validate

Prompt file imported from guimatheus92/validate (.github/prompts/validate.prompt.md). Copyright stays with the author.

Validate the recent work in this repository. You may not declare it correct — only evidence may. If this repo is a checkout of the validate plugin's source or ships its skill, read skills/validate/SKILL.md and follow it; otherwise apply this compressed contract:

1. Scope: the full range of the session's work — all its commits plus uncommitted changes (never just HEAD~1) — unless the user gave an explicit range or paths.
2. Declare coverage first: before executing, enumerate the session's claims (from the request, commit messages, and diff — one line each, with the tier that will prove it), state what you WILL validate (per tier) and what you canNOT validate with the reason (no tooling, no credentials, not locally executable — check what parses or dry-runs, declare the rest), and name the nonfunctional dimensions (security, performance, scale, compatibility, reliability, deployment) no claim covers. The final report must match this declaration, claim by claim.
3. Tier 1 — static: typecheck, lint, build, using the project's own scripts/Makefile/CI definitions. Absent = SKIP with reason. On every stack, also run diff hygiene: git diff --check, leftover conflict markers, unexpected or generated files flagged against the stated task, every changed file expected or flagged — and every changed surface reachable: code nothing calls, routes to, or registers is a finding, and runtime must not validate it as live.
4. Tier 2 — tests: run the relevant suites. For a bug fix, prove the regression: the covering test must fail on the pre-fix code (throwaway git worktree at base) and pass on the fixed tree — capture both. For any other test the session added — over pre-existing behavior or a new feature — prove it can fail: first search history for the commit that introduced or fixed the behavior and replay (the new test fails at that commit's predecessor, passes intact); no such boundary → tamper-check: break the covered code's behavior in a throwaway worktree of HEAD; the new test must fail there while unrelated control tests stay green, and pass intact (SKIP with stated reason only when infeasible). Filtered runs count only via runner-reported executions — a filter matching zero tests exits green and proves nothing. Missing or failed proof = Tier 2 FAIL, however green the suite.
5. Tier 3 — runtime: run the real thing and observe the changed flow. CLI → run it and capture output; API → send real requests; web UI → drive a browser if tooling exists, else assert on served HTML and mark purely visual claims SKIP (infeasible), never PASS.
6. Deployed evidence (when applicable): if the work fixes code that predates this branch, the request or commits mention production/incident/telemetry/customers, the surface is a service route, worker, or deploy config, or the user asks whether it happens in real data — after the tiers, check the deployed system too (docs-only, local-only, and never-deployed new behavior are exempt). Discover a telemetry/log/data source (recipe, project docs, the code that emits the operation names); never invent a cluster, table, operation name, or window — ask the user for non-secret coordinates or an explicit waiver instead (they authenticate in their own terminal; an explicit decision already in their request consumes the ask). Derive query identifiers from the emitting code, and read failure status from payload/description fields, not only transport results. Caller first: identify the real deployed producer and read caller-side request telemetry (outgoing/client requests, gateway logs) before interpreting service-side operation rows — reconcile the two sides, classify provenance from a primary signal (an environment column reading production is not one), and when no caller-side source exists, name the gap and proceed on the service side. A zero-row result counts only beside a positive control proving the source live, and reads "NOT OBSERVED in over " — never "never happens" or "no customer impact"; zero eligible operations is "NOT MEASURABLE — 0 eligible target operations", never 0/0. A production fact the user or commits asserted is a claim: zero rows in a fit source, or a fit source that does not support the assertion = that claim FAIL; source unavailable and not waived = BLOCKED — either way an asserted claim's verdict controls the overall one. A dimension you added on your own is a declared gap with a runbook, not a claim — a scoped PASS is allowed. When this phase applies, the verdict line carries a scope: "causal correctness PASS; deployed incidence NOT OBSERVED; customer impact UNPROVEN".
7. Verdicts: PASS / FAIL / BLOCKED / SKIP per claim and per tier, overall verdict first. Evidence = output, exit codes, response bodies, screenshots captured THIS run; memory is not evidence; ambiguous output = FAIL. BLOCKED names exactly what is missing — and never invents it: no fabricated env values, credentials, or stubs to force a check to run. SKIP is always labeled: SKIP (not applicable) / SKIP (infeasible) / SKIP (user-waived) — the last only from the user's own words, quoted; never self-waived.
8. Honesty: never weaken a check to pass it. At most 3 fix-and-rerun attempts, then stop and report FAIL. A rerun spends an attempt when it changes an assertion, or the code under validation, to turn a failing check green; repairing a weak proof spends one too. Relocating a fixture that a change already validated this run made inert spends nothing — provided no assertion is loosened, removed, or skipped, the input keeps its strength, and you declare the relocation with its diagnosis and diff. An undeclared relocation spends an attempt. One self-review batch — the findings of one run of a review, lint, or audit tool driven against this diff as the Tier 3 exercise — counts once, not once per finding. One exception past the ceiling: a declared attempt 4 (consequential), when the last round strictly reduced the red checks, the remaining failure is a diagnosed consequence of a validated change, and only test fixtures/setup or docs are touched. Label it as such and carry its diagnosis and diff; an undeclared fourth round makes the report itself FAIL, and there is no attempt 5. Banned in your conclusions: "should work", "probably", "seems to", "appears to", "likely", "I believe", … (canonical list: skills/validate/reference/evidence.md).

Report: overall verdict, per-tier table (Tier | Verdict | What ran | Evidence), a claims table (Claim | Verdict | Evidence — one row per declared claim), a deployed-evidence status table and a Verdict scope line when that phase ran, a "Not validated" line matching the declared gaps and naming the nonfunctional dimensions no claim covered, a "Fix-and-rerun attempts used" line (always present, even at 0/3) plus a declared line per consequential relocation and for any attempt 4 (consequential), then the evidence appendix with captures quoted.