| name | ci-review |
| description | CI-optimized code review: multi-agent parallel review with confidence scoring and atomic GitHub PR review posting. Runs 6 (lean) or 9 (full) specialized review agents including one unconstrained deep-reviewer, scores findings for confidence, filters false positives, and submits a single atomic GitHub PR review with inline comments via gh api. Supports --single for cost-effective single-agent reviews with confidence scoring, --agent profile for AI-authored PRs (surfaces more findings since fixes are cheap). Use --min-severity to control finding threshold. Use when reviewing PRs in CI pipelines, GitHub Actions workflows, or locally. Triggers: /ci-review, review PR, CI code review, automated PR review. Use --model to override the reviewer model (e.g., --model opus for deeper analysis). Use: /ci-review <PR#> [focus text] [--full|--lean|--single] [--agent] [--model sonnet|opus] [--min-severity <level>] |
| user-invocable | true |
| allowed-tools | ["Bash","Read","Grep","Glob","Agent","AskUserQuestion"] |
| argument-hint | <PR#> [focus text] [--full|--lean|--single] [--agent] [--model sonnet|opus] [--min-severity low|medium|high|critical] |
CI Review
Multi-agent code review for pull requests. Posts findings as an atomic GitHub PR review with inline comments.
Before running, read references/REVIEW-POSTING.md now for the review posting format and error handling chain.
Profiles
| Profile | Agents | Use When |
|---|
| single | single-reviewer (one comprehensive agent) | Routine PRs, CI budgets, small diffs โ ~6x cheaper than lean |
| lean (default) | deep-reviewer, guidelines-checker, bug-detector, security-reviewer, silent-failure-hunter, code-simplifier | Every PR โ balanced cost and coverage |
| full | All lean agents + test-analyzer, comment-analyzer, type-analyzer | Critical PRs, large changes, pre-release |
| agent | Same agents as full | PR authored by an AI agent โ surfaces more findings since fixes are cheap |
Agent Profile
The --agent flag activates the agent profile, designed for reviewing AI-authored code. It uses the full agent set but with different thresholds and additional prompt context:
- Uses the full agent set (all 9 review agents)
- Lowers the default
--min-severity to low (surface everything actionable)
- Injects context into each agent prompt: "This PR was authored by an AI agent. Surface all valid findings including minor ones โ the cost of fixing is negligible. Pay attention to AI-specific patterns: over-engineering, hallucinated API usage, unnecessary abstractions, verbose boilerplate, and cargo-culted patterns."
Severity Filter
Use --min-severity <level> to control which findings appear in the review:
| Level | Includes | Default For |
|---|
low | All findings (low, medium, high, critical) | --agent profile |
medium | medium + high + critical | Default for --single, --lean, --full |
high | high + critical | โ |
critical | critical only | โ |
If not specified, --min-severity defaults to medium for normal runs and to low when --agent is used. The severity filter is applied AFTER confidence scoring โ it removes real but low-priority findings, not false positives.
Review Posting Rules (Inlined for Reliability)
These rules are critical. They are also detailed in REVIEW-POSTING.md but inlined here as defense-in-depth:
- Always use event
"COMMENT" โ never "APPROVE" or "REQUEST_CHANGES"
- Build ONE
gh api call that creates the review with all inline comments at once
line is the line number on the new version of the file. Always use side=RIGHT
- Only post actionable inline comments โ no confirmations, no "looks good"
- Do not repeat correctly addressed items
- If no inline comments, omit the comments array and just post the body
Timing Logs
This section is reference guidance. Do not execute anything from it directly โ timing markers are only emitted from within each numbered Step's own Bash invocations below. This section describes the two variants and when to apply each.
Every Step in the ## Workflow section emits a phase-start marker at its beginning and a phase-end marker at its end so the GitHub Actions log shows where time is spent. GitHub Actions renders ::group:: / ::endgroup:: as collapsible sections in the run UI; local runs see them as plain text. Track each Step's elapsed seconds and report them in the Step 8 summary as Phase timings (s): s0=... s1=... ... total=....
