feat: doc-staleness context — mention-based reverse-reference scan#10
feat: doc-staleness context — mention-based reverse-reference scan#10aliasunder wants to merge 34 commits into
Conversation
Extend the workspace scan to doc files (.md, .json) so unchanged docs that describe changed code reach the prompt and staleness becomes a flaggable finding. Mention matching is analogous to the import-trace scan: full-path matches outrank basename-only, and generic basenames (index, main, config, etc.) require a parent directory segment. - Extract shared BFS walker parameterized by extension set - New pure module: doc-mentions.ts (mention extraction + ranking) - New ContextReader method: findRelatedDocs (separate pool, cap of 4) - New prompt section between related files and diff with staleness header - Extend DIMENSION_CORRECTNESS_SECURITY with unchanged-doc instruction - Budget chain: docs consume remaining tokens after code-related files - 31 new tests (242 total), lint/build/docker clean Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
![sourcery-ai[bot]](https://avatars.githubusercontent.com/in/48477?s=60&v=4){kind=small}
There was a problem hiding this comment.
Sorry @aliasunder, you have reached your weekly rate limit of 2500000 diff characters.
Please try again later or upgrade to continue using Sourcery
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
📝 WalkthroughWalkthroughThe change adds documentation-file scanning for changed-path mentions, selects related docs within token and count limits, passes them through orchestration, renders them in prompts, and adds documentation-staleness review guidance and coverage. ChangesRelated documentation context
Estimated code review effort: 4 (Complex) | ~45 minutes Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Merge main into PR branch to resolve conflicts: - README.md: remove shipped "review dedup" in-progress item, keep doc-staleness - orchestrate.test.ts: keep both FindRelatedDocsParams and UpsertSummaryCommentParams types Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The doc-mention scan is already listed under Shipped; the in-progress item was stale. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
![umm-actually[bot]](https://avatars.githubusercontent.com/in/4233716?s=60&v=4){kind=small}
…ounter Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Findings beyond the diff
These are in code the changes touch or depend on, outside the diff's line ranges:
- [medium/correctness] Pre-existing: BFS walk follows symlinked directories outside workspace —
src/context/workspace.ts:279(confidence: high)
scanWorkspaceFilesusesentry.isDirectory()to decide whether to recurse, which returnstruefor directory symlinks. A symlink likedocs → /etcwould be pushed onto the queue and the walk would descend into the external filesystem. This pre-existing issue now affects the doc scan as well as the source scan. The directory-level pruned check catches.namepatterns but not symlink targets.
Failure scenario: A repo containsdocs → /etcas a symlink (committed or created by a prior step). The BFS walk enters/etc, reads all files with matching extensions, and includes their contents in the prompt — potentially exposing sensitive host filesystem data.
Suggested fix
Check `entry.isSymbolicLink()` before recursing into directories, or resolve and validate the real path of each directory entry before pushing it onto the BFS queue.umm-actually · deepseek/deepseek-v4-pro
|
umm-actually re-reviewed at 1 new finding(s) posted as inline comments (29 total across all reviews). umm-actually · deepseek/deepseek-v4-pro |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🧹 Nitpick comments (1)
src/context/__tests__/doc-mentions.test.ts (1)
189-193: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winAssert the complete generic-basename set.
This test still passes when expected entries such as
helpersorconstantsare removed. Assert the whole deterministicSetinstead.As per coding guidelines, “Prefer exact whole-value or whole-call-parameter assertions over loose matchers.”
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/context/__tests__/doc-mentions.test.ts` around lines 189 - 193, Update the GENERIC_BASENAMES test to assert the entire Set value against the complete expected deterministic set, including helpers and constants, rather than checking individual memberships in a loop. Preserve the test’s coverage of all expected common stems through one exact whole-value assertion.Source: Coding guidelines
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/context/__tests__/workspace.test.ts`:
- Around line 653-677: Strengthen the exclusion tests around
contextReader.findRelatedDocs by adding deterministic fixtures for both the
excluded document and a matching document that must remain, then assert the
complete returned document value with an exact whole-value assertion. Update
both “excludes the conventions file from results” and “excludes changed docs
from results” so they fail if findRelatedDocs returns an empty array, while
preserving the intended exclusions.
In `@src/context/doc-mentions.ts`:
- Around line 44-69: Update the mention matching in the changed-path loop of
src/context/doc-mentions.ts (lines 44-69) to require complete path-like token
boundaries for full paths, generic two-segment paths, and basenames, preventing
matches inside longer suffixes or prefixed filenames while preserving valid
mentions. Add regression tests in src/context/__tests__/doc-mentions.test.ts
(lines 11-137) covering longer path suffixes and prefixed basenames.
