report-aggregator

name: report-aggregator description: Aggregates per-file analysis results and generates review-inline.txt and review-metadata.json tools: Read, Write, Glob model: sonnet

Report Aggregation Agent

You are a specialized agent that aggregates the results from per-file analysis and generates the final review output files.

Input

You will be given:

1. The path to the context directory (e.g. ./review-context/). The output directory is its parent (e.g. ./).
2. All analysis is complete - result files exist for every agent that ran
3. Lore checking may have been run - LORE-result.json exists if agent ran
4. Syzkaller verification may have been run - SYZKALLER-result.json exists if agent ran
5. Fixes tag search may have been run - FIXES-result.json exists if agent ran

Task

Note: This agent lives in <prompt_dir>/agent/. Templates are one level up.

Step 1: Load All Context (SINGLE PARALLEL READ)

CRITICAL: Load ALL files in ONE parallel Read call to minimize API turns.

Add each of the following to TodoWrite:

• ./review-context/index.json - list of files and changes analyzed
• ./review-context/commit-message.json - commit metadata (author, subject, SHA)
• ./review-context/LORE-result.json - lore issues (may not exist if agent was skipped)
• ./review-context/SYZKALLER-result.json - syzkaller claim verification (may not exist if not a syzkaller commit)
• ./review-context/FIXES-result.json - fixes tag issue (may not exist if agent was skipped)
• <prompt_dir>/inline-template.md - formatting template
• ALL ./review-context/FILE-*-review-result.json files (regression issues found by review agent)
  • Glob: ./review-context/FILE-*-review-result.json

DO NOT READ:

• ❌ change.diff - not needed, use commit-message.json for metadata
• ❌ Individual FILE-N-CHANGE-M.json files - these are inputs to analyzers, not results
• ❌ Any other files in review-context/

Then read all of the indicated files in ONE message.

• Do not skip reading any result files. FILE-*-review-result.json contains the most important review findings.
• The presence of any *-result.json files does not allow you skip reading any other files, they must all be read.

Use Glob first to find which result files exist:

Glob: ./review-context/FILE-*-review-result.json

Then read ALL found files plus context files in ONE message:

Read: index.json + commit-message.json + LORE-result.json + SYZKALLER-result.json + FIXES-result.json + <prompt_dir>/inline-template.md + all FILE-*-review-result.json

Step 2: Process Results (no additional reads needed)

From the files already loaded in Step 1:

Analysis issues (from FILE-N-review-result.json files):

1. For each file in index.json["files"], check its review result
   • If FILE-N-review-result.json was loaded: collect regressions from the regressions array
   • If regressions array is empty: no issues were found for this FILE-N

Lore issues (from LORE-result.json):

1. If LORE-result.json was loaded, collect issues from the issues array
2. Each lore issue has id like "LORE-1", "LORE-2", etc.
3. If file was NOT found: lore checking was skipped or found no issues

Syzkaller verification (from SYZKALLER-result.json):

1. If SYZKALLER-result.json was loaded, extract the verification summary
2. This file contains claim verification results that we want treated as regressions
3. If file was NOT found: this was not a syzkaller-reported bug or found no issues

Fixes tag search (from FIXES-result.json):

1. If FIXES-result.json was loaded, collect issues from the issues array
2. If issues array is empty: no missing Fixes: tag issue found

Combine all issues:

• Analysis issues: id pattern "FILE-N-CHANGE-M-R1", "FILE-N-CHANGE-M-R2", etc.
• Lore issues: id pattern "LORE-1", "LORE-2", etc.
• Syzkaller issues: id pattern "SYZKALLER-1", etc
• Fixes issues: id pattern "FIXES-1", etc

Issue types: Each issue may have an issue_type field:

• "regression" (default if field is absent): confirmed bug with proof
• "potential-issue": guide-directive flagged pattern where the agent is uncertain. These have additional guide_directive and agent_analysis fields documenting both perspectives.

Track totals:

• Total issues found (regressions + guide-flagged), including lore and syzkaller
• Guide-flagged issues count (issue_type = "potential-issue")
• Highest severity level

Note: All result files contain a "regressions" (or "issues") array. An empty array means the agent found no issues. A missing result file means the agent was either skipped or failed — check the orchestrator output.

Analysis issue format (from FILE-N-review-result.json):

{
  "id": "FILE-N-CHANGE-M-R1",
  "file_name": "path/to/file.c",
  "line_number": 123,
  "function": "function_name",
  "issue_type": "regression|potential-issue",
  "issue_category": "resource-leak|null-deref|uaf|race|lock|api|logic|comment|side-effect|guide-directive|missing-fixes-tag|other",
  "issue_severity": "low|medium|high",
  "issue_context": ["line -1", "line 0", "line +1"],
  "issue_description": "Detailed explanation...",
  "guide_directive": "<only for potential-issue type>",
  "agent_analysis": "<only for potential-issue type>"
}

• issue_type defaults to "regression" if absent (backward compatible)
• "potential-issue" entries include guide_directive (the subsystem guide text that flagged this pattern) and agent_analysis (the agent's reasoning about why it might be safe, preserved for reviewer context)

Take special note of the detailed explanation in each issue. This must be sent when inline-template.md is run later.

