track-review-findings
Skill0 starsChanged 17 days agoMITView on GitHub
Use when the same review, audit, or automated check runs repeatedly against content that keeps changing (a PR re-reviewed on every push, a recurring security scan, a repeated lint or test-quality pass) and unchanged issues get re-flagged as noise on every round, fixed issues vanish from the report with no confirmation they were actually fixed, or a regression that was already reported once slips back in without anyone noticing it returned.
In this repository23 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- agent-orchestrationSkill
- ask-for-approvalSkill
- claim-verification-loopSkill
- code-councilSkill
- code-reviewSkill
- context-as-imagesSkill
- grillingSkill
- grow-test-coverageSkill
- investigateSkill
- loop-until-verifiedSkill
- plan-and-buildSkill
- prompt-cachingSkill
- proofreadSkill
- review-allSkill
- review-before-mergeSkill
- review-panelSkill
- to-questionnaireSkill
- trim-contextSkill
- unslopSkill
- verify-findingsSkill
- whatnowSkill
- wizardSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 23
Stop condition: every current finding is matched or added, every previous finding has an explicit verdict, and the new state is written …
- Distinctive: line 36
No durable state at all: every run reinvents the list, nothing is ever confirmed resolved.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Track review findings
Overview
When a check repeats against changing content, do not treat each run's output as a fresh list. Give every finding a stable identity, carry a small state record forward, and reconcile: add new findings, leave matched ones alone, resolve missing ones only when a full check just ran, and reopen anything that comes back after being resolved.
When to use
- A review, audit, or scan repeats on the same evolving target (PR pushes, nightly scans, fix loops).
- Unchanged issues should stay quiet instead of re-alerting every round.
- You need to know what was actually fixed, not just what went undetected this time.
Skip this for a one-shot check that never runs again on the same target, or when there is nowhere durable to store state between runs (see Common mistakes).
The pattern
- Fingerprint, don't diff text. Derive each finding's ID from stable fields (file path plus a whitespace-normalized, lowercased body), never from exact wording or line numbers. A rewrap or a shifted line must not mint a new ID.
- Persist state somewhere durable and visible. Store
{finding_id, state: ACTIVE|RESOLVED, first_seen, last_seen, resolved_at?, reopened_count}per finding. With no database, embed it as a hidden, versioned marker inside the artifact people already read (an HTML comment in a PR comment) and parse it back next run. - Reconcile every run: current-only findings become ACTIVE; matched findings get
last_seenrefreshed, flipping RESOLVED back to ACTIVE withreopened_count += 1if they returned; previous-only findings become RESOLVED, but only when this run fully covered the same scope and the target changed (compare head SHAs), otherwise leave them ACTIVE since a partial run proves nothing. A run where any reviewer died, errored or returned nothing did not cover the full scope, so nothing is resolved from it. Report only what changed state. - Cap what you keep. Bound stored RESOLVED entries (for example, 20 most recent); never cap ACTIVE ones.
- Stop condition: every current finding is matched or added, every previous finding has an explicit verdict, and the new state is written back to where it was read from.
Quick reference
- New -> ACTIVE,
first_seen = now. - Still present -> keep, refresh
last_seen. - Was RESOLVED, reappeared -> ACTIVE,
reopened_count += 1. - Missing, full fresh check -> RESOLVED.
- Missing, partial or unchanged check -> leave ACTIVE.
- Unchanged this round -> do not re-report.
Common mistakes
- Fingerprinting on exact text or line number orphans a finding's history on a harmless rewrap or line shift.
- Marking something resolved without confirming full scope and a changed target; a narrow or failed run then reports false fixes.
- No durable state at all: every run reinvents the list, nothing is ever confirmed resolved.
- Unbounded resolved history drowns the active list over time.
Source
Repo: https://github.com/The-PR-Agent/pr-agent, commit 1f11725b4d96afbaad8ed956afd074ecf4642426, license MIT.
File read: pr_agent/algo/review_finding_state.py (fingerprinting, state marker parse/serialize, reconcile_review_findings).
Adapted: generalized pr-agent's per-comment ACTIVE/RESOLVED reconciliation into a project-agnostic pattern. Wording rewritten, not copied, per MIT attribution.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