Single-call variant โ use when the entire Step fits in one Bash invocation (Steps 0, 1, 2). Capture the start epoch into a shell variable and compute elapsed inline, so no state needs to cross tool calls. Preserve the step's exit status by capturing $? immediately after <step commands> and re-emitting it at the end โ otherwise the trailing echo commands would always return 0 and mask prerequisite or eligibility failures (e.g., gh auth status failing in Step 0). All echoes go to stdout โ per GitHub's workflow-command spec, ::group:: / ::endgroup:: must appear on stdout to render as collapsible sections in the Actions log. For steps whose <step commands> also produces parseable output (e.g., Step 2's gh pr view --json), the output stream will interleave markers with the parseable line; the model reads both and extracts the parseable content visually (the JSON line is self-evident). Pattern: echo "::group::[ci-review] Step N: <name>"; START=$(date +%s); <step commands>; STATUS=$?; echo "[ci-review] Step N done elapsed=$(( $(date +%s) - START ))s"; echo "::endgroup::"; exit $STATUS.
Multi-call variant โ use when the Step spans multiple Bash invocations or waits on subagent tool calls (Steps 3, 3.5, 4, 5, 6, 7). All marker echoes go to stdout (GitHub's workflow-command spec). Any value the model needs to remember for later calls (epochs, SHAs, owner/repo) must be printed to stdout as a labeled line so it survives across tool calls. In the first Bash call of the Step, print echo "::group::[ci-review] Step N: <name>" and date +%s โ remember the printed epoch in your working state. In the last Bash call of the Step, substitute the remembered epoch into echo "[ci-review] Step N done elapsed=$(( $(date +%s) - <REMEMBERED_EPOCH> ))s"; echo "::endgroup::". For Steps with agent fan-out (Steps 4, 5), place the phase-end marker after all agents have returned โ the elapsed value will include their wall-clock time, which is exactly what we want to measure.
Workflow
Step 0: Prerequisites
Verify gh CLI is available and authenticated. Single Bash invocation using the Timing Logs single-call variant:
echo "::group::[ci-review] Step 0: Prerequisites"
START=$(date +%s)
gh auth status
STATUS=$?
echo "[ci-review] Step 0 done elapsed=$(( $(date +%s) - START ))s"
echo "::endgroup::"
exit $STATUS
If gh is not found (exit 127), abort with: "gh CLI is required. Install: https://cli.github.com/"
If not authenticated (non-zero exit from gh auth status), abort with: "gh is not authenticated. Run: gh auth login"
Step 1: Parse Arguments
Emit timing markers via a minimal Bash invocation even though argument parsing itself has no other shell work โ s1=<parse> in the Step 8 summary must be a measured elapsed value, never a placeholder:
echo "::group::[ci-review] Step 1: Parse Arguments"
START=$(date +%s)
echo "[ci-review] Step 1 done elapsed=$(( $(date +%s) - START ))s"
echo "::endgroup::"
Extract from the argument string:
- PR identifier (required): a number (e.g.,
123) or GitHub URL. If a URL, extract the PR number and store as PR_NUMBER. If the URL points to a different repository than the current one, abort with: "Cross-repo URLs are not supported. Run this skill from the target repo, or pass just the PR number."
- Focus text (optional): free text describing what to focus the review on (e.g., "auth flow", "error handling")
- Profile flag (optional):
--single, --lean, --full, or --agent. Default: --lean. --agent implies --full agent set. --single uses one comprehensive agent instead of the specialist fan-out.
- Model override (optional):
--model sonnet|opus. Default: use each agent's built-in model (Sonnet for reviewers, Haiku for scorers). When set, overrides the model for review agents only โ confidence scorers always use Haiku.
- Severity filter (optional):
--min-severity low|medium|high|critical. Default: medium (or low when --agent).
If no PR number is provided, abort with: "Usage: /ci-review <PR#> [focus text] [--full|--lean|--single] [--agent] [--model sonnet|opus] [--min-severity ]"
Step 2: Eligibility Check
Single Bash invocation using the Timing Logs single-call variant. Preserves the gh pr view exit status so non-zero (PR not found, permissions) propagates for the abort check:
echo "::group::[ci-review] Step 2: Eligibility Check"
START=$(date +%s)
gh pr view <PR#> --json state,number,title,url
STATUS=$?
echo "[ci-review] Step 2 done elapsed=$(( $(date +%s) - START ))s"
echo "::endgroup::"
exit $STATUS
- If
gh pr view fails (non-zero exit), abort: "PR #N not found or insufficient permissions."
