<critical_rule priority="9999"> BOTH workstreams MUST be launched in a SINGLE message (parallel invocation). Do NOT wait for one to complete before launching the other.

EXCEPTION: If PR_CONTEXT.linked_issues is empty AND PR_CONTEXT.ci_status != "fail" AND no dependency files changed, the @researcher workstream may be SKIPPED (it would have nothing meaningful to research). In this case, launch only @pr-context-gatherer. </critical_rule>

Skip-Researcher Conditions

Skip the @researcher workstream when ALL of these are true:

PR_CONTEXT.linked_issues is empty
PR_CONTEXT.ci_status is NOT "fail"
No dependency manifest files changed (package.json, requirements.txt, go.mod, Cargo.toml, pom.xml, Gemfile, etc.)
No SECURITY.md or security-related files changed

When skipping researcher, set these REVIEW_CONTEXT fields to empty:

linked_issues_detail: []
dependency_advisories: []
ci_failure_analysis: []
related_prs: []

WORKSTREAM A: PR CONTEXT (@pr-context-gatherer)

DELEGATE TO: @pr-context-gatherer

**TASK**: Analyze all changed files in PR #[pr_number] for code review context

**CHANGED FILES** ([files_changed] files):
[List all files from PR_CONTEXT.changed_files, one per line]

**EXPECTED OUTCOME**:
- Full content of every changed file (post-change state)
- Diff hunks for each file
- Import/export analysis for each file
- Callers of any changed functions/exports
- Associated test files for each changed file
- Git history (recent authors, change frequency) for each file
- Dependency graph (what depends on what)
- Codebase conventions (naming, structure, error handling, test patterns)

**MUST DO**:
- Read EVERY changed file in full (not just diff hunks) — reviewers need full context
- For each changed file, identify its imports and exports
- For each changed export/public function, find callers in the rest of the codebase
- For each changed file, find associated test files (by convention: `*.test.*`, `*.spec.*`, `__tests__/`, or co-located)
- Run `git log --oneline -5 <file>` for each changed file to get recent history
- Build a dependency graph: which changed files depend on each other, and which unchanged files depend on changed files
- Detect codebase conventions by examining 3-5 existing files near the changed files:
  - Naming conventions (camelCase, snake_case, PascalCase)
  - File/directory structure patterns
  - Error handling patterns (try/catch style, Result types, error codes)
  - Test patterns (test framework, assertion style, mocking approach)
  - Import ordering conventions

**MUST NOT DO**:
- Modify any files
- Run tests or builds
- Skip any changed file (even if it looks trivial)
- Guess at conventions without examining actual code

**CONTEXT**:
- Repository: [repo]
- Base branch: [base_branch]
- Head branch: [head_branch]

**TO GET THE DIFF**:
```bash
gh pr diff [pr_number] --repo [repo]

REPORT FORMAT:

### File Map

#### [file_path]
- Language: [lang]
- Diff: [+N/-M lines]
- Imports: [list]
- Exports: [list]  
- Callers: [list of file:function that call into this file]
- Test files: [list or "none found"]
- Git history: last modified [date], recent authors: [list], frequency: [high/medium/low]

[Repeat for each file]

### Dependency Graph
[file_a] → depends on → [file_b]
[file_c] → depended by → [file_d]

### Conventions Detected
- Naming: [description]
- File structure: [description]
- Error handling: [description]
- Test patterns: [description]
- Import order: [description]

### Test Coverage
- Files with tests: [list]
- Files without tests: [list]


---

## WORKSTREAM B: EXTERNAL CONTEXT (@researcher)

**DELEGATE TO**: @researcher

```markdown
**TASK**: Gather external context for PR #[pr_number] review

**PR METADATA**:
- Title: [title]
- Description: [description or "No description provided"]
- Author: @[author]
- Labels: [labels]
- Linked issues: [linked_issues]
- CI status: [ci_status]
- Changed files: [changed_files — abbreviated if > 20, showing first 10 + "and N more"]

**EXPECTED OUTCOME**:
- Detailed info on linked issues (acceptance criteria, context)
- Dependency advisory check (if dependency files changed)
- CI failure analysis (if CI is failing)
- Related PRs (recent PRs touching same files)

