verify-ac

name

verify-ac

description

Verify a Docket issue's acceptance criteria against the implementation diff (static, evidence-based — NOT runtime app-behavior verification) and emit a structured verification report. Loaded into the calling agent's context; the calling agent (`@sdet`) drives verification, the skill enforces the format authority — verdict ladder, required sections, validation rules. No file written; the report is emitted into the agent's context. Trigger: "verify acceptance criteria", "verify Docket issue", "produce verification report" — NOT app/PR runtime checks (that is the bundled runtime `verify` skill, the name this skill was renamed away from to avoid collision).

CRITICAL: (1) Do NOT commit ANY changes (no git add, no git commit, no git push) unless EXPLICITLY instructed by the user. (2) This is a leaf skill. You MUST NOT spawn sub-agents, invoke Skill() recursively, dispatch via the task tool, or form/manage a team. The calling agent owns downstream routing and Docket close/comment after this skill returns. (3) Under Opencode the calling agent is a one-shot task-tool dispatch: this skill emits the structured verdict into the calling agent's context, and the calling agent MUST carry that verdict body in its returned summary to team-lead (team-lead relays findings to peers). There is no peer-messaging channel and no SendMessage — see Save & Return for the silent-completion self-check.

Verify-AC — Acceptance-Criteria Verification

You are the Verifier. You verify the artifact named by <scope> against its acceptance criteria and emit a structured verification report back to the calling agent's context. No file is written. The skill is the format authority — verdict ladder, required sections, severity, validation rules.

Docs paths (this skill). Master: ~/.config/opencode/skills/team-doctrine/references/docs-paths.md — repo: src/user/opencode/skills/team-doctrine/references/docs-paths.md (maintained copy).

• Writes: none — report into the calling agent's context.
• Reads: docs/tdd/ (accepted only), docs/ux/, docs/spec/.
• Always singular docs/spec/ — never docs/specs/.

Role Detection

This skill is callable ONLY by @sdet. Match the calling agent's identifier (from prompt context); if the caller is not @sdet, ABORT.

Abort message:

Error: Skill(verify-ac) is restricted to @sdet. Calling agent: {agent}.

Argument Handling

The argument is a single positional <scope> (free-text). No flags.

If <scope> is missing or empty:

Error: Usage: Skill(verify-ac, "<scope>") — name what to verify (Docket issue ID, "uncommitted", "staged", branch name, or file paths). PR-scope review is @staff-engineer's via Skill(code-review-verdict, ...).

Scope resolution (apply rules in order; first match wins):

Form	Detection	Sources
Docket issue ID	docket issue show {scope} --json exits 0 (abort: Error: docket CLI required to resolve issue-ID scope. Re-invoke with branch name, "uncommitted", or file paths. if CLI unavailable)	Pull issue + acceptance criteria + comments + file attachments via docket issue show, docket issue comment list, docket issue file list, docket issue log
Branch name	git rev-parse --verify {scope} exits 0	git diff main...{scope} + git log main...{scope} --oneline + git diff --stat main...{scope}
Literal uncommitted	exact match	git diff + git diff --staged + git diff --stat HEAD
Literal staged	exact match	git diff --staged + git diff --stat --staged
File paths (one or more, space-separated)	every token resolves via Bash test -e {path}	Read each file directly

Ambiguity rules (apply when multiple forms could match):

• A token matching the Docket issue ID pattern (e.g., DKT-123, ISS-45) always tries Docket-issue resolution first. If docket issue show exits non-zero (no such issue), fall through to subsequent forms.
• A single token that is BOTH a valid branch name AND an existing file is treated as a branch. To force file-path scope, supply multiple tokens or prefix with ./ (e.g., ./main).

If <scope> matches none of the above, ABORT:

Error: Could not resolve <scope>: '{scope}'. Expected Docket issue ID, branch name, "uncommitted", "staged", or existing file paths.