- If PR state is not
OPEN, abort: "PR #N is {state}. Only open PRs can be reviewed."
Print: "Reviewing PR #N: {title} ({url}) โ profile: {single|lean|full|agent}"
Step 3: Gather Context (Parallel)
Multi-call timing variant. Emit the phase-start marker in a dedicated Bash call before launching the parallel operations below:
echo "::group::[ci-review] Step 3: Gather Context"
date +%s
Then launch these four operations in parallel:
3a. Fetch PR diff:
gh pr diff <PR#>
3b. Fetch PR metadata:
gh pr view <PR#> --json title,body,headRefName,baseRefName,files,additions,deletions,changedFiles
3c. Discover CLAUDE.md files:
Search for CLAUDE.md files in the repo root and in directories containing changed files. Use Glob to find **/CLAUDE.md and Read to load the root CLAUDE.md and any others relevant to the changed files. Preserve scope: prefix each file's contents with its path (e.g., ### /CLAUDE.md, ### /packages/api/CLAUDE.md) so agents know which rules apply to which directories.
3d. Fetch existing PR comments:
Run the fetch-pr-comments.sh script to retrieve all existing comments on this PR (inline review comments, PR-level comments, and review bodies from all authors):
sh ../../scripts/fetch-pr-comments.sh <PR#>
Store the JSON output as EXISTING_COMMENTS. Comment bodies are truncated to 2000 characters by the script to limit context size on comment-heavy PRs โ this is sufficient for content-signal matching since actionable content appears early in comments.
If the script fails (non-zero exit, or stderr contains {"error":...}), log a warning and continue โ cross-run dedup is best-effort. On failure, set EXISTING_COMMENTS to:
{"inline_comments":[],"pr_comments":[],"review_bodies":[],"summary":{"total_inline":0,"total_pr_comments":0,"total_review_bodies":0}}
Print: "Fetched {total_inline} inline comments, {total_pr_comments} PR comments, {total_review_bodies} review bodies for cross-run dedup."
Compile the context bundle:
PR_DIFF: the full diff text
PR_META: title, body, branch names, file list, stats
CLAUDE_MD: path-prefixed CLAUDE.md contents (or "No CLAUDE.md found")
FOCUS: the focus text (or empty)
EXISTING_COMMENTS: structured JSON of all existing PR comments (for cross-run dedup in Step 5)
Large diffs: If the diff exceeds 10,000 lines, warn: "Large diff ({N} lines) โ review quality may degrade for files not near the top of the diff." Do not truncate โ let agents handle context naturally. They can always Read individual files for deeper investigation.
Emit the Step-3 phase-end marker in a final Bash call after the context bundle is compiled (substitute the epoch you remembered from the phase-start call):
echo "[ci-review] Step 3 done elapsed=$(( $(date +%s) - <STEP3_START_EPOCH> ))s"
echo "::endgroup::"
Step 3.5: Ensure PR Branch is Checked Out
Multi-call timing variant. Emit the phase-start marker in a first Bash call that prints everything subsequent calls need โ Bash tool state does not persist across calls, so any value you need later must be echoed to stdout here:
echo "::group::[ci-review] Step 3.5: Ensure PR Branch is Checked Out"
date +%s
gh pr view <PR#> --json headRefOid --jq '"PR_HEAD_SHA=" + .headRefOid'
echo "HEAD_SHA=$(git rev-parse HEAD)"
Review agents use Read, Grep, and git blame to examine the actual code โ not just the diff. Decide the checkout action based on the SHAs:
- If
HEAD_SHA matches PR_HEAD_SHA โ already on the right commit, do nothing.
- Otherwise โ run
gh pr checkout <PR#>.
- If checkout fails โ warn but continue. Pass a note to agents: "File access unavailable โ review from diff only." This lets agents skip
Read/git blame rather than wasting turns on failing tool calls.
