Selective integration verification driven by git commits or working-tree changes: analyze the impact surface, prepare the environment, verify changed behavior through real pages, and record reusable module knowledge and per-run evidence separately. P0 smoke checks and change acceptance are tracked independently, so necessary scenarios are never skipped just because a commit baseline already exists.
Trigger it with requests like "test the recent commits/changes", "run integration tests", "verify this batch of commits / this change", or "regression check". It is not for writing unit tests or pure code review. Requires a host that supports skills (ZCode / Codex / OpenCode) and an available browser tool.
flowchart LR
A["1 Scope<br/>commits / worktree / goal"] --> B["2 Impact analysis<br/>match modules & call chains"]
B --> C["3 Environment<br/>verify version, reuse or start services"]
C --> D["4 Execute<br/>P0 smoke + change acceptance scenarios"]
D --> E["5 Report & capture<br/>evidence to runs, stable knowledge to modules"]
E -. "baseline advances by coverage" .-> A
| Phase | What happens | Details |
|---|---|---|
| 1 Scope | Use the commits/worktree/goal the user stated; list P0 smoke and change acceptance separately | analysis.md |
| 2 Impact analysis | Match modules by each repo's final diff; trace services, APIs, and page entries on a miss | analysis.md |
| 3 Environment | Verify environment and target version; prefer reusing already-running matching services | env.md |
| 4 Execute | Real-page scenarios plus evidence matching the acceptance goal (screenshots, requests, data checks) | execute.md |
| 5 Report & capture | Run evidence goes to runs/; only verified stable knowledge updates modules/ | execute.md |
- Knowledge and evidence stay separate: steps and past results in modules do not prove this run; every run is recorded independently.
- Baselines advance by coverage: a P0 pass at the same HEAD cannot skip pending change-acceptance scenarios; anything uncovered means the batch is not accepted.
- Real-page evidence: builds, API checks, and read-only database queries can supplement evidence but never stand in for having walked the real page.
- Honest records: failures, blocks, degradations, and skipped steps are all stated; missing evidence is marked as not retained, never fabricated.
SKILL.md Entry point: workflow, capabilities, and authorization boundaries (progressive loading)
phases/ Details: analysis (scope & impact) / env (environment) / execute (browser execution)
TEMPLATE.md Template for module knowledge files
LICENSE MIT
projects/ Local accumulation area (git-ignored, not synced)
└── <project>/
├── env.md Linked repos, start commands, addresses, credential sources, and environment limits
├── tested-commits.md Per-repo versions with scope/depth/scenario coverage and relevant dirty state
├── modules/*.md Stable knowledge: trigger scope, call chains, entries, scenarios, and evidence criteria
└── runs/*.md Actual operations, results, evidence, and uncovered items for each run
This repository is the skill source directory. Clone it and link it into your host's skills directory:
| Host | Skills directory |
|---|---|
| ZCode | ~/.zcode/skills/ |
| Codex | ~/.codex/skills/ |
| OpenCode | ~/.config/opencode/skills/ |
git clone https://github.com/ThinkDonk/commit-test.git
# Windows (junction)
mklink /J "%USERPROFILE%\.zcode\skills\commit-test" <clone-path>\commit-test
# macOS / Linux (symlink)
ln -s "$PWD/commit-test" ~/.zcode/skills/commit-testOn the first run, the knowledge directory for a project is initialized on the spot under projects/ (env, modules, baselines, and run records).
- In git: SKILL.md, phases/, TEMPLATE.md, and both READMEs (methodology, synced across machines)
- Not in git:
projects/(local environment, baselines, module knowledge, and run evidence; excluded by.gitignore). An ignored directory is not a credential store — reports never copy passwords or tokens.
The browser layer uses whatever tooling the current host actually provides; no specific auxiliary skill is required. Detailed execution rules, result evidence, and baseline updates are documented in SKILL.md and its phase files.
Issues and suggestions are welcome.