Tighten acceptance evidence discipline across agents

[shinpr]

Summary

• Introduce acceptance evidence discipline so test results cited as evidence reflect real behavior — not zero-match selectors, skipped tests, TODO/placeholder bodies, or always-passing assertions. The rule is scoped to test evidence so non-test verification (build, typecheck, CLI) and refactor tasks without test evidence are unaffected.
• Extend the task-executor frontend with sections that previously existed only on the backend variant: Test Environment Check, Reference Representativeness (with a concrete "feature area" definition), and the four-bullet Iron Rule. Add Dependency Version Uncertain (2-4) and Test Environment Not Ready (2-7) escalation schemas, and register test_environment_not_ready in the orchestration guide's allowed escalation_type list so the workflow contract stays consistent.
• Compress prompt files by ~280 lines without losing behavior-binding signals: collapse repeated preambles, deadwood "Specific Boundary" recaps, illustrative bullet lists, verbose JSON examples, and ENFORCEMENT boilerplate; rewrite negative-form instructions in positive form; restore anchor values and disambiguation phrasing where the first pass over-compressed.

What changes per agent

task-executor.md / task-executor-frontend.md

• Add runnableCheck.result field spec and an Exit Gate item that count only test runs cited as evidence; non-test verification is exempt.
• Add Test Environment Check (frontend), Reference Representativeness (frontend), and expand the Iron Rule to four bullets (frontend). Both files now share the same structural sections, while domain vocabulary stays specific to their stack.
• Add Escalation Response 2-4 Dependency Version Uncertain (frontend) and 2-7 Test Environment Not Ready (both files), with structured missingComponent + description fields so the orchestrator can parse the cause.
• Compress: collapse File Scope Constraint to direct rules, fold Similar Function/Component Duplication Check indicators into one inline list with explicit "Exactly the pair (a+c) or (b+c) → Escalation" thresholds, drop deadwood preambles and the redundant Execution Principles footer.
• Add Change Category-style stop conditions to the Iron Rule's escalation paths in positive form: "Route any new library/pattern decision through Escalation Response 2-4" instead of "Do not introduce a new library or pattern".

quality-fixer.md / quality-fixer-frontend.md

• Add a conditional Substance check in Step 3 and an approved criterion that only apply when a test run is cited as evidence for the task's intended behavior.
• Split the blocked JSON example into two per-variant blocks (specification conflict / missing prerequisites) so the LLM does not produce hybrid responses, and keep concrete sample values as anchors ("500 error" / "Button disabled" etc.).
• Frontend: rewrite Frontend-Specific Quality Criteria to be repository-local (prefer dominant local patterns, mock at network/API boundaries, allow browser primitive doubles, exercise the component under test through real renders). The "new library/pattern" rule routes through the agent's actual blocked output contract.
• Compress: drop low-binding Important Principles, generic Debugging Hints, the Fix Continuation Determination Conditions duplicate of Step 5, and a 17-line mermaid diagram that restated text rules.

integration-test-reviewer.md

• Add a substantive-assertion criterion to Quality Assessment and a Hollow or Placeholder Assertion entry to Common Issues, limited to text-inspectable patterns.
• Tighten the JSON schema example while keeping meaningful placeholder text.

code-reviewer.md

• Extend Test Coverage for Acceptance Criteria to flag tests that grep finds but do not exercise the AC (skipped, TODO-only, or always-true assertions), with explicit examples and an absence-verification clarification.
• Compress the schema and example sections; drop the generic Special Considerations block.

skills/subagents-orchestration-guide/SKILL.md

• Add test_environment_not_ready to the task-executor escalation_type list so the new typed escalation has a documented downstream handler.

Line-count impact

File	Before	After	Δ
agents/task-executor.md	444	406	-38
agents/task-executor-frontend.md	395	411	+16
agents/quality-fixer.md	330	230	-100
agents/quality-fixer-frontend.md	440	328	-112
agents/integration-test-reviewer.md	146	137	-9
agents/code-reviewer.md	275	243	-32
agents/ total	2030	1755	-275

All target files now sit at or below 450 lines; task-executor and quality-fixer are both well under the 400-line target.

Test plan

• node scripts/sync-plugins.mjs regenerates the three plugin directories cleanly (handled by the pre-commit hook).
• claude plugin validate passes for marketplace.json and each plugin manifest (handled by the pre-commit hook).
• node scripts/check-skills-index.mjs reports all 11 skills consistent.
• Run a sample task through task-executor and confirm the new Exit Gate item rejects zero-match / skipped / placeholder evidence while leaving non-test verification alone.
• Run a sample frontend refactor task (no test evidence cited) and confirm the conditional substance check in quality-fixer-frontend does not falsely block it.
• Trigger a missing-test-environment condition and confirm the agent returns escalation_type: "test_environment_not_ready" and that the orchestration guide recognizes it.

🤖 Generated with Claude Code