Close Step 3.5 with a phase-end Bash call (required on every branch โ already-on-commit, checkout-succeeded, and checkout-failed-but-continuing):
echo "[ci-review] Step 3.5 done elapsed=$(( $(date +%s) - <STEP3_5_START_EPOCH> ))s"
echo "::endgroup::"
Step 4: Launch Review Agents
Multi-call timing variant. This is typically the longest phase โ the elapsed value tells the user exactly how much time agent reasoning consumed. Emit the phase-start marker in a dedicated Bash call before launching agents:
echo "::group::[ci-review] Step 4: Launch Review Agents"
date +%s
Select agents based on profile:
Single agent (when --single):
ci-review:single-reviewer โ one comprehensive agent covering all domains
Lean agents (default, when --lean or no flag):
ci-review:deep-reviewer
ci-review:guidelines-checker
ci-review:bug-detector
ci-review:security-reviewer
ci-review:silent-failure-hunter
ci-review:code-simplifier
Full and agent profile agents (added when --full or --agent):
7. ci-review:test-analyzer
8. ci-review:comment-analyzer
9. ci-review:type-analyzer
Launch ALL selected agents in parallel using the Agent tool. If --model was specified, pass it as the model parameter on each Agent tool call to override the agent definition's default. Each agent receives the same prompt:
Review this pull request. Report findings in your defined output format:
## Findings
1. **[severity]** `file:line` โ description + **Recommendation:**
If no issues found, output "No findings."
## PR Metadata
Title: {title}
Branch: {head} โ {base}
Changed files: {count} ({additions}+ / {deletions}-)
## CLAUDE.md Rules
{CLAUDE_MD contents}
## Focus
{FOCUS text, or "No specific focus โ review broadly."}
## Agent Context
{If --agent: "This PR was authored by an AI agent. Surface all valid findings including minor ones โ the cost of fixing is negligible. Pay attention to AI-specific patterns: over-engineering, hallucinated API usage, unnecessary abstractions, verbose boilerplate, and cargo-culted patterns."}
{If not --agent: omit this section entirely}
## PR Diff
{PR_DIFF}
Agent failure handling: If an agent fails, times out, or returns unparseable output, log the failure and continue with findings from the remaining agents. Do not abort the review because one agent failed. Record the failure in the Step 8 summary (e.g., "Agents: 8 launched, 7 completed, 1 failed").
After all agents have returned (or been recorded as failed), emit the Step-4 phase-end marker (substitute the remembered epoch):
echo "[ci-review] Step 4 done elapsed=$(( $(date +%s) - <STEP4_START_EPOCH> ))s"
echo "::endgroup::"
Step 5: Confidence Scoring
Multi-call timing variant. Emit the phase-start marker in a dedicated Bash call before launching scorers:
echo "::group::[ci-review] Step 5: Confidence Scoring"
date +%s
Collect all findings from all agents into a single list. For each finding, record:
- The finding text (severity, file, line, message, recommendation)
- Which agent produced it
Launch one ci-review:confidence-scorer agent per finding, all in parallel. Haiku is cheap and fast โ independent scoring per finding ensures no cross-contamination and survives partial failures.
Each scorer receives:
Score this review finding from 0-100 for confidence (how certain are you that this finding is factually correct?).
Check the diff to verify the finding is on a changed line, then read the actual code to verify it's real.
Confidence is about accuracy, not importance โ a minor but real style issue should score high.
## Finding
Source agent: {agent-name}
{the finding text: severity, file, line, message, recommendation}
## PR Diff
{PR_DIFF}
Wait for all scorers to complete. Collect their scores.
Scorer failure handling: If a scorer fails or times out, default to confidence score 50 (uncertain). This drops the finding from the confidence filter, erring on the side of excluding unverifiable findings rather than including them.
Confidence filter: Remove all findings with confidence score below 65. A score of 65 means "more likely real than not" โ a reasonable bar that avoids discarding real findings that happen to be minor.
Severity filter: Remove findings whose severity is below the resolved --min-severity threshold (default: medium, or low when --agent). Severity order: critical > high > medium > low.
Deduplicate: Multiple agents often flag the same underlying issue. Before building the review, scan all surviving findings and merge duplicates:
- Exact match: same file and same line โ keep the finding with the higher confidence score
- Near match: same file and lines within 5 of each other, describing the same root cause โ keep the higher-scored one, note the other agent in the
Found by: tag (e.g., Found by: bug-detector, security-reviewer)
- Semantic overlap: different files or lines but describing the same conceptual issue (e.g., "missing null check on user input" flagged independently by bug-detector and security-reviewer at two call sites) โ these are NOT duplicates โ keep both, they're separate instances of the same pattern
Existing comment dedup: After within-run dedup, check each surviving finding against EXISTING_COMMENTS to skip findings already covered by prior comments on this PR. This prevents duplicate postings across multiple ci-review runs, and avoids piling on when humans or other bots already flagged the same issue.
