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. ...
Source: kdunee/intentguard — distributed by TomeVault.
1---2name: kdunee-intentguard-intentguard3description: Change Verification4---56# Change Verification78Verify 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.910## Workflow11121. Inspect worktree first with `git status` and relevant diff.132. Read changed files and nearby tests before choosing commands.143. Infer blast radius: docs-only, local behavior, shared abstraction, or public API.154. Run smallest useful validation first; widen only if risk is broader.165. Report findings first, then commands run and residual risk.1718## Rules1920- Never guess what changed. Read diff first.21- Ignore unrelated dirty files unless they affect verification.22- For review requests, prioritize bugs, regressions, and missing tests over summary.23- If nothing fails, say `No findings.` and still note untested risk.24- Do not modify code during pure verification unless user asked for fixes too.2526## Validation Ladder2728Choose lowest rung that gives real confidence.2930- Inspect only: docs, comments, tiny non-executable changes.31- Targeted checks: default for most edits. Prefer touched tests, path-scoped lint, narrow type checks.32- Broad checks: use for shared abstractions, config, public API, or prompt/cache/judge/provider flow.3334## IntentGuard Notes3536Keep normal work focused on runtime package unless user asks otherwise.3738- Main code: `intentguard/`39- Main tests: `tests/`40- Usually out of scope: `validation/`, `ai_research/`4142High-risk changes:4344- Public API surface: `IntentGuard`, `IntentGuardOptions`45- Runtime flow across prompt factory, cache, inference provider, and judge46- `Judge` uses strict majority, so ties must evaluate false47- Cache keys and cache versioning must stay compatible4849Useful commands when scope fits:5051```bash52uv run ruff check <changed-paths>53uv run ruff format --check <changed-paths>54uv run mypy intentguard55uv run python -m unittest discover tests56make test57```5859## Reporting6061Keep response short, concrete, evidence-based.6263For review tasks, use:6465```markdown66Findings:671. [severity] `path:line` - issue and impact6869Open Questions:701. ...7172Verification:731. `command` - result7475Residual Risk:761. ...77```7879If no issues found:8081```markdown82No findings.8384Verification:851. `command` - result8687Residual Risk:881. ...89```9091---92> Source: [kdunee/intentguard](https://github.com/kdunee/intentguard) — distributed by [TomeVault](https://tomevault.io).93<!-- tomevault:4.0:skill_md:2026-06-18 -->
tomevault-io