| name | change-verification |
| description | Use for code-change verification: diff review, smoke tests, regression checks, and pre-merge confidence on modified code. Trigger on prompts like 'check my changes', 'review this diff', 'verify this fix', or any request to validate recent edits instead of implementing new code. |
Change Verification
Verify changed code with high signal, low waste. Start from actual diff, run smallest checks that can prove or disprove safety, then report findings and remaining risk.
Workflow
- Inspect worktree first with
git status and relevant diff.
- Read changed files and nearby tests before choosing commands.
- Infer blast radius: docs-only, local behavior, shared abstraction, or public API.
- Run smallest useful validation first; widen only if risk is broader.
- Report findings first, then commands run and residual risk.
Rules
- Never guess what changed. Read diff first.
- Ignore unrelated dirty files unless they affect verification.
- For review requests, prioritize bugs, regressions, and missing tests over summary.
- If nothing fails, say
No findings. and still note untested risk.
- Do not modify code during pure verification unless user asked for fixes too.
Validation Ladder
Choose lowest rung that gives real confidence.
- Inspect only: docs, comments, tiny non-executable changes.
- Targeted checks: default for most edits. Prefer touched tests, path-scoped lint, narrow type checks.
- Broad checks: use for shared abstractions, config, public API, or prompt/cache/judge/provider flow.
IntentGuard Notes
Keep normal work focused on runtime package unless user asks otherwise.
- Main code:
intentguard/
- Main tests:
tests/
- Usually out of scope:
validation/, ai_research/
High-risk changes:
- Public API surface:
IntentGuard, IntentGuardOptions
- Runtime flow across prompt factory, cache, inference provider, and judge
Judge uses strict majority, so ties must evaluate false- Cache keys and cache versioning must stay compatible
Useful commands when scope fits:
uv run ruff check <changed-paths>
uv run ruff format --check <changed-paths>
uv run mypy intentguard
uv run python -m unittest discover tests
make test
Reporting
Keep response short, concrete, evidence-based.
For review tasks, use:
Findings:
1. [severity] `path:line` - issue and impact
Open Questions:
1. ...
Verification:
1. `command` - result
Residual Risk:
1. ...
If no issues found:
No findings.
Verification:
1. `command` - result
Residual Risk:
1. ...