For each surviving finding that has a file and line:
- Search
EXISTING_COMMENTS.inline_comments for any comment where (skip comments with line: null):
path matches the finding's file, AND
line is numeric and within ยฑ5 of the finding's line, AND
- At least one content signal matches (see below)
- If all three conditions match โ mark the finding as already-commented and exclude it
- If no inline match found, search
EXISTING_COMMENTS.pr_comments and EXISTING_COMMENTS.review_bodies:
- The comment body contains the finding's file path, AND
- At least one content signal matches
For findings without a file (body-only findings):
- Search
EXISTING_COMMENTS.pr_comments and EXISTING_COMMENTS.review_bodies for comments where:
- The comment body mentions a file path related to the finding's context (or the finding's description references the same area), AND
- A key phrase match exists (see below)
- Both conditions are required โ body-only matching without a location anchor is too loose and would suppress unrelated architectural findings
Content signal matching โ to avoid false negatives from overly broad matches, require a key phrase match or two or more of the following signals (case-insensitive):
- Severity tag: the comment body contains the finding's severity in a tag-like context (e.g.,
[high], **[high]**)
- Type keyword: the comment body contains the finding's type keyword (
bug, security, error-handling, quality, review, guidelines, test-coverage, comment-accuracy, type-design)
- Key phrases (strongest signal): extract 2โ3 significant noun phrases from the finding description (the core issue โ e.g., "SQL injection", "null check", "race condition", "missing validation") and check if any appear in the comment body. A key phrase match alone is sufficient โ it is specific enough to confirm the same issue.
A single generic signal (severity tag alone or type keyword alone) is NOT sufficient โ common words like high, bug, or review appear in many unrelated comments and would suppress distinct findings near the same location.
Matching policy:
- Match generously โ better to skip a potential duplicate than to re-post noise
- All existing comments are checked regardless of author (human reviewers, bots, previous ci-review runs)
- Resolution status is irrelevant โ whether resolved or unresolved, the issue was already flagged
- String matching is case-insensitive
Track the count of findings excluded by this pass as EXISTING_DEDUP_COUNT.
If no findings survive filtering: Build a no-findings review body using the template from REVIEW-POSTING.md section 3 ("No Findings"), then skip to Step 7 to post it.
After all scorers have returned and filtering/deduplication is complete, emit the Step-5 phase-end marker:
echo "[ci-review] Step 5 done elapsed=$(( $(date +%s) - <STEP5_START_EPOCH> ))s"
echo "::endgroup::"
Step 6: Build Review Payload
Multi-call timing variant. Emit the phase-start marker in a dedicated Bash call before reading references/REVIEW-POSTING.md and deriving types:
echo "::group::[ci-review] Step 6: Build Review Payload"
date +%s
Read references/REVIEW-POSTING.md now for the detailed format specification.
Derive <type> from the source agent. The inline comment format uses **[<severity>] <type>**. Map the agent name to its type:
| Agent | Type |
|---|
| guidelines-checker | guidelines |
| bug-detector | bug |
| security-reviewer | security |
| silent-failure-hunter | error-handling |
| code-simplifier | quality |
| deep-reviewer | review |
| single-reviewer | review |
| test-analyzer | test-coverage |
| comment-analyzer | comment-accuracy |
| type-analyzer | type-design |
For each surviving finding:
-
Determine if it's inline-eligible: Check if the finding's file appears in the PR diff and the line is in a changed hunk. If yes โ inline comment. If no โ body-only finding.
-
Build inline comment object:
{
"path": "<file>",
"line": <line_number>,
"side": "RIGHT",
"body": "**[<severity>] <type>**\n\n<description>\n\n**Recommendation:** <recommendation>\n\n`Found by: <agent-name>`"
}
-
Build review body:
- Summary with profile, finding counts by severity
- If focus text was provided, mention it
- List any body-only findings (not in diff) under "### Findings Not in Diff"
After the payload is built, emit the Step-6 phase-end marker:
echo "[ci-review] Step 6 done elapsed=$(( $(date +%s) - <STEP6_START_EPOCH> ))s"
echo "::endgroup::"
Step 7: Post Review
Multi-call timing variant โ this step spans multiple Bash invocations (OWNER/REPO resolution, payload build, gh api post, and potential error-handling retries). Fold the phase-start marker into the first Bash call (OWNER/REPO resolution below), and emit the phase-end marker at the end of every exit path of the error-handling chain: successful post, the invalid-comments retry, the drop-all-inline-comments fallback, the gh pr comment fallback, and the final stdout-print fallback.