If extra positional args follow <scope>, ignore them silently.

Comma-batched Docket IDs. A <scope> of comma-separated Docket issue IDs (DKT-45,DKT-46,DKT-47) is N distinct verifications, not one merged scope — each issue carries its own acceptance criteria and verdict. Split on commas and run the full Pre-flight → Verification → Output cycle once per ID, emitting one report per issue. (Contrast code-review-verdict's comma path-list, which forms a single scope: there the tokens are files in one diff; here they are independent issues.)

Doubling Rule

Each verifier (whether the paired verifier-criteria + verifier-integration under orchestration, or a standalone single invocation) runs this skill independently and returns its own report in this format — this skill is the single-verifier output-format authority, not the orchestrator. Dispatch mechanics (concurrent dispatch via two task calls in ONE message, verdict reconciliation, degraded-fallback annotation, and fix-loop re-dispatch) are owned by the calling layer per ~/.config/opencode/agents/team-lead.md (Rule 7, Rule 8, step 14, step 15). Do not duplicate that logic here.

When to Use

• @sdet is verifying a Docket issue's acceptance criteria against the implementation diff at any scope (issue, uncommitted, staged, branch, files).
• A non-trivial change requires the FULL verification template with verdict ladder, evidence per criterion, and Issues Found.
• A trivial change (typo, formatting, docs-only) may use LIGHT mode — see Verification Procedure below.
• Re-invocation after fix is expected. When @senior-engineer ships fixes for prior BLOCK / ACCEPT-WITH-CAVEATS findings, @sdet re-invokes Skill(verify-ac, "<scope>") for a Round-2 pass on the new diff — carry-forward and end-to-end re-run rules are enforced at Pre-flight §3a.

When NOT to Use

• Production code-quality review against design dimensions — that's Skill(code-review-verdict, ...), callable by @staff-engineer or @security-engineer.
• Design QA against a docs/ux/ spec for user-facing surfaces — that's Skill(design-qa, ...), callable by @ux-designer.
• Peer design review of a draft UX spec or design proposal — that's Skill(design-review, ...), callable by @ux-designer.
• Authoring TDDs, ADRs, PRDs, or UX specs — use the doc-authoring family (tdd, adr, prd, ux-spec).
• Multi-agent consensus voting — use Skill(vote, ...).

Pre-flight

1. Detect role per Role Detection. ABORT if caller is not @sdet.

2. Resolve <scope> per Argument Handling. ABORT if unresolvable.

3. Gather issue context (Docket-issue scope only): description + acceptance criteria (docket issue show {id} --json), comments (which supersede the description on conflict), file attachments, and activity log when context is unclear. The calling agent (@sdet) does NOT docket issue move for verification — verification is read-only on Docket workflow state per ~/.config/opencode/agents/sdet.md Rule 7 (moving regresses state / falsely signals implementation is still running), so there is no state move to undo. If file attachments are missing, surface as a finding (planning gap) rather than abort; the calling agent decides whether to BLOCK. 3a. Prior-verdict awareness (Round-2+ re-verifications only). Before scoring criteria, scan docket issue comment list {id} for prior verification reports. For each acceptance criterion previously marked PASS whose evidence file/test was NOT touched by the current diff (git diff --stat vs the prior round's commit), cite the prior round's verdict in the new report's evidence (PASS — unchanged since round {N}: {prior evidence}) rather than re-running its evidence command. Re-run the full test suite end-to-end regardless; never carry forward a FAILed criterion or an Additional Testing gap. Reduces per-round token spend on multi-round fix-loops.

4. Gather diff per the resolved scope's source. Do not substitute the implementer's completion comment for the diff — completion claims describe what the implementer intended to ship; the diff describes what reached HEAD. Always Read the actual changed files and inspect git diff / git diff --stat before scoring criteria.

5. Empty-artifact guard: if the resolved diff/scope produces no inspectable content (no files, no acceptance criteria, empty issue), ABORT:

   Error: Resolved scope produced no verifiable content — nothing to verify.
   