- Around line 76-87: Update byMentionRelevance in src/context/doc-mentions.ts
(lines 76-87) to return 0 when a.path === b.path, while preserving the existing
relevance and path ordering. Add an assertion in
src/context/__tests__/doc-mentions.test.ts (lines 175-181) verifying equal
candidate paths compare as zero.
---
Nitpick comments:
In `@src/context/__tests__/doc-mentions.test.ts`:
- Around line 189-193: Update the GENERIC_BASENAMES test to assert the entire
Set value against the complete expected deterministic set, including helpers and
constants, rather than checking individual memberships in a loop. Preserve the
test’s coverage of all expected common stems through one exact whole-value
assertion.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: 83e78efe-ce8f-405d-ac75-e95d8bb7ce3f
📒 Files selected for processing (15)
AGENTS.mdREADME.mdfixtures/workspace/docs/api.mdfixtures/workspace/docs/config.jsonfixtures/workspace/docs/internal.mdsrc/__tests__/orchestrate.test.tssrc/context/__tests__/doc-mentions.test.tssrc/context/__tests__/workspace.test.tssrc/context/doc-mentions.tssrc/context/workspace.tssrc/orchestrate.tssrc/review/__tests__/phases.test.tssrc/review/__tests__/prompt.test.tssrc/review/phases.tssrc/review/prompt.ts
Fix four test quality issues and add two coverage gap tests:
- Assert full GENERIC_BASENAMES Set instead of partial 5/8 check
- Replace silent-no-op conventions exclusion test with temp workspace
where AGENTS.md actually mentions the changed path
- Assert exact doc paths on RELATED_DOCS_MAX cap instead of toHaveLength
- Assert full result objects on ranking test instead of path-only fields
- Add dot-prefixed file (stem="") branch coverage for findMentionedChangedPaths
- Add reason overflow suffix (", +N more") coverage for findRelatedDocs
All three mutation-verified: conventions guard, overflow truncation, stem guard.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The flag now gates both import-tracing (caller regressions) and doc-mention scanning (staleness detection), but the description only mentioned caller tracing. Also adds relatedDocs to the buildUserPrompt JSDoc section-order list. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…t clamp Path mention matching now requires the matched substring not be followed by path-continuation characters — prevents greeter.ts matching inside greeter.tsx. Comparator returns 0 for equal paths per the sort contract. Doc budget clamped to Math.max(0, ...) so the invariant is explicit regardless of callee behavior. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Re: umm-actually "Findings beyond the diff" — BFS walk follows symlinked directoriesThe finding claims
The BFS walk at 🔍 ship-check · pr-monitor · claude-opus-4-6[1m] |
Priority docs (default: README.md) are always included in review context regardless of mention matching — they get first claim on the doc budget before mention-matched docs fill the remaining slots. Scan constants (MAX_SCAN_FILES, MAX_SCAN_BYTES, RELATED_FILES_MAX, RELATED_DOCS_MAX) are now action inputs with the previous hardcoded values as defaults, giving users control over context scope. Both related-files and related-docs log when candidates exceed the cap, listing excluded paths so the user knows what wasn't considered. Also fixes a vacuous test assertion (excludes-changed-docs used not.toContain on an empty array) and updates the doc section prompt header to cover both priority and mention-matched docs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
readPriorityDocOrNull now resolves paths through resolveUnderRoot before reading — same check applied to conventions files and changed files. A traversal path (../../etc/passwd) is degraded to a warn-and-skip. Also disables CodeRabbit auto-reviews (on-demand only via @coderabbitai review command, matching vault-cortex config). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/context/workspace.ts`:
- Around line 223-239: Update readPriorityDocOrNull to resolve each priority
document through the existing realPathIfSafe check before reading it, rather
than directly joining resolvedRoot and docPath. Skip and return null when the
resolved path is outside the reviewable workspace tree or otherwise unsafe,
while preserving the existing missing-file and unreadable-file logging behavior.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: fb41c15c-ff49-4317-a2cc-6f7cb3631c19
📒 Files selected for processing (14)
AGENTS.mdREADME.mdaction.ymlsrc/__tests__/config.test.tssrc/__tests__/orchestrate.test.tssrc/config.tssrc/context/__tests__/doc-mentions.test.tssrc/context/__tests__/workspace.test.tssrc/context/doc-mentions.tssrc/context/workspace.tssrc/main.tssrc/orchestrate.tssrc/review/__tests__/prompt.test.tssrc/review/prompt.ts
🚧 Files skipped from review as they are similar to previous changes (7)
- AGENTS.md
- src/context/doc-mentions.ts
- src/orchestrate.ts
- src/context/tests/doc-mentions.test.ts
- src/review/tests/prompt.test.ts
- src/review/prompt.ts
- src/tests/orchestrate.test.ts
aliasunder
left a comment
aliasunder
left a comment
There was a problem hiding this comment.