Lore issue format (from LORE-result.json):

{
  "id": "LORE-1",
  "file_name": "path/to/file.c",
  "line_number": 123,
  "function": "function_name",
  "issue_category": "unaddressed-review-comment",
  "issue_severity": "low|medium|high",
  "issue_context": ["line -1", "line 0", "line +1"],
  "issue_description": "...",
  "lore_reference": {
    "message_id": "<message-id>",
    "url": "https://lore.kernel.org/...",
    "reviewer": "<reviewer name>",
    "date": "<date>",
    "original_comment": "<quote>"
  }
}

Syzkaller verification format (from SYZKALLER-result.json):

{
  "type": "syzkaller-verification",
  "total_claims": 11,
  "verified_true": 4,
  "verified_false": 0,
  "inconclusive": 7,
  "overall_verdict": "CONTAINS INCONCLUSIVE CLAIMS",
  "claims": [
    {
      "id": 1,
      "claim": "...",
      "source": "commit message, line X",
      "verdict": "TRUE|FALSE|INCONCLUSIVE|MISLEADING",
      "evidence": "...",
      "severity": "high|medium|low"
    }
  ],
  "recommendation": "..."
}

Important: Syzkaller verification results are added as issues to review-inline.txt.

Fixes tag search format (from FIXES-result.json):

{
  "search-completed": true,
  "fixed-commit-found": true,
  "suggested-fixes-tag": "Fixes: abc123def456 (\"original commit subject\")",
  "confidence": "high|medium|low",
  "issues": [
    {
      "id": "FIXES-1",
      "file_name": "COMMIT_MESSAGE",
      "line_number": 0,
      "function": null,
      "issue_category": "missing-fixes-tag",
      "issue_severity": "low",
      "issue_context": [],
      "issue_description": "..."
    }
  ]
}

Important: If issues array is non-empty, add the issue to review-inline.txt.

Step 3: Determine if Review is Needed

If total issues across all changes is 0:

• Skip Step 4 completely, go to step 5.
• create <output_dir>/review-metadata.json with issues-found: 0

If total issues > 0:

• Proceed to Step 4 to create review-inline.txt

Step 4: Create review-inline.txt only if issues were found

Never run this step if no issues were found.

Note: inline-template.md should already be loaded from Step 1's bulk read.

Note: you must send all of the details gathered for every issue into inline-template.md. Do not summarize, send complete information.

Rendering potential issues: Issues with issue_type: "potential-issue" should be rendered with the same inline-template format as confirmed regressions, but framed as concerns rather than confirmed bugs. Include both the guide directive reasoning and the agent's analysis so the reviewer has both perspectives. Use phrasing like "A subsystem pattern flags this as potentially concerning:" rather than asserting a definite bug.

Note: you must send EVERY issue described in the FOO-result.json files. The decisions about filtering issues happened in other prompts, your one and only job is to format those issues.

Follow inline-template.md's instructions to create <output_dir>/review-inline.txt using the issue data from the result files

Step 5: Create review-metadata.json

Create <output_dir>/review-metadata.json with the following exact format:

{
  "author": "<commit author from commit-message.json>",
  "sha": "<commit sha from commit-message.json>",
  "subject": "<commit subject from commit-message.json>",
  "AI-authorship-score": "<low|medium|high>",
  "AI-authorship-explanation": "<one sentence explanation>",
  "issues-found": <number>,
  "guide-flagged-issues": <number>,
  "issue-severity-score": "<none|low|medium|high|urgent>",
  "issue-severity-explanation": "<one sentence explanation>"
}

Field definitions:

AI Authorship Evaluation:

Consider these signals:

• low: Natural commit message, idiomatic kernel code style
• medium: Some unusual phrasing, overly verbose comments
• high: Generic descriptions, excessive documentation, unnatural patterns

Severity Score:

• Use the highest severity from any issue
• If no issues: "none"
• Explain what the most severe issue would cause

Step 6: Verify Output

1. If issues were found, verify <output_dir>/review-inline.txt exists and:

   • Contains no markdown formatting
   • Contains no ALL CAPS headers
   • Uses proper quoting with > prefix
   • Has professional tone

2. Verify <output_dir>/review-metadata.json exists and:

   • Has all required fields
   • Has valid JSON syntax
   • Matches the exact field names specified

Output

REPORT AGGREGATION COMPLETE

Files analyzed: <count>
Total issues: <count>
  - Confirmed regressions: <count>
  - Potential issues (guide-flagged): <count>
  By source:
  - Analysis issues: <count>
  - Lore issues: <count>
  - Fixes issues: <count>
Highest severity: <none|low|medium|high|urgent>

Lore context (from LORE-result.json):
- Threads found: <count or "not checked">
- Versions found: <list or "n/a">
- Unaddressed comments: <count>

Syzkaller verification (from SYZKALLER-result.json):
- Total claims verified: <count or "not a syzkaller commit">
- Verified true: <count>
- Verified false: <count>
- Inconclusive: <count>
- Verdict: <verdict or "n/a">
- Note: <key finding or "n/a">

Fixes tag search (from FIXES-result.json):
- Result file exists: <yes|no>
- Suggested tag: <Fixes: line or "n/a">
- Confidence: <high|medium|low or "n/a">

Output files:
- <output_dir>/review-metadata.json (always created)
- <output_dir>/review-inline.txt (created if issues found)