agentleFS
WeaverseWeaverse/.agents/skills/agent-work-verification/SKILL.md

agent-work-verification

Skill0 starsChanged 6 days agoMITView on GitHub

Verify a coding agent's returned branch against Git, gate and artifact evidence. Use when delegating bounded implementation or accepting agent-built work, especially when reports claim clean state, baseline failures or published artifacts.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    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
  1. Distinctive:

    If a claimed commit is missing, inspect reflog/object history without resetting or writing lost-found output.

    line 35
  2. Distinctive:

    Keep logs tied to their command's working directory, exit and SHA.

    line 53
  3. Distinctive:

    Verify a missing-runtime/tool claim before authorizing a workaround.

    line 55
  4. Distinctive:

    Check deletion effects before reverting or accepting them.

    line 57
  5. Distinctive:

    Use native git worktree remove within authorization, never force removal as an inspection shortcut.

    line 72
  6. Distinctive:

    Keep unverified claims explicitly separate.

    line 75

≈ “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.

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

Accepting agent-built work

Preserve the task contract

  • Give one writer ownership of the outcome across source, schema, tests and docs. Include repository, branch, exact base, authoritative dirty work, required surviving capabilities, forbidden side effects and expected commands/results. Never put a second writer into a live owner's checkout.
  • Pass a brief file rather than long Markdown in a double-quoted shell argument: backticks can execute and transports can truncate it. A mid-run correction should identify only the changed contract and affected boundary.
  • A deletion-only brief can remove the sole creation path across repositories. Name the capability to preserve, and check replacement SDK types/actual dispatch before accepting a keyword match as feature parity.

Authenticate the return

git branch --show-current
git rev-parse HEAD
git status --short --untracked-files=all
git log --oneline -5
git reflog -10
git diff --stat <base>...HEAD
git diff --check <base> HEAD
  • A clean checkout's bare git diff --check checks no committed change. Include new/untracked files in the inventory. Check staged bytes against what will actually commit, not only working-tree text.
  • A canonical checkout may be sitting on the candidate too. Authenticate both candidate and exact base before comparing test totals. Read the target's current tip and inspect whether landing would discard intervening work.
  • If a claimed commit is missing, inspect reflog/object history without resetting or writing lost-found output. Preserve the complete recovered chain on a task-owned ref only within authority; do not reset to the first orphan found and lose its descendants.
  • Inspect push configuration and use an explicit SHA:refs/heads/branch with follow-tags disabled when pushing is authorized. Do not infer a destination from the local branch name or upstream alone. A mistaken push needs immediate disclosure and an approved recovery, not an automatic force-push.
  • Missing SSH allowed-signers configuration makes git verify-commit inconclusive, not proof of an invalid signature. Establish the verifier's trusted identity without weakening signing or verification guards.

Evidence stronger than a completion report

  • Claimed image publication, resource creation and releases need registry/platform readback at the exact identity. A build log, local tag or intended URL does not prove the consumer can fetch the artifact.
  • For "pre-existing failures", execute the same failing gate on the exact base in an isolated checkout using its own correctly resolved workspace dependencies. Regenerate ignored prerequisites without applying schema or copying the candidate fix. A symlinked dependency graph can secretly execute the candidate.
  • Compare failure signatures, not just totals. Rotating failures can be timing/load-sensitive. A failure present at base establishes attribution; it does not waive a release gate or justify saying the gate passed.
  • Keep logs tied to their command's working directory, exit and SHA. A wrapper's trailing echo or a cached log can mask failure. An errored/partial reviewer supplies no clean verdict; historical SHA reviews need rechecking.
  • Verify a missing-runtime/tool claim before authorizing a workaround. Never source an agent's environment file into the persistent shell; inspect it and use the approved isolated environment for reproduction.
  • Check deletion effects before reverting or accepting them. A removed module may be inert, or it may be the only remaining lifecycle path; import counts alone cannot prove dynamically dispatched behavior.

Boundaries to inspect explicitly

  • Re-scan the exact migration diff and ledger. A follow-up correction can silently edit applied history. database-migration-review owns numbering, snapshots, live-catalog and unresolved-operation proof.
  • A config/filter change can make the target test pass by excluding unrelated tests. Compare executed file sets, not only total counts. For a Dockerfile change, build the real target or report image behavior unproven.
  • Copy/clone/move/import operations need sourceId === targetId and tenant-boundary cases; a nontransactional read-then-write can duplicate children or erase settings even when ordinary inputs pass.
  • A source fix and green test are not enough if production cannot reach the branch. Use production-debugging for payload-shape, path-reachability and negative-control proof.
  • Retire a worktree only when its work is retained, it is clean (including untracked files) and no live owner uses it. Ignored dependency trees are not protected by a clean status; preserve anything valuable first. Use native git worktree remove within authorization, never force removal as an inspection shortcut.

Report verified branch/base/SHA/tree, changed paths, actual gate exits, artifact identities and remaining gaps. Keep unverified claims explicitly separate; this playbook does not authorize recovery writes or publication.

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.