verification-before-completion

name

verification-before-completion

description

Before claiming work is complete, run tea-rags imports/churn signals on just-edited files to surface collateral-damage candidates (files whose dependents could break), then run verification commands so success claims are backed by evidence and blast-radius awareness. Triggers on "ready to commit", "all done", "should be fixed now", "готово", "проверь перед commit", "все работает". NOT for mid-work sanity checks without a completion claim. Wraps superpowers:verification-before-completion with a tea-rags collateral-damage scan.

dinopowers: verification-before-completion

Wrapper over superpowers:verification-before-completion. Adds collateral-damage scan: before claiming complete, query tea-rags blast-radius signals (imports count, churn, ownership) on just-edited files; surface which dependents to verify too.

Iron Rule

tea-rags collateral-damage scan MUST run BEFORE claiming "done" — whenever session edited ≥1 existing file.

Correct tool (semantic_search), correct impact rerank ("blastRadius" when codegraph on, {imports 0.5, churn 0.3, ownership 0.2} fallback when off), correct params (brace-expanded pathPattern over git diff --name-only, metaOnly: true), correct verdict (surface high-fanIn/imports files with explicit "verify dependents" rec) = core value.

Only new files created (no edits to existing): skip scan, verdict SAFE (no existing-file edits). Don't fabricate pathPattern.

Verdict Ladder (PRESCRIPTIVE — apply before claiming done)

🛑 STOP — read the ladder, then act.

Read blast signal from fanIn when codegraph on (real dependents), else imports proxy.

Verdict	Triggers
HIGH-BLAST	fanIn/imports top 10% of result set (or absolute: >20 dependents)
MEDIUM-BLAST	fanIn/imports top 30% (or 5-20 dependents)
LOW-BLAST	fanIn/imports ≤ 5 dependents

Block is prescriptive, not informational — DO NOT skip ladder evaluation.

Chaining rule: see CHAINING.md — every dinopowers:X redirects superpowers:X. NEVER bypass the wrapper.

Index freshness: see FRESHNESS.md and tea-rags/rules/index-freshness.md. No background reindex hook — worktree-plan freshness explicit (clone + per-task reindex in dinopowers:executing-plans); run mcp__tea-rags__index_codebase manually to search code edited but not committed, BEFORE first tea-rags call.

Step 1 — Collect edited file set

From git status --short or git diff --name-only, collect:

Source	Example
Modified files (M) in working tree	src/core/domains/explore/reranker.ts
Staged files (A/M in index)	tests/explore/reranker.test.ts
Files changed in session (uncommitted)	any file touched by Edit/Write/MultiEdit

Output:

• editedFiles: relative paths with actual content changes (exclude pure renames, new-only files)
• intent: one sentence — what session changed

editedFiles empty (pure new-file session): skip to Step 4, verdict SAFE (no existing-file edits).

Step 2 — Collateral-damage scan call

Issue ONE mcp__tea-rags__semantic_search call — SAME idiom as dinopowers:writing-plans and dinopowers:executing-plans:

project:     <alias from list_projects — RECOMMENDED, omit path when set>
path:        <current project path — fallback when no alias is registered>
query:       <intent from Step 1>
pathPattern: "{editedFile1,editedFile2,...}"   ← brace expansion
rerank:      "blastRadius"               ← codegraph on; OFF fallback below
limit:       <editedFiles.length * 3>
metaOnly:    true

Codegraph gating for rerank: "blastRadius" (real fanIn) when prime ## Enrichment lists codegraph.symbols; else fall back { custom: { imports: 0.5, churn: 0.3, ownership: 0.2 } } (import-proxy).

Do NOT substitute:

Wrong tool	Why wrong
mcp__tea-rags__hybrid_search	Custom impact rerank tied to semantic_search
Named preset "hotspots" / "codeReview" / "impactAnalysis"	impactAnalysis does not exist; these miss the blast-radius dimension. "blastRadius" IS correct when codegraph is on
mcp__tree-sitter__trace_impact	Structural trace is complementary (call graph), not blast-radius on git signals; use tea-rags first
One call per file	Brace expansion covers all
Built-in grep -r "import.*<file>"	Misses the ranked imports overlay signal; slower
git log --oneline <files>	Shows history of edited files, not their dependents

Do NOT pass:

• metaOnly: false — verdict inputs are signals, not content
• Different weights — must match dinopowers:writing-plans / executing-plans / tea-rags:data-driven-generation Step 6 for cross-skill comparability
• filter narrowing — pathPattern already scopes

Results empty (files brand-new committed this session, or not in git yet): skip to Step 4, verdict UNVERIFIABLE (edited files not indexed). Don't fabricate collateral-damage signals.