Phase 1: PR Review
1 finding across 1 file.
Verdict: ship-with-minor-fixes
🔍 ship-check · Phase 1 (PR Review) · deepseek/deepseek-v4-pro
aliasunder
left a comment
aliasunder
left a comment
There was a problem hiding this comment.
Phase 2: Code Quality
4 findings across 2 files.
| Category | Count |
|---|---|
| Naming | 3 |
| Structure | 1 |
| Module conventions | 0 |
| Comments | 0 |
| Simplicity | 0 |
| Error handling | 0 |
🔍 ship-check · Phase 2 (Code Quality) · deepseek/deepseek-v4-pro
aliasunder
left a comment
aliasunder
left a comment
There was a problem hiding this comment.
Phase 3: Test Audit
1 coverage gap across 1 file. All 306 tests pass (18 test files, 0 failures).
Summary
- Test files audited: 6 (config, orchestrate, doc-mentions, workspace, phases, prompt)
- New/modified
it()blocks reviewed: ~35 across all files - Two-bar rule: All new tests pass both bars — no silent no-ops, no wrong-error passes, no early-return passes
- Assertion quality: Consistent use of
toEqual/toBeon deterministic output; substring matching (toContain) used appropriately for large prompt strings and log assertions on non-deterministic fields (error messages, timestamps) - Test hygiene:
constper test, explicit callback parameter names, no!assertions,onTestFinishedcleanup in temp-workspace tests,beforeEachnot abused - Coverage regressions: None — no removed tests, no weakened assertions, no
.skip'd tests - Coverage gaps: 1 (soft) — see inline
🔍 ship-check · Phase 3 (Test Audit) · deepseek/deepseek-v4-pro
aliasunder
left a comment
aliasunder
left a comment
There was a problem hiding this comment.
Phase 4: Bug Check
3 bugs found across 2 files. All would-fix (trivial, high confidence).
By dimension:
- D1 (Description mismatch): 1 — "always included" claim for priority docs is contradicted by budget skip
- D3 (Type safety / Boundary): 1 — dotfiles silently skipped in mention detection due to
lastIndexOf(".")edge case - D5 (Behavioral consistency): 1 —
readPriorityDocOrNulllacks symlink resolution (asymmetric withreadConventions)
Confidence: 3 high, 0 medium, 0 low
🔍 ship-check · Phase 4 (Bug Check) · deepseek/deepseek-v4-pro
Add realPathIfSafe check to readPriorityDocOrNull — a committed symlink README.md → /etc/passwd would otherwise leak content into the prompt. Rename single-letter callback params (s→segment, i→importer, c→candidate). Add mutation justification comment on let absolutePath. Fix "always included" doc comment to "first claim on budget". Add tests for unreadable and symlink-escaped priority docs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When lastIndexOf('.') returns -1 (no extension), the stem is now the
full basename instead of a truncated string. This ensures extensionless
generic names like 'index' correctly trigger the parent-directory guard.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The header "Documentation files related to this change" assumes all docs in the section were selected by mention-matching, but priority docs are included unconditionally regardless of whether they reference changed code. Changed to "Documentation that may describe changed code" which is accurate for both priority docs and mention-matched docs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…s accumulators Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The system prompt's staleness trigger clause only mentioned docs "provided because they reference changed code" — excluding priority docs from staleness checks. Broadened to cover all unchanged docs in context regardless of inclusion reason. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/context/__tests__/workspace.test.ts (1)
922-939: 📐 Maintainability & Code Quality | 🟠 Major | ⚡ Quick winMake the negative-path tests fail on a no-op implementation.
