_codewhale_only: false, |
| 3726 | plugin_registry: None, |
| 3727 | mode: crate::tui::app::AppMode::Agent, |
| 3728 | }, |
| 3729 | ), |
| 3730 | ); |
| 3731 | |
| 3732 | assert!( |
| 3733 | prompt.contains("## Concise Output Discipline"), |
| 3734 | "Concise Output Discipline should be appended" |
| 3735 | ); |
| 3736 | } |
| 3737 | |
| 3738 | /// #2953 — the Calm overlay (`CALM_PERSONALITY`) stays out of the default |
| 3739 | /// model-prompt path to keep the static prefix slim. Voice and tone |
| 3740 | /// guidance travels via the constitution preamble instead. |
| 3741 | #[test] |
| 3742 | fn default_prompt_does_not_include_calm_personality_overlay() { |
| 3743 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3744 | let calm_text = CALM_PERSONALITY; |
| 3745 | let first_calm_line = calm_text.lines().find(|l| !l.is_empty()).unwrap_or(""); |
| 3746 | assert!( |
| 3747 | !prompt.contains(first_calm_line), |
| 3748 | "default agent prompt must not include the calm personality overlay" |
| 3749 | ); |
| 3750 | } |
| 3751 | |
| 3752 | #[test] |
| 3753 | fn live_prompt_path_returns_world_state_blocks_with_markers() { |
| 3754 | let tmp = tempdir().expect("tempdir"); |
| 3755 | let prompt = system_prompt_for_mode_with_context_skills_session_and_approval( |
| 3756 | tmp.path(), |
| 3757 | None, |
| 3758 | None, |
| 3759 | None, |
| 3760 | PromptSessionContext { |
| 3761 | user_memory_block: Some("## Memory\n- remember the cutover"), |
| 3762 | goal_objective: Some("ship WorldState Blocks"), |
| 3763 | project_context_pack_enabled: false, |
| 3764 | locale_tag: "en", |
| 3765 | translation_enabled: false, |
| 3766 | model_id: "deepseek-v4-pro", |
| 3767 | context_window_override: None, |
| 3768 | verbosity: Some("concise"), |
| 3769 | skills_scan_codewhale_only: false, |
| 3770 | plugin_registry: None, |
| 3771 | mode: crate::tui::app::AppMode::Agent, |
| 3772 | }, |
| 3773 | ); |
| 3774 | |
| 3775 | let SystemPrompt::Blocks(blocks) = prompt else { |
| 3776 | panic!("live prompt assembly must return SystemPrompt::Blocks"); |
| 3777 | }; |
| 3778 | assert!( |
| 3779 | blocks.len() >= 3, |
| 3780 | "constitution + at least one WorldState fragment + authority trailer" |
| 3781 | ); |
| 3782 | assert!( |
| 3783 | !blocks[0].text.contains("<!-- cw:ctx:"), |
| 3784 | "constitution block must stay marker-free for prefix cache stability" |
| 3785 | ); |
| 3786 | assert!( |
| 3787 | blocks[0].text.contains("## Core Execution"), |
| 3788 | "constitution retains static core execution guidance" |
| 3789 | ); |
| 3790 | |
| 3791 | let flat = system_prompt_flat_text(&SystemPrompt::Blocks(blocks.clone())); |
| 3792 | assert!(flat.contains(crate::model_context::FragmentId::Workspace.marker())); |
| 3793 | assert!(flat.contains(crate::model_context::FragmentId::Route.marker())); |
| 3794 | assert!(flat.contains("## Environment")); |
| 3795 | assert!( |
| 3796 | flat.contains("verbosity: concise") && flat.contains("translation: off"), |
| 3797 | "route fragment keeps verbosity/translation but drops the model id" |
| 3798 | ); |
| 3799 | assert!(!flat.contains("model: deepseek-v4-pro")); |
| 3800 | assert!(flat.contains("<session_goal>")); |
| 3801 | assert!(flat.contains("ship WorldState Blocks")); |
| 3802 | assert!(flat.contains("remember the cutover")); |
| 3803 | assert!(flat.contains("## Authority Recap")); |
| 3804 | assert!( |
| 3805 | !flat.contains(crate::model_context::FragmentId::SkillsTools.marker()), |
| 3806 | "skills remain in constitution, not a volatile SkillsTools fragment" |
| 3807 | ); |
| 3808 | } |
| 3809 | |
| 3810 | #[test] |
