verify-setup
/verify-setup
Give a repository its own executable verification kit under .agents/verify: doctor, deterministic checks, per-feature drivers that observe the real surface and keep evidence tied to the exact tree, and an enforcer table. Use when a repo has no way for an agent to prove a change works, when setting up verification for a new repo, when adding a driver for a feature, when a done claim was wrong because the wrong surface was checked, or when the verify-before-done hook blocks for lack of evidence. Do not use to review a diff (gate) or to record a case (correct).
Assessment and security
- Files in this package
- 16
- Scripts
- 10
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 16.10.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
- 10 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: assets/verify/check.sh, assets/verify/doctor.sh, assets/verify/drive/example.sh, assets/verify/fresh.sh, assets/verify/lib.sh, assets/verify/observe.sh, scripts/scaffold.mjs, scripts/test-fresh.sh, scripts/test-observe.sh, scripts/test-scaffold-check.sh
What it does and when to use it
Give a repository its own executable verification kit under .agents/verify: doctor, deterministic checks, per-feature drivers that observe the real surface and keep evidence tied to the exact tree, and an enforcer table. Use when a repo has no way for an agent to prove a change works, when setting up verification for a new repo, when adding a driver for a feature, when a done claim was wrong because the wrong surface was checked, or when the verify-before-done hook blocks for lack of evidence. Do not use to review a diff (gate) or to record a case (correct).
Original author description
Give a repository its own executable verification kit under .agents/verify: doctor, deterministic checks, per-feature drivers that observe the real surface and keep evidence tied to the exact tree, and an enforcer table. Use when a repo has no way for an agent to prove a change works, when setting up verification for a new repo, when adding a driver for a feature, when a done claim was wrong because the wrong surface was checked, or when the verify-before-done hook blocks for lack of evidence. Do not use to review a diff (gate) or to record a case (correct).
How it works
- Kit
- Procedure
- Rules
- Maintenance
SKILL.md preview
--- name: verify-setup description: "Give a repository its own executable verification kit under .agents/verify: doctor, deterministic checks, per-feature drivers that observe the real surface and keep evidence tied to the exact tree, and an enforcer table. Use when a repo has no way for an agent to prove a change works, when setting up verification for a new repo, when adding a driver for a feature, when a done claim was wrong because the wrong surface was checked, or when the verify-before-done hook blocks for lack of evidence. Do not use to review a diff (gate) or to record a case (correct)." compatibility: "Requires git, bash, and Node or Bun for the scaffold. Drivers use whatever the surface needs, such as agent-browser, curl, or the repo's own CLI." --- # Verify setup A repo is verifiable when an agent that has never seen it can run one command and get evidence, from the surface a user touches, that a change works. Prose that says how to check is not enough. This skill ships scripts. ## Kit `node scripts/scaffold.mjs <repo-root> [--local]` copies the kit into `<repo>/.agents/verify/` and never overwrites. `--local` adds `.agents/verify/` to `.git/info/exclude`, so the kit works on this machine without ever being versioned. Use it for repos you do not own: | File | Contract | |---|---| | `doctor.sh` | exit 0 only when every precondition the drivers need exists | | `check.sh` | types, lint, unit tests; non-zero on any failure | | `observe.sh <feature> [--runs N]` | runs `drive/<feature>.sh` with `EVIDENCE_DIR` set, N times when asked; passes only if every run passes; `meta.json` records tree digest, exit, runs, passed. `observe.sh check` runs `check.sh` | | `fresh.sh` | exit 0 when every contract feature passed on its latest observation of the current tree contents; `observe: f x5` also needs at least 5 runs; the Stop hook calls it | | `lib.sh` `artifact_current <artifact> <src>...` | fails when a built artifact is older than its sources, so a driver never observes another build | | `drive/<feature>.sh` | reaches the feature as a user does and writes proof into `EVIDENCE_DIR` | | `features/<feature>.md` | where it lives, how a user reaches it, how to drive it, what usually lies | | `ENFORCERS.md` | rule to enforcer table | `evidence/` is gitignored. The tree digest covers tracked changes and untracked files, so evidence from before an edit never counts after it. ## Procedure 1. Read the repo's README, CI config, and agent instructions. List the surfaces: CLI, HTTP API, web UI, desktop, library. 