**MUST DO**:

1. **Linked Issues** (if any):
   For each issue in [linked_issues]:
   ```bash
   gh issue view <number> --repo [repo] --json title,body,labels

Extract: title, body summary, labels, and any acceptance criteria (look for checkbox lists, "acceptance criteria" headers, or "requirements" sections)

Dependency Advisories (if dependency files changed): Check if any of these files are in the changed list: package.json, package-lock.json, yarn.lock, pnpm-lock.yaml, requirements.txt, Pipfile.lock, go.sum, Cargo.lock, Gemfile.lock

If yes, check for known advisories:
```
# For npm projects
npm audit --json 2>/dev/null | jq '.vulnerabilities | to_entries[:5]'
# For other ecosystems, use web search for recent advisories
```
CI Failure Analysis (if ci_status == "fail"): For each failing check:
- Attempt to fetch logs if URL is available
- Categorize: test failure, build error, lint error, other
- Identify which changed files are likely related
- Provide error summary

Related PRs: Find recent PRs that touch the same files:

gh pr list --repo [repo] --state all --limit 10 --json number,title,files \
  --jq '[.[] | select(.files[]?.path as $p | [CHANGED_FILES] | any(. == $p))][:5]'

If the above is complex, use a simpler approach:

# Search for PRs mentioning the same directories
gh pr list --repo [repo] --state merged --limit 20 --search "path:[key_directory]"

MUST NOT DO:

Modify any files
Run tests or builds
Spend more than 2 queries per linked issue
Fetch full CI logs (summaries only)

REPORT FORMAT:

### Linked Issues

#### #[number]: [title]
- Labels: [labels]
- Summary: [1-2 sentence summary]
- Acceptance Criteria: [extracted list or "none found"]

### Dependency Advisories
[List any advisories found, or "No dependency files changed" / "No advisories found"]

### CI Failure Analysis
[For each failing check: name, failure type, error summary, related files]
[Or "CI passing" / "No CI checks"]

### Related PRs
[List recent PRs touching same files with title and relevance]
[Or "No related PRs found"]


---

## MERGE WORKSTREAM RESULTS

After both workstreams complete (or the single workstream if researcher was skipped), assemble the `REVIEW_CONTEXT` object.

### Assembly Rules

1. **File map**: Direct from @pr-context-gatherer output. Each changed file must have an entry.
2. **Dependency graph**: Direct from @pr-context-gatherer output.
3. **Conventions**: Direct from @pr-context-gatherer output.
4. **Test coverage**: Derived from @pr-context-gatherer's test file associations.
5. **Linked issues detail**: From @researcher (or empty if skipped).
6. **Dependency advisories**: From @researcher (or empty if skipped).
7. **CI failure analysis**: From @researcher (or empty if skipped).
8. **Related PRs**: From @researcher (or empty if skipped).

### Partial Failure Handling

<critical_rule priority="999">
  If ONE workstream fails but the other succeeds, DO NOT abort.
  
  - @pr-context-gatherer fails: This is CRITICAL. The review cannot proceed without
    file context. RETRY ONCE. If retry fails, ABORT with error.
  
  - @researcher fails: This is NON-CRITICAL. Proceed with empty researcher
    fields. Add a note to REVIEW_CONTEXT: "External context gathering failed.
    Review proceeds without linked issue analysis, CI failure analysis, or
    dependency advisory checks."
</critical_rule>

### Validation

Before proceeding to R2, verify REVIEW_CONTEXT:

1. `file_map` has an entry for every file in `PR_CONTEXT.changed_files`
2. `conventions` has at least one non-empty field
3. `dependency_graph` exists (may be empty for unrelated files)

If validation fails, log warning and proceed (degraded review is better than no review).

---

## GATE ENFORCEMENT

<gate id="r1-exit" priority="9999">
  R1 MUST produce a REVIEW_CONTEXT with a populated file_map before proceeding to R2.
  
  VALID REVIEW_CONTEXT requires:
  1. file_map is non-empty
  2. Every file in PR_CONTEXT.changed_files has a file_map entry
     (or a documented reason for absence, e.g., binary file, deleted file)
  3. conventions object exists
  
  If file_map is empty → ABORT (cannot review without file context)
  If file_map is partial → WARN and proceed (degraded review)
</gate>

---

## EDGE CASES

### Binary Files
- Binary files (images, compiled assets) cannot be read as text.
- Add to file_map with `full_content: null`, `language: "binary"`.
- Skip from all R2 review passes.
- If the PR is ONLY binary files, skip to R3 with a note: "All changes are binary files. No code review applicable."

### Deleted Files
- Deleted files have no post-change content.
- Add to file_map with `full_content: null`, note `deleted: true`.
- R2 reviewers should check: are callers of deleted exports updated?

### Renamed/Moved Files
- `gh pr diff` shows renames. Track old → new path.
- Read the new path for full content.
- Check that imports referencing the old path are updated.

### Very Large Files (> 5000 lines)
- Still read in full (reviewers need context).
- But note in file_map: `large_file: true`.
- R2 passes may focus primarily on changed hunks for very large files.

### Submodule Changes
- Submodule pointer changes show as a single-line diff.
- Add to file_map with `language: "submodule"`, `full_content: null`.
- @researcher should check the submodule repo for what changed (if accessible).