Resolve the repository owner and repo โ Bash tool state does not persist across calls, so print a labeled OWNER_REPO=<owner/repo> line to stdout and remember it. The model substitutes the <OWNER> and <REPO> placeholders in the gh api block below with the remembered value โ no reliance on shell variables from this call:
echo "::group::[ci-review] Step 7: Post Review"
date +%s
gh repo view --json nameWithOwner --jq '"OWNER_REPO=" + .nameWithOwner'
Build the JSON payload using jq and post via gh api. Substitute the owner/repo you remembered from the phase-start call into <OWNER>/<REPO> below:
PAYLOAD=$(jq -n \
--arg event "COMMENT" \
--arg body "$REVIEW_BODY" \
--argjson comments "$COMMENTS_JSON" \
'{event: $event, body: $body, comments: $comments}')
REVIEW_URL=$(echo "$PAYLOAD" | gh api \
"repos/<OWNER>/<REPO>/pulls/<PR#>/reviews" \
--method POST \
--input - \
--jq '.html_url')
If no inline comments, omit the comments array:
PAYLOAD=$(jq -n \
--arg event "COMMENT" \
--arg body "$REVIEW_BODY" \
'{event: $event, body: $body}')
Error handling chain (follow in order):
- If
gh api fails due to invalid inline comments โ remove the invalid comment, rebuild payload, retry (up to 3 times)
- If still failing โ drop all inline comments, move all findings to review body, retry
- If review API fails entirely (403/401) โ fall back to
gh pr comment <PR#> --body "$REVIEW_BODY_WITH_ALL_FINDINGS"
- If everything fails โ print the review body to stdout so the user can post manually
Print the review URL on success, then emit the Step-7 phase-end marker. The same two-line phase-end block must follow every exit path above (successful post, retry success after invalid-comment cleanup, drop-inline fallback success, gh pr comment fallback, and the stdout-print terminal fallback) so the Step-7 group is always closed:
echo "[ci-review] Step 7 done elapsed=$(( $(date +%s) - <STEP7_START_EPOCH> ))s"
echo "::endgroup::"
Step 8: Summary
No ::group:: wrapper for this step โ the summary should be visible at the top level of the log. Still include the total elapsed time.
Print a brief summary for the CI log:
CI Review complete for PR #N
Profile: single|lean|full|agent
Agents: N launched, N completed [, N failed]
Raw findings: N collected
After confidence scoring (โฅ65): N survived
After severity filter (โฅ{min-severity}): N survived
Already commented (skipped): N
Posted: N inline comments + review body
Review: <URL>
Phase timings (s): s0=<prereq> s1=<parse> s2=<eligibility> s3=<context> s3_5=<checkout> s4=<agents> s5=<scoring> s6=<payload> s7=<post> total=<sum>
The Phase timings line is the most load-bearing output for post-run analysis โ list every step's elapsed seconds (using the values you tracked across the Timing Logs markers) plus a total= that is the sum of all phases. Every Step must still be measured and reported when it runs โ even when no branching action occurs (e.g., Step 3.5 when HEAD_SHA already matches PR_HEAD_SHA and no checkout runs, the Step still executed gh pr view + git rev-parse, so its elapsed time should be recorded and reported). Because date +%s has 1-second granularity, legitimately fast steps (especially Step 1's no-op argument parse) may report s<N>=0 โ that is a valid measured value, not a "skipped" marker. A Step's elapsed should be omitted only when the Step was truly not entered (which for Steps 0โ7 would indicate a prompt bug โ they run on every invocation).
Agent Output Format
All review agents output findings in this format (enforced by their agent definitions):
## Findings
1. **[severity]** `file/path.ts:line`
Description of the issue.
**Recommendation:** How to fix it.
2. **[severity]** `file/path.ts:line`
...
Parse each finding to extract: severity, file, line, message (description + recommendation + any extra fields like **Rule:** or **Evidence:**), source agent name.
If an agent outputs "No findings." โ record zero findings from that agent.