Anchor Vet Agent Skill

anchor-vet

Force the skipped move in code review: prove the reviewed diff's scope, blast

radius, claims, findings, and verdict are anchored to git evidence.

Trigger Check

Use this skill only when there is a concrete code change to judge: a PR, branch,

commit range, or staged review target.

Boundary test: can you name a base...head pair of git states whose difference

is the artifact under judgment? If yes, use anchor-vet. If no, get the branch or ref

first; do not review pasted diffs or prose descriptions.

Do not use anchor-vet for design feedback without a diff, whole-codebase audits,

feature implementation, bug fixing, formatting-only tasks, or document review.

If the user asks you to fix findings too, render the review verdict first; fixes

are a separate task, and behavioral fixes route to anchor-repro.

Forced Procedure

• Anchor the scope. Check out the reviewed head locally. Capture external

claims before init: for PRs, save the title/body from `gh pr view

--json title,body --jq '.title + "\n\n" + (.body // "")'` to a claims file.

Run:

python3 skills/anchor-vet/scripts/vet_gate.py init --project <root> --review <slug> --base <ref> --head <ref> --claims-file <path>

or, only when there are genuinely no external claims, --no-claims.

• Read every hunk. Inspect the full diff and every hunk in scope.json.

If the diff is too large to read honestly, stop and propose a split.

• Walk the blast radius. For each candidate symbol, read the references

found by the gate. Record examined refs with effect and a one-line note.

Examined means you opened/read the reference and judged how the change

affects it.

• Account weak or symbolless hunks. For weak/symbolless hunks, use a

verified enclosingUnit when a named unit exists, or file-level with a

reason only for genuinely unit-less code/config/prose.

• Cross-examine claims. Disposition every anchored claim id from

scope.json: verified, contradicted, not-verifiable-from-diff, or

non-claim. Verified/contradicted claims need citations.

• Hunt findings with severity honesty. Use only blocking, concern, and

nit. Cite findings. A blocking finding needs R2 cited evidence or an R3

repro probe. When a behavioral assertion is cheap to demonstrate, record a

failing anchor-repro ledger entry against the reviewed head and cite it with

entrySha256.

• Write and check review.json. Run:

python3 skills/anchor-vet/scripts/vet_gate.py check --project <root> --review <slug>.

Fix check failures by doing the missing review work, not by padding JSON.

• Render the verdict through the gate. Run:

python3 skills/anchor-vet/scripts/vet_gate.py close --project <root> --review <slug> --verdict approve|request-changes|blocked.

Report the verdict, cite verdict.json, and name each finding's evidence

rung.

Script Contract

Artifacts are written under the target project:

.techne/review/<slug>/
  scope.json    # computed by init
  review.json   # authored by the reviewer
  report.json   # computed by check
  verdict.json  # computed by close

Generated .techne/ output belongs to target projects. Do not commit it to this

repository.

Review skeleton:

{
  "schema": "techne.vet/1",
  "symbols": [
    {
      "id": "s-...",
      "symbol": "changedName",
      "disposition": "reviewed",
      "refs": [
        {
          "file": "src/caller.py",
          "line": 42,
          "effect": "unaffected",
          "note": "caller passes the same contract"
        }
      ]
    }
  ],
  "hunks": [],
  "claims": [
    {
      "ids": ["c-..."],
      "disposition": "verified",
      "citations": [{"file": "src/change.py", "line": 10}]
    }
  ],
  "findings": [],
  "testsAcknowledgment": "No tests changed; the diff is docs-only or existing coverage is enough because ..."
}

Use reference.md for the full JSON schema, rubrics, exclusions, weak spots,

and probe guidance.

Stop Conditions

• Stop if no git-anchored base...head diff is available.
• Stop if init refuses because of dirty worktree, missing merge base, or head

mismatch; fix the setup or record blocked.

• Stop if the review is too large to read in full; propose a split instead of

sampling silently.

• Stop before approve if refs, hunks, claims, or test acknowledgment remain

unaccounted.

• If head moves mid-review, do not fight the tool. Close the old slug as

blocked --reason ... or leave it, then start a new slug for the new head.