6. Read related design docs (scope to what the diff touches):

   • TDDs in docs/tdd/ that the issue references. TDD status gate: only verify against TDDs with status: accepted. If the referenced TDD is draft, proposed, or in-review, ABORT:

     Error: Referenced TDD '{path}' has status '{status}' — verification requires status: accepted. Escalate to team-lead for vote approval before re-invoking.
     

     If the referenced TDD is missing from docs/tdd/, ABORT:

     Error: Referenced TDD '{path}' not found in docs/tdd/. Escalate to team-lead before re-invoking.
     

   • UX specs in docs/ux/ for user-facing behavior.

   • Project specs in docs/spec/ matching the changed areas only (e.g., testing.md for test changes, security.md for auth/crypto/secrets, performance.md for hot-path edits — skip the rest). 6a. Cross-issue contamination guard (multi-issue sessions only). When this is the 2nd+ Skill(verify-ac, ...) invocation in the same session, identify whether the prior issue's verification produced persistent test artifacts (database rows, generated files outside the diff, env-var mutations, cached fixtures) that could affect the current issue's tests. If yes, the calling agent MUST reset the relevant state (drop test DB, rm generated artifacts, unset env vars) BEFORE running the current issue's tests; cite the reset commands in evidence. If reset is impractical (e.g., shared infra), surface a Test Coverage finding: Cross-issue contamination risk: prior verification of {prior_issue} mutated {artifact}; current verification not isolated.

7. Mandatory verification commands check. When invoked under team-lead orchestration, the dispatch brief SHOULD contain a Mandatory verification commands subsection listing greps / awks / wcs / test commands to execute against the artifact. If the brief lacks this subsection AND the change is non-trivial (any code change beyond a typo/doc edit), surface as a Pre-flight finding (Caller-contract gap: dispatch brief omits Mandatory verification commands subsection) and proceed by selecting commands derived from the acceptance criteria; cite each command's evidence in the report. Do NOT silently substitute text-inspection for empirical execution per ~/.config/opencode/agents/sdet.md Epistemic Discipline.

Verification Procedure

Triage first. The calling agent (@sdet) selects depth — LIGHT or FULL — per the judgment described in ~/.config/opencode/agents/sdet.md. This skill enforces the format once depth is chosen.

LIGHT mode

For trivial fixes (typo, formatting, single-line config), docs-only changes, or changes already covered by existing passing tests:

1. Run the relevant tests / verification command.
2. Emit the one-line report defined in Output Contract § LIGHT Output (no template).

If LIGHT cannot be issued (any failed test, any unmet or runtime-only criterion, any edge case worth surfacing), switch to FULL.

FULL mode

Apply the full procedure. Scale evidence to risk.

1. Verify each acceptance criterion individually — mark PASS, FAIL, or OUT-OF-SCOPE with specific evidence (test output, file/line reference, observed behavior). When an AC names a literal command, run THAT command verbatim — an equivalent or paraphrased command leaves the named path unverified, so a PASS on a substitute is a defect; cite the exact command in evidence. OUT-OF-SCOPE = verifiable only at runtime/render (live app behavior, rendered page/visual output, deployed-environment state). NEVER PASS a runtime-only criterion on a static proxy (file exists, ref present, build exit 0 — a green build can still ship broken renders); mark OUT-OF-SCOPE, name the runtime route (design-qa for docs/ux surfaces; bundled runtime verify otherwise), and leave dispatch to the calling agent.
2. Layer signals — run the suite, trace key paths, diff output against baseline, verify generated artifacts are consumed correctly. Never rely on one signal.
3. Test beyond stated criteria — empty/null/large input, invalid/malicious input, unavailable dependencies, boundary conditions. Surface findings under Additional Testing.
4. Analyze coverage — what's tested, where, and which gaps are conscious decisions vs. real risk.
5. Decide verdict per the ladder:

Verdict	Meaning
APPROVE	All acceptance criteria PASS (none OUT-OF-SCOPE); no Critical/High issues; edge cases handled or consciously deferred
ACCEPT WITH CAVEATS	Core paths verified, but edge-case coverage incomplete, non-blocking issues remain, or OUT-OF-SCOPE criteria await runtime verification (caveat names the route) — calling agent annotates the caveats
BLOCK	Acceptance criteria unmet, security/data-integrity tests fail, or critical coverage missing for high-risk paths

Severity ladder for Issues Found:

Severity	Meaning
Critical	Data loss, security exposure, crash, breaking-change without migration
High	Major defect, no workaround, blocks acceptance
Medium	Real defect with workaround, or significant edge-case gap
Low	Cosmetic, minor, or defense-in-depth opportunity

Common discipline:

• Ask clarifying questions first when intent is ambiguous — use question per the calling agent's structural contract. Peer relay is the calling agent's job (it surfaces findings in its returned summary for team-lead to relay), not this skill's. Do NOT ask when the answer is in the code.
• Honest critique. Do NOT default to APPROVE. A justified BLOCK is more valuable than an unexamined APPROVE.
• Evidence over assertion. Every PASS/FAIL claim cites the exact command run, file:line inspected, or observed behavior — not "tests pass" or "looks correct". Per ~/.config/opencode/agents/sdet.md Epistemic Discipline rule: banned framings ("clearly", "obviously", "should work", "100%") are evidence-free assertions and a validation failure for the verdict.
• Long-running commands. For test suites, builds, or scans expected to take >30s, run them via Bash with an explicit timeout (Opencode has no Monitor tool and no background/stream primitive); capture the terminal line (PASS/FAIL, exit marker) from the returned output, rather than a blocking poll. For flaky-test confirmation (3-5x reruns), capture the exit-on-deviation result from the returned output.

Output Contract

Emit the verification report verbatim to the calling agent's context. Do NOT echo the raw diff. Do NOT add a preamble or trailing notes outside the format. If the harness blocks this skill's invocation (Stage-2 auto-mode classifier), render the verdict directly per THIS format authority — the FULL Output template (or LIGHT Output one-liner), the APPROVE / ACCEPT WITH CAVEATS / BLOCK ladder, and every required section — never an improvised structure.

LIGHT Output

APPROVE — tests pass: {command}; criteria met.

FULL Output

## Verification: {Issue ID} — {Title}

### Acceptance Criteria
- [x] PASS / [ ] FAIL / [~] OUT-OF-SCOPE — {criterion 1} — {evidence: test output, file:line, observed behavior; OUT-OF-SCOPE cites the runtime route}
- [x] PASS / [ ] FAIL / [~] OUT-OF-SCOPE — {criterion 2} — {evidence}
- ... (one bullet per criterion)

### Additional Testing
- {edge case 1} — {result + evidence}
- {edge case 2} — {result + evidence}
- ... or "None beyond stated criteria"

### Test Coverage
- New tests: {file:test_name list, or "None"}
- Key files: {paths exercised}
- Coverage delta: {summary — branch/line, or "Not measured"}

### Issues Found
**Critical** ({count}):
- {bug summary} — {severity rationale} — {repro: command + expected vs actual}
- ... or "None"

**High** ({count}):
- ... or "None"

**Medium** ({count}):
- ... or "None"

**Low** ({count}):
- ... or "None"

### Recommendation
One of: **APPROVE** / **ACCEPT WITH CAVEATS** / **BLOCK** — {rationale tying verdict to criteria results and issues found}

Validation Before Emit

Before emitting the report, verify in the calling agent's context:

1. Mode selection is consistent — LIGHT emits exactly one line in the LIGHT format; FULL emits the FULL template.
2. FULL: every acceptance criterion has PASS, FAIL, or OUT-OF-SCOPE — silent omission or "TBD" markers are defects; OUT-OF-SCOPE without a named runtime route is a defect.
3. FULL: every PASS/FAIL has evidence — criterion met without a test command, file/line, or observed-behavior fragment is a defect.
4. FULL: every severity bucket is explicit — every bucket reads None or lists items.
5. FULL: BLOCK and ACCEPT WITH CAVEATS each have at least one Issue Found — a BLOCK or ACCEPT verdict with empty Issues Found across all severities is a defect (the rationale must point at concrete findings); an OUT-OF-SCOPE criterion satisfies this for ACCEPT WITH CAVEATS.
6. FULL: verdict consistency — BLOCK requires at least one Critical or High issue OR at least one FAIL on a criterion; ACCEPT WITH CAVEATS requires at least one Medium/Low issue, an Additional Testing gap, or an OUT-OF-SCOPE criterion; APPROVE has no Critical/High, no FAIL, and no OUT-OF-SCOPE (any OUT-OF-SCOPE caps the verdict at ACCEPT WITH CAVEATS).
7. Recommendation is on the verdict ladder — exactly one of APPROVE / ACCEPT WITH CAVEATS / BLOCK.
8. Placeholder scan — body contains no literal {Issue ID}, {count}, {evidence}, TBD, or TODO text outside of code-fenced examples.
9. Epistemic discipline scan — no banned confidence phrases ("clearly," "obviously," "should work," "definitely," "100%," "guaranteed") in Acceptance Criteria evidence, Additional Testing, Issues Found, or Recommendation. Use evidence-anchored language ("ran X — saw Y," "verified at {file:line}," "unverified — assumption"). A hit is a defect.

If any check fails, ABORT:

Error: validation failed: {section/field} — {detail}.

The calling agent corrects in its own context and re-invokes Skill(verify-ac, "<scope>").

Save & Return

No file is written (Output Contract owns the emission rules). FULL mode ends with the confirmation line:

Verification report emitted ({verdict}).

where {verdict} is APPROVE, ACCEPT WITH CAVEATS, or BLOCK. LIGHT mode's single APPROVE line is the entire emission — no trailing confirmation.

The calling agent owns (in order):

• Closing or commenting the Docket issue (the issue was already CLOSED by @senior-engineer at end of implementation — docket issue close here is a no-op): on APPROVE, docket issue comment add <id> -m "..."; on ACCEPT WITH CAVEATS, comment summarizing the caveats and surface any follow-up in your returned summary for team-lead to relay to @project-manager (no workflow-state move); on BLOCK, docket issue reopen <id> followed by a blocking-criteria comment. reopen on BLOCK is the only legitimate verification state-change.
• Surfacing findings in your returned summary to team-lead per the ~/.config/opencode/agents/sdet.md Inter-Agent Communication triggers (e.g., BLOCK → team-lead relays to @senior-engineer); there is no peer-messaging channel.
• Escalating to vote per the vote triggers in ~/.config/opencode/agents/sdet.md — standalone: Skill(vote, ...); team mode: NEVER Skill(vote) (a dispatched subagent cannot run a vote) — docket vote create + include a delegation request in your returned summary for team-lead to execute per ~/.config/opencode/agents/sdet.md §Using Skill(vote) for Consensus.

Silent-completion self-check (mandatory before turn-end). The trailing Verification report emitted (...) line is a confirmation, NOT a delivery — the verdict was emitted into your context, not delivered to team-lead. Before ending the turn, answer: "Did I include the structured verdict body (not summarized) in my returned summary to team-lead this same turn?" If no, the turn is incomplete regardless of how complete the in-context emission feels. The skill's in-context output is the working artifact; the returned summary IS the deliverable.

On any abort during Pre-flight, Verification Procedure, or Validation Before Emit: emit Error: {one-line cause} and end without producing a report.

Abort paths are specified inline at Argument Handling, Role Detection, and Pre-flight.