AGENT SKILL

verify-review-reservation

Recheck the bounded identity-ledger reservation invariant after lock, hash-check or publication changes.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
0

01 前言 Front Matter

name
verify-review-reservation
description
Recheck the bounded identity-ledger reservation invariant after lock, hash-check or publication changes.

03 SKILL.md 正文

Verify review reservation

This opt-in verification profile is for the sdd-forge source repository. It is supplemental evidence; it never replaces normal tests, review or task gates.

Property and scope

Two contenders holding the same original ledger hash cannot both succeed in one reservation attempt. The user harm is two reviewers admitted against conflicting identity history. tests/fixtures/project-skills/formal-review-reservation/mapping.json pins the source and maps atomic mkdir, the post-lock hash comparison, staged append/rename and lock release to model actions. The finite model has two actors, one attempt each, and no crash, retry, stale-lock recovery, malicious filesystem or Windows semantics. All writers must honor the same lock; mkdir and same-directory rename must be atomic and hashes collision-free. A finite-model PASS is not a proof of all implementation schedules or operating systems.

The covered source is plugins/sdd-quality-loop/scripts/validate-review-context-set.sh in the combined integration candidate, SHA-256 8d4397f7120d07b6eb0513049c519367de7e9cd1d61ea2814c1b8ffcd28dedcf. The unchanged acquire/hash-check/append/rename/release tail has SHA-256 099073aafef9082cab371bfab889a19afe4e40a3f9d791739ceb820052a4a92c. Evidence for the combined candidate is in reports/verification/formal-reservation-combined-20261028/. The separately recorded historical asset check (id 683491413) is in reports/verification/formal-reservation-asset-683491413-20261029/; earlier main-based evidence remains historical. Use the local pstack-maintain-formal-verification-skill procedure when the mapping drifts.

Tool choice and setup

TLC from the fixed v1.7.4 release asset explores the interleavings directly and is smaller here than translating shell/jq/filesystem operations into a code-level verifier. Java must be installed. Release asset 184694200 is available from https://github.com/tlaplus/tlaplus/releases/download/v1.7.4/tla2tools.jar; query gh api repos/tlaplus/tlaplus/releases/assets/184694200 for metadata. Download the jar outside the repository and verify SHA-256 936a262061c914694dfd669a543be24573c45d5aa0ff20a8b96b23d01e050e88. The helper rejects all other bytes before creating an evidence directory. These bytes report TLC2 Version 2.19 of 08 August 2024 (rev: 5a47802).

The previously selected v1.8.0 asset 683491413 returned HTTP 404 on 2026-10-31; assets under that tag had already been replaced repeatedly. The stable release is sufficient for this unchanged finite model and avoids following the moving asset. This does not guarantee perpetual upstream availability: if the exact asset disappears or differs, stop rather than accepting a new jar implicitly. Historical v1.8.0 evidence remains unchanged. The new tool's negative, baseline and actual-code race observations are recorded separately in reports/verification/formal-reservation-stable-20261031/. Bash, jq and sha256sum are required for the Linux real-code replay. No network call occurs during verification.

Run and observe

From the source repository, execute:

python3 tests/verify-review-reservation.py --jar /path/to/tla2tools-1.7.4.jar --output /path/to/new-evidence-directory

The output directory must not exist. The helper first verifies the source pin, then runs the lock-removed negative model and the intact baseline. It retains TLC logs, counterexample traces, tool hash, source hash and measured time. It also extracts the actual reservation transaction from the pinned Bash script into isolated scratch state and invokes two competing processes. The negative mutates that extracted code's acquire/release guard; both calls must succeed. The intact code must accept only one. The fixture replaces unrelated input-validation setup, so this proves the reservation tail's observed behavior, not the whole review CLI. For the negative replay the parent releases both contenders only after both reach the post-hash-check barrier; timeout is a synchronization failure.

Maintenance and cleanup

A source-pin mismatch is blocked. Inspect changed code/model correspondence before updating the pin; do not update it solely to make the command pass. A baseline counterexample is a product/model investigation, never permission to weaken the property. Scratch mutations stay outside the product checkout and are removed when each replay finishes; retained evidence survives. No scheduled invocation or host-support claim is implied. Re-run normal project tests independently.

04 同仓库更多 Skills