agentleFS
edxleonedxleon/runspector/.claude/skills/deep-review/SKILL.md

deep-review

Skill11 starsChanged 13 days agoMITView on GitHub

Deep review of a change before it is committed - for the write path to Garmin, the agent's fence, the MCP tool set, the workflows, a migration, or any change over ~1,000 lines. Fresh reviewers, every important finding proven by a red test, a second round on the fixes, a live run on throw-away data, a fresh reader last.

47 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use react, testing-strategy and 8 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • VercelSkill

    react

    React renderer for json-render that turns JSON specs into React components. Use when working with @json-render/react, building React UIs from JSON, creating component catalogs, or rendering AI-generated specs.

    Also in 3 projects that use this

  • BlockSkill

    testing-strategy

    Guidelines for writing effective tests in this project

    Also in 3 projects that use this

  • SentrySkill

    deslop

    Strip AI-generated noise from the diff before pushing — extra comments a human wouldn't write, defensive try/catch in trusted code paths, casts to any, inline imports in Python, and other style inconsistencies with the surrounding file. Use this immediately before commit so the diff stays clean.

    Also in 3 projects that use this

  • SentrySkill

    review

    Fixture review skill.

    Also in 5 projects that use this

  • AnthropicSkill

    debug

    Structured debugging session — reproduce, isolate, diagnose, and fix. Trigger with an error message or stack trace, "this works in staging but not prod", "something broke after the deploy", or when behavior diverges from expected and the cause isn't obvious.

    Also in 3 projects that use this

  • NVIDIASkill

    testing

    Run Python and Rust tests for physicsnemo-curator using uv + pytest with coverage reporting via pytest-cov, and cargo-nextest for Rust tests. Includes benchmark workflows with pytest-benchmark and criterion.

    Also in 3 projects that use this

  • +1 more from getsentry
  • Other versions47 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 47

    1. lobehub/lobe-chat★ 86ksame name
    2. lobehub/lobehub★ 86ksame name
    3. dyad-sh/dyad★ 22ksame name
    4. coder/coder★ 17ksame name
    5. tikv/tikv★ 17ksame name
    6. apache/cassandra★ 10ksame name
    7. gitkraken/vscode-gitlens★ 10ksame name
    8. compozy/compozy★ 2.9ksame name
    9. pedronauck/skills★ 655same name
    10. microsoft/regorus★ 375same name

    37 more not listed

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository3 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 41 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Only what makes the change wrong for a runner:

    line 22
  2. Very specific:

    Every finding names file:line and a concrete failure scenario.

    line 34
  3. Distinctive:

    Start subagents in parallel.

    line 40
  4. Distinctive:

    Done when: every reviewer returned a list, empty or not.

    line 59
  5. Very specific:

    Refuted: only with the file:line that contradicts it.

    line 70
  6. Distinctive:

    When round 2 finds a new important finding of a class from round 1, close the class instead of the case: one function every caller asks, or …

    line 79
  7. Distinctive:

    Ship: when no proven finding is open, the live run matched, and the reader had no concrete scenario.

    line 109

“Common” is measured: kept by at least 6 in 10 of the 41 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Deep review

CI proves the code does what its tests say. This review looks for what no test says yet. It runs before the commit, on the uncommitted diff (git diff HEAD), or on the last commit (git show HEAD) when the diff is empty.

When

ChangeReview
plan.py (apply, undo, claims), garmin.py write calls, web/agent.py, mcp_server.py tool set, .github/workflows/**, src/runspector/migrations/*.sql, or over ~1,000 linesfull: steps 1-5
a rule (logic.py, planning.py, plan.py, habits.py) or a prompt (skills/*.md)light: steps 1-4, one reviewer
anything elsethe checks in AGENTS.md are enough

What counts

Only what makes the change wrong for a runner:

  • a correctness bug: a concrete input or state that gives a wrong result or a crash - above all a wrong call for the day or a wrong week;
  • a security hole: injection (Garmin's free text and the athlete's own text reach prompts and the page), a secret in a log, argv or URL, a tool an unattended card run must not reach, a write to Garmin without the athlete's yes;
  • data loss, or a silently wrong number (training load, HRV, a pace, minutes of a week);
  • a test that no longer tests: an assertion deleted or weakened, a new skip, a mock that hides the change;
  • a migration that breaks an applied schema or deletes data.

Every finding names file:line and a concrete failure scenario. Style, naming, docs, speculative performance and unchanged code are not findings, unless the change makes old code reachable. At most five nits; an empty review is a valid result.

Step 1 - Fresh reviewers

Start subagents in parallel. They judge the code and the requirement; they do not get your reasoning, earlier results or any claim about quality ("fixed", "safe", "tested") - those steer a reviewer away from the bug.

  • Correctness gets the diff and the requirement as it stood before the code (the issue, the roadmap section, the plan line), as a claim to check against the code.
  • Security and data (full only) gets the diff and goes through "What counts" point by point: what an unattended card run can reach, what reaches the watch, what free text reaches a prompt.
  • Over ~1,000 lines: one correctness reviewer per group of files (rules, store and migrations; tools, MCP server and web; simulator and evals).

The prompt for each:

Review this diff against "What counts" in .claude/skills/deep-review/SKILL.md. Report what it
calls important and at most five nits. Each finding: file:line, a concrete failure scenario
(input or state -> wrong result), pre-existing yes/no. Nothing found = an empty list.

Done when every reviewer returned a list, empty or not.

Step 2 - Prove

The referee is a red test, not a second opinion.

  • Important finding: a test written from the finding's scenario, not from the code. It counts only when it was seen red at its own assertion while the fix was missing - run it before the fix, or take the fix lines out briefly and put them back.
  • Not reproducible: "plausible, not reproduced"; the live run or the maintainer decides. It is never counted as refuted.
  • Refuted only with the file:line that contradicts it.
  • Nit: fix it, or drop it with one sentence why.

Step 3 - Fix, round 2

Fix each proven finding until its test is green. Then a fresh reviewer gets the fixes, one line per class of error found in round 1 ("a check-in was not counted as an input of the coach card"), and the question whether a fix removed a check, a guard or a test.

When round 2 finds a new important finding of a class from round 1, close the class instead of the case: one function every caller asks, or a property test over a grid of inputs ("no check-in ever makes a week bigger"). Where that is not possible, the maintainer decides.

Step 4 - Live run

At least one real story end to end, on data no test generated:

  • a throw-away RUNSPECTOR_HOME - a copy of a real database (sqlite3 <db> ".backup <copy>", without the garmin/ token folder), deleted afterwards - or runspector serve --demo;
  • the page in a browser, runspector today, a real Claude chat over runspector mcp, and for a fence change a real card run that tries what it must not do;
  • a write to a real Garmin account only after the maintainer's explicit yes, with undo_workout as the way back.

Done when what you saw matches the requirement; otherwise back to step 2.

Step 5 - Fresh reader (full only)

A new agent sees only the final diff - no findings, no history, no claim about quality:

Name the most likely way this breaks in production: file:line, input or state -> wrong result.
If you find none, write "nothing found"; "unsure: <what I could not check>" is allowed.

A concrete scenario goes back to step 2. "Nothing found" and "unsure" end the review.

Finish

  • Ship when no proven finding is open, the live run matched, and the reader had no concrete scenario.
  • Code comments give the reason for a line; the review's story goes into the commit message, in plain words: what was found, what was fixed, and an "unsure:" line for what the reader or the live run could not settle. A bug found only in step 4 or 5 gets its regression test and a line there.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.