Step 3 — Compute collateral-damage verdict

For each unique relativePath in results, extract from payload.git.file.* + overlay:

• imports score — how many modules import this file (blast radius)
• commitCount — churn indicator
• blameDominantAuthorPct (with adaptive label) — live-line silo risk (use label, not magic percentage — silo / deep-silo are codebase-relative)

Verdict ladder per edited file: See Verdict Ladder near top.

Aggregate:

### dinopowers collateral-damage scan

| Edited file | imports | churn | verdict |
|---|---|---|---|
| src/core/contracts/errors.ts | 47 | 23 commits | HIGH-BLAST |
| src/adapters/qdrant/client.ts | 12 | 8 commits | MEDIUM-BLAST |
| tests/explore/strategies.test.ts | 0 | 2 commits | LOW-BLAST |

**Verify before claiming done:**
- `errors.ts` imported by 47 modules → run the project's test suite covering all dependents
- `client.ts` imported by 12 modules → run integration tests touching the qdrant adapter
- `strategies.test.ts` → no dependents; the file's own tests are sufficient

Phrasing stays generic — never name specific runners (vitest, jest, pytest, rspec, etc.). Reading agent resolves project's actual test command from project context (package.json scripts, Makefile, CI config).

Step 3a — Tests-at-risk lookup (targeted verification)

Invoke Skill(tea-rags:tests-as-context) with:

recipe: "tests-at-risk"
affectedFiles: <editedFiles from Step 1>
intent: <intent from Step 1>

Recipe surfaces DSL leaf test chunks semantically bound to change. For each HIGH-BLAST/MEDIUM-BLAST file whose result list non-empty, augment verdict block with targeted rec:

**Targeted scenarios for high-blast files:**
- src/core/contracts/errors.ts → run the tests for these scenarios:
  - tests/errors/typed-errors.test.ts:42 — TypedError > propagates cause
  - tests/api/handler-errors.test.ts:78 — handler > error response shape

Output stays runner-agnostic. Phrasing: "run the tests for these files", "execute these scenarios via the project's standard test command". Never name specific runner.

Recipe returned SKIP (no DSL test chunks indexed): leave verdict block as-is, add **Targeted scenarios:** unavailable (no DSL test chunks indexed) — verification scope guided by blast-radius signals only.

Recipe returned empty list (preflight passed, no semantic match): add **Targeted scenarios:** no obvious test bindings — run general verification covering blast-radius dependents.

Step 4 — Invoke superpowers:verification-before-completion

Invoke Skill tool with superpowers:verification-before-completion. Prepend collateral-damage block as context. Phrase handoff as:

"Before claiming done, the edited files have these blast-radius signals: …… Run verification commands that exercise HIGH-BLAST dependents, not just files directly edited. Evidence before assertions.

Chaining rule reminder: when your cycle would next invoke superpowers:finishing-a-development-branch (or any wrapped superpowers:Y), invoke dinopowers:Y instead — see the Chaining rule section above."

Let superpowers:verification-before-completion run its standard verification cycle (tests, type-check, lint, build). Wrapper doesn't replace it — informs scope of verification.

Red Flags — STOP and restart from Step 2

• "Only tests need to run, I know the change is isolated" → if ANY edited file has imports > 0, run Step 2
• "git diff is small (2 lines), skip the scan" → 2-line change in high-blast file = catastrophe. Run Step 2.
• Substituted grep -r / git log → redo with semantic_search + custom rerank
• Named preset instead of custom weights → redo
• Skipped verification after surfacing HIGH-BLAST → wrapper informs scope, never substitutes for verification
• metaOnly: false → restart
• Let superpowers:verification-before-completion chain into a raw superpowers:finishing-a-development-branch without redirecting to dinopowers:finishing-a-development-branch → intercept and invoke the wrapper instead (see Chaining rule)

Common Mistakes

Mistake	Reality
Claim "done" after tests in edited files pass	Dependents may still break. imports > 5 = verify their tests too.
Use rerank: "hotspots" instead of custom impact	hotspots returns BUG-prone zones (history), not BLAST-radius (structural)
Silent downgrade of HIGH-BLAST to "probably fine"	Verdict informs scope; downgrading without evidence hides risk
Run full test suite regardless of verdict	LOW-BLAST files don't need full suite; the scan's point is targeted verification
Skip scan for "trivial" renames	Renames break imports; ALL dependents may need updates
Run scan but ignore block in verification	Block is prescriptive, not informational — it says WHICH tests to run