| 3811 | fn live_prompt_world_state_diff_retains_unchanged_fragments() { |
| 3812 | let tmp = tempdir().expect("tempdir"); |
| 3813 | let session = PromptSessionContext { |
| 3814 | user_memory_block: None, |
| 3815 | goal_objective: None, |
| 3816 | project_context_pack_enabled: false, |
| 3817 | locale_tag: "en", |
| 3818 | translation_enabled: false, |
| 3819 | model_id: "codewhale", |
| 3820 | context_window_override: None, |
| 3821 | verbosity: None, |
| 3822 | skills_scan_codewhale_only: false, |
| 3823 | plugin_registry: None, |
| 3824 | mode: crate::tui::app::AppMode::Agent, |
| 3825 | }; |
| 3826 | let first = system_prompt_for_mode_with_context_skills_session_and_approval( |
| 3827 | tmp.path(), |
| 3828 | None, |
| 3829 | None, |
| 3830 | None, |
| 3831 | session.clone(), |
| 3832 | ); |
| 3833 | let second = system_prompt_for_mode_with_context_skills_session_and_approval( |
| 3834 | tmp.path(), |
| 3835 | None, |
| 3836 | None, |
| 3837 | None, |
| 3838 | PromptSessionContext { |
| 3839 | goal_objective: Some("only goal changed"), |
| 3840 | ..session |
| 3841 | }, |
| 3842 | ); |
| 3843 | |
| 3844 | let extract_world = |prompt: &SystemPrompt| -> crate::model_context::WorldState { |
| 3845 | let SystemPrompt::Blocks(blocks) = prompt else { |
| 3846 | panic!("expected Blocks"); |
| 3847 | }; |
| 3848 | let mut state = crate::model_context::WorldState::new(); |
| 3849 | for block in blocks.iter().skip(1) { |
| 3850 | for id in crate::model_context::FragmentId::all() { |
| 3851 | let marker = id.marker(); |
| 3852 | if let Some(rest) = block.text.strip_prefix(marker) { |
| 3853 | let body = rest.trim_start_matches('\n'); |
| 3854 | state.upsert(crate::model_context::ModelContextFragment::new( |
| 3855 | *id, |
| 3856 | id.role(), |
| 3857 | body, |
| 3858 | )); |
| 3859 | } |
| 3860 | } |
| 3861 | } |
| 3862 | state |
| 3863 | }; |
| 3864 | |
| 3865 | let previous = extract_world(&first); |
| 3866 | let next = extract_world(&second); |
| 3867 | let diff = next.render_diff(Some(&previous)); |
| 3868 | assert!( |
| 3869 | diff.retained |
| 3870 | .iter() |
| 3871 | .any(|marker| marker == crate::model_context::FragmentId::Route.marker()), |
| 3872 | "unchanged route fragment must be retained: {diff:?}" |
| 3873 | ); |
| 3874 | assert!( |
| 3875 | diff.updated |
| 3876 | .iter() |
| 3877 | .any(|fragment| fragment.id == crate::model_context::FragmentId::Workspace), |
| 3878 | "goal change must update workspace fragment: {diff:?}" |
| 3879 | ); |
| 3880 | } |
| 3881 | |
| 3882 | #[test] |
| 3883 | fn default_prompt_stays_under_2953_static_baseline() { |
| 3884 | const ISSUE_2953_BASELINE_CHARS: usize = 30_461; |
| 3885 | let prompt = compose_prompt_with_approval_model_and_shell(Personality::Calm, "codewhale"); |
| 3886 | |
| 3887 | assert!( |
| 3888 | prompt.chars().count() < ISSUE_2953_BASELINE_CHARS, |
| 3889 | "default static prompt should stay below the #2953 baseline" |
| 3890 | ); |
| 3891 | } |
| 3892 | } |
| 3893 | #[test] |
| 3894 | fn core_execution_profile_is_runtime_only() { |
| 3895 | for required in [ |
| 3896 | "repository instructions", |
| 3897 | "inspect the narrow owner", |
| 3898 | "verify it", |
| 3899 | "Report changed files", |
| 3900 | ] { |
| 3901 | assert!(CORE_EXECUTION_PROFILE_PROMPT.contains(required)); |
| 3902 | } |
| 3903 | for forbidden in [ |
| 3904 | "footer", |
| 3905 | "color", |
| 3906 | "hotbar", |
| 3907 | "panel", |
| 3908 | "Fleet", |
| 3909 | "Workflow", |
| 3910 | "OpenHands", |
| 3911 | ] { |
| 3912 | assert!(!CORE_EXECUTION_PROFILE_PROMPT.contains(forbidden)); |
| 3913 | } |
| 3914 | } |
| 3915 |