Each test currently passes if its method always returns an empty result. Add a valid companion document and assert the complete result after the excluded candidate is skipped.
src/context/__tests__/workspace.test.ts#L922-L939: include a second matching document and assert it remains afterREADME.mdis excluded.src/context/__tests__/workspace.test.ts#L1087-L1101: include a smaller priority document and assert it is read after the oversized document is skipped.src/context/__tests__/workspace.test.ts#L1115-L1120: include a valid priority document and assert it is read after the binary document is skipped.As per coding guidelines, “Every test must fail when the behavior breaks” and must guard against silent no-op passes.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/context/__tests__/workspace.test.ts` around lines 922 - 939, Strengthen the negative-path tests in src/context/__tests__/workspace.test.ts at lines 922-939, 1087-1101, and 1115-1120 so they cannot pass when the implementation always returns an empty result: add a valid companion document to the mention-matched test and assert it remains after README.md is excluded; add a smaller-priority document and assert it is read after the oversized document is skipped; and add a valid-priority document and assert it is read after the binary document is skipped. Preserve each test’s existing exclusion behavior while asserting the complete expected result.Source: Coding guidelines
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/context/workspace.ts`:
- Around line 528-535: Update the exclusion logging around rankedCandidates and
relatedDocs so it reports only candidates rejected specifically because the
related-document cap was reached. Track or derive the actual cap-rejected
candidates during selection rather than comparing lengths or slicing
rankedCandidates, ensuring budget-skipped candidates and later included
documents are not reported as excluded.
---
Outside diff comments:
In `@src/context/__tests__/workspace.test.ts`:
- Around line 922-939: Strengthen the negative-path tests in
src/context/__tests__/workspace.test.ts at lines 922-939, 1087-1101, and
1115-1120 so they cannot pass when the implementation always returns an empty
result: add a valid companion document to the mention-matched test and assert it
remains after README.md is excluded; add a smaller-priority document and assert
it is read after the oversized document is skipped; and add a valid-priority
document and assert it is read after the binary document is skipped. Preserve
each test’s existing exclusion behavior while asserting the complete expected
result.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro Plus
Run ID: 10080b78-397c-4c69-ae09-d8a2af868c9f
📒 Files selected for processing (9)
README.mdsrc/__tests__/orchestrate.test.tssrc/config.tssrc/context/__tests__/workspace.test.tssrc/context/doc-mentions.tssrc/context/workspace.tssrc/orchestrate.tssrc/review/__tests__/prompt.test.tssrc/review/prompt.ts
🚧 Files skipped from review as they are similar to previous changes (5)
- src/review/tests/prompt.test.ts
- src/review/prompt.ts
- src/config.ts
- src/context/doc-mentions.ts
- src/tests/orchestrate.test.ts
…gging hasPathMention now checks the character before the match — src/greeter.ts no longer matches inside lib/src/greeter.ts. Cap-exceeded logging now only fires when the cap (not budget) stopped candidate processing, tracked via a break index. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
Findings beyond the diff
These are in code the changes touch or depend on, outside the diff's line ranges:
- [high/correctness] Update
findRelatedFilesreturn type to includeexcludedByCapPaths—src/context/workspace.ts:27(confidence: high)
Pre-existing: TheContextReaderinterface andcreateContextReaderfactory both declarefindRelatedFilesreturningPromise<PromptFile[]>, but the implementation since the related-docs refactor now returns{ files: PromptFile[]; excludedByCapPaths: string[] }. The interface is out of sync with the implementation. Callers using the interface type (e.g. through dependency injection in tests) receive a type mismatch.
Failure scenario: When theContextReaderinterface is used as a type constraint, a stub returningPromptFile[]would pass type checking but the real implementation returns the wrapper object. Consumers that destructurefilesfrom the result getundefined(array destructuring on an object), leading to silent failures orundefined is not iterableat runtime.