2. Scaffold the kit. Fill `doctor.sh` with the real binaries, env vars and services. Fill `check.sh` with the commands CI runs, fastest first. 3. Pick the three to five features changes touch most often (git log, open issues). For each, write the driver and the feature file. 4. A driver observes the outcome, not a proxy. When it loads a built artifact (binary, WASM, bundle), it first proves the artifact came from this tree: `artifact_current`, or rebuild and compare with the committed copy. A benchmark driver measures the baseline in the same run (origin/main and this tree, same browser) and is declared with `x<N>` in the contract, since one run of a noisy number is a coin flip. A CLI driver runs the built binary and asserts on output. A web driver uses agent-browser to screenshot and asserts on the DOM or text. An API driver curls the running service and asserts on the body, not only the status. 5. Fill "What usually lies" from real failures: cases in the canonical repo under `cases/<repo>/`, review findings, and papercuts. 6. Prove the kit before handing it over: `doctor.sh`, `check.sh`, then `observe.sh <feature>` for every driver, on a clean tree. Open each screenshot and look at it. Then break the feature on purpose (or revert a known fix) and confirm the driver fails. A driver that cannot fail is not a driver. 7. Report the commands run, exit codes, and evidence paths. ## Rules - Never mark a driver done because it exited 0. It is done when it has been seen to fail on a broken state and pass on a fixed one. - Drivers are deterministic and idempotent. They clean up what they start, also on failure. - Do not mock the surface. If the real surface needs credentials or a service that is unavailable, `doctor.sh` fails with that reason; never fall back to a fake. - Label each claim in reports as measured, inferred, or guess. - In a repo you do not own, scaffold with `--local`. Proposing the kit upstream is a human decision. ## Maintenance Run `node scripts/scaffold.mjs <repo-root> --check` before relying on a kit: it fails when `lib.sh`, `observe.sh`, or `fresh.sh` differ from this skill's copies. Copy the current ones over; repo-specific files are left alone. When a change adds or moves a feature, update its driver and feature file in the same diff. When a mistake escapes, add a check that fails on that mistake and a row in `ENFORCERS.md`.
Files in this package (16)
- SKILL.md 5 KB
- assets/verify/.gitignore 1 KB
- assets/verify/ENFORCERS.md 1 KB
- assets/verify/check.sh 1 KB
- assets/verify/doctor.sh 1 KB
- assets/verify/fresh.sh 2 KB
- assets/verify/lib.sh 1 KB
- assets/verify/observe.sh 2 KB
- assets/verify/drive/example.sh 1 KB
- assets/verify/features/README.md 1 KB
- evals/evals.json 2 KB
- evals/triggers.json 1 KB
- scripts/scaffold.mjs 2 KB
- scripts/test-fresh.sh 2 KB
- scripts/test-observe.sh 3 KB
- scripts/test-scaffold-check.sh 1 KB
Versions
The database has no other skills named verify-setup.
Database snapshot: 13.11.2026 · package version dbc03104c5b9 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100setup
Set up Rung's verification kit in a repository: install the CLI, record the project's linters, tests and security scans as checks, gate pull requests in CI, and teach coding agents to verify their work. Use when asked to "set up rung", "add rung to this repo", "verify this repo… - 95rung-setup
Set up Rung's verification kit in a repository so Hermes can prove its code changes work: record the project's tests, linters and security scans as checks, run them, and report what they found. Use when asked to "set up rung", "verify this repo with rung", or when a repository has no rung.toml yet. - 100create-verification-skill
Generate a project-local verification skill (.omp/skills/verify-<app>/) that drives the app the way a user does and captures proof: control CLI, launch/doctor/drive/evidence/cleanup, and a feature map. Any language or platform. Use for /skill:create-verification-skill, \"make a… - 86verify-create
Build a project's own verify skill that launches the real app, drives each user-facing feature in a real browser, and keeps a feature map. Use when a project has a UI or front end and no scripted way to prove it works, or when the user asks for a verification skill. - 100verify-changes
Verification toolkit for agent-hq. Use when validating a change before save, before merge, or when a general verify step defers to the project. A decision matrix maps the shape of the diff to the relevant phases — pick the minimum set that exercises the actual risk surface. Not a fixed sequence. - 100setup-kit
Set up or update AI-agent files (optionally with a minimal, team or strict profile), agent guards, lint/CI checks and repository settings in the current project, in any language. Use when the user runs /setup-kit, asks to "set up the kit", "add AGENTS.md and checks", "configure…