Acceptance

Advisory gate — findings are reported, user decides how to proceed.

Step 1: Resolve Target

• Argument → TaskGet(taskId).
  • Epic (metadata.type == "epic") → epic mode (validate all descendants)
  • Non-epic task → task mode (validate this task only)
• No argument → TaskList() filtered by metadata.type == "epic" AND status in ["completed", "in_progress"]. Pick highest ID → epic mode.
• None found → stop: "No task or epic found. Pass a task ID."

Step 2: Gather Plan Context

Epic mode: From TaskGet(epicId): extract metadata.design. If metadata.plan_file set → ct plan latest --task-file <plan_file>. Falls back to metadata.design. Not found → "No plan found — checking criteria only."

Task mode: Walk metadata.parent_id to find root epic. Extract metadata.design from root for context. If root has metadata.plan_file → ct plan latest --task-file <plan_file>. Falls back to task's own description.

Step 3: Gather Criteria

Epic mode: Recursively collect all descendants via metadata.parent_id chains. No descendants → stop: "Run /scope first."

Orphaned task detection: flag any task where status == "pending" but parent status == "completed". Advisory warning only — never blocks acceptance. See ${CLAUDE_SKILL_DIR}/references/scenarios.md.

Group criteria by subtree: grouping nodes become section headers, leaves have ## Acceptance Criteria extracted. Missing section → ⚠ No acceptance criteria defined. Format:

Phase N: <phase subject>
  Task <id>: <leaf subject>
  - [ ] criterion 1

Flat epics (no grandchildren) produce flat Task <id>: ... output.

Task mode: Extract ## Acceptance Criteria from the target task's description. Missing → use full description as criteria context.

Step 4: Get Diff

git diff HEAD + git diff --cached in parallel. Both empty → git diff HEAD~1..HEAD. Still empty → verify against HEAD state.

Step 5: Spawn Verifier and Breaker

Two parallel Task(subagent_type="general-purpose") agents:

Verifier evaluates each criterion against the diff: PASS/FAIL/PARTIAL/N/A with line-level evidence. All criteria N/A → PASS with note "no applicable criteria found." Adds "Plan Deviations" section noting justified vs problematic divergences. Ends with one-line verdict.

Breaker assumes the implementation is subtly wrong. Hunts five angles: implied requirements, edge cases, integration points, technically-met-but-incomplete criteria, missing negatives. Each finding rated HIGH/MEDIUM/LOW with related criterion. Must list which angles were checked with at least one concrete observation per angle. "No findings" is only valid after showing analysis of all 5 angles — not as a bare assertion.

Breaker validation (Step 6 pre-check): Before applying reconciliation rules, verify the breaker report contains per-angle analysis. Bare "no findings" without angle-by-angle evidence → re-run breaker with explicit instruction to analyze all 5 angles. Do not auto-pass on a structurally invalid breaker report.

Step 6: Reconcile and Present

Reconciliation rules (applied before presenting):

1. Verifier FAIL → FAIL
2. Verifier PARTIAL → PARTIAL
3. Verifier PASS + breaker HIGH findings on PASS/N/A criteria → PARTIAL (breaker caught what verifier missed)
4. Verifier PASS + no breaker HIGH findings → PASS

Present both reports with clear labels. PASS → concise green summary (max 5 lines, no per-criterion breakdown), done.

PARTIAL/FAIL + --auto → automatically Fix gaps (option 1). After 2 fix iterations with no improvement, Commit anyway (override) (option 2).

PARTIAL/FAIL without --auto → AskUserQuestion with exactly 4 options:

1. Fix gaps → spawn fix agent with FAIL/PARTIAL criteria as scope, re-run from Step 4
2. Commit anyway (override) → note override in findings, proceed to Step 7
3. File as follow-up tasks → TaskCreate per gap (metadata: {type: "bug", priority: "P2", parent_id: epicId}), proceed to Step 7
4. Re-run after manual fixes → wait for user signal, re-run from Step 4

Step 7: Store Findings

ct plan create --prefix "acceptance" --topic "<target subject>" --project "$(git rev-parse --show-toplevel)". TaskUpdate target (epic or task) with acceptance_result: {verdict, criteria_count, verifier_report, breaker_report} and plan_file: "$PLAN_FILE" (omit if empty).