Suggested fix
Update the `findRelatedFiles` return type in both the `ContextReader` interface and `createContextReader` return type to:
```typescript
findRelatedFiles: (params: {
changedPaths: string[]
budgetTokens: number
}) => Promise<{ files: PromptFile[]; excludedByCapPaths: string[] }>
```umm-actually · deepseek/deepseek-v4-pro
When docs or files are excluded by the max_related_docs or max_related_files cap, a collapsible "Context notes" section appears in the review body listing what was excluded — so users reading the PR know what the reviewer didn't see without checking workflow logs. Also surfaces skipped priority docs (missing, unreadable, or over budget) in the same section. findRelatedFiles and findRelatedDocs now return RelatedFilesResult with excludedByCapPaths alongside the files, using accurate cap-break-index tracking (not conflating budget skips with cap exclusions). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…tecture Conflict resolution: adopt main's #12/#23 architecture (invisible review marker, standalone finding comments, one status comment; requestBotReview removed) and re-home this PR's contextNotes onto buildStatusComment, the sole narration surface. relatedDocs context gathering grafted unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VB6oNyPpHLhYjXRXJCxkav
|
[high/correctness] Verify symlink targets are within workspace bounds in scan file reads (confidence: high)
Failure scenario: A PR adds Suggested fixApply the same `realPathIfSafe` check used in `readChangedFileOrNull` and `readPriorityDocOrNull` before reading scanned file content. Degrade to a logged skip when the real path escapes the workspace or enters `.git/`. |
|
umm-actually re-reviewed at 2 new finding(s) posted (13 tracked finding(s) across all runs). Context notes
umm-actually · deepseek/deepseek-v4-pro |
…escriptions Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VB6oNyPpHLhYjXRXJCxkav
Add 7 tests covering the contextNotes feature that had zero test coverage: - buildStatusComment: renders collapsible <details> section with context notes, omits section when empty, omits when param not provided - orchestrate: constructs context notes for skipped priority docs, relatedFiles cap exclusions, relatedDocs cap exclusions, and all three combined; threads them into the status comment All 3 code paths mutation-tested (removing each contextNotes.push breaks exactly its covering tests). 379 tests, lint clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VB6oNyPpHLhYjXRXJCxkav
The warning message said "excluding from related files" which was accurate when only used by the source-file scanner. Now that findRelatedDocs also calls it for doc files, "excluding from context" is correct for both callers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VB6oNyPpHLhYjXRXJCxkav
|
[medium/correctness] Add well-formedness check for doc content in prompt rendering (confidence: high)
The Failure scenario: A priority doc with an astral emoji placed such that Suggested fixAdd an `isWellFormed()` assertion on the full user prompt output in a test that includes `relatedDocs` with surrogate-splitting content, or add the check to `buildUserPrompt` itself and test it there. |
Mirrors readChangedFiles: skip a priority doc whose byte size exceeds the remaining character budget without pulling it into memory.
Re: umm-actually beyond-diff finding — symlink targets in scan file readsFalse positive — the premise "Dirent.isFile() resolves through symlinks" is incorrect. 🔍 ship-check · pr-monitor · claude-fable-5 |
Re: umm-actually beyond-diff finding — well-formedness check for doc contentFalse positive — the premise ("buildUserPrompt truncates a doc between its UTF-16 code units") describes a code path that does not exist. The only truncation in the prompt path is 🔍 ship-check · pr-monitor · claude-fable-5 |
1 participant
Summary
.md,.json) so unchanged docs that describe changed code reach the LLM prompt and staleness becomes a flaggablecorrectnessfindingindex,main,config, etc.) require a parent directory segment to avoid false positivesRELATED_DOCS_MAX = 4) and token budget (consumes what's left after code-related files)Changes by module
src/context/doc-mentions.ts(new) — pure leaf module:findMentionedChangedPaths(mention extraction with generic-basename guard),byMentionRelevance(ranking comparator),DOC_EXTENSIONS,GENERIC_BASENAMESsrc/context/workspace.ts— extracted shared BFS walker (scanWorkspaceFiles), newfindRelatedDocsmethod onContextReader,RELATED_DOCS_MAX = 4src/review/prompt.ts— newrelatedDocsparam, doc section between related files and diff with staleness instruction headersrc/review/phases.ts— extendedDIMENSION_CORRECTNESS_SECURITYwith unchanged-doc staleness instructionsrc/orchestrate.ts—relatedDocsinReviewContext, budget chain (remaining after related files → doc scan), wiring through tobuildUserPromptAGENTS.md— module map updatedREADME.md— doc-staleness moved from "In progress" to "Shipped"Test plan
doc-mentions.test.ts— full-path match, basename match, generic guard, ranking, constantsworkspace.test.ts— find docs, exclude conventions/changed docs, cap, budget, ranking, binary skip, JSON extensionprompt.test.ts— section ordering with docs, absent when empty, staleness headerphases.test.ts— staleness instruction presentorchestrate.test.ts— budget chain, conventionsFile passthrough, relatedDocs in review context, conditional skipnpm test— 242 passing (31 new)npm run lint— cleannpm run build— cleandocker build .— succeeds🤖 Generated with Claude Code
Summary by CodeRabbit
relatedDocsto the review context and included a related-docs section in prompts, with stronger documentation verification guidance.priority_docsand configurable scan/related caps.