Agent acceptance gate

agent-acceptance-gate

The last gate before merging multi-agent output. The reconciler describes what each agent did; the acceptance-gate decides whether the round is mergeable.

It runs the success_criteria declared in .coord/plan.yml, aggregates risks, optionally audits prose, checks budget, and produces a single PASS / FAIL / RETRY verdict per task and overall.

Why use this instead of "I'll grep it myself"

The gate replaces ad-hoc inline shell verification with one structured verdict. Concrete differences observed in production (measured-benefits.md R5 + Phase D):

Inline shell verification	This skill (R5 measured)
Main session runs 5+ grep / diff commands, interprets each output, drafts a manual PASS/FAIL — ~10k tokens of shell-debugging in context	Subagent runs all checks + returns ~2k token structured report → ~5× token saving in main session
Easy to forget a check ("did I grep for banned phrases this time?")	Preset YAML encodes the full check set — no missed checks
Catches one drift case; misses cross-language echoes	cross_document_link_text_parity v0.2.3 caught 9+ real drift bugs the human audit missed

The DeepSeek-R2 fabrication that almost shipped

2026-05-13 session: Codex generated a paragraph claiming "DeepSeek-R2 reaches 94.2% on GPQA Diamond" — pure fabrication. Manual grep on the diff missed it because the claim was internally consistent. Only a third-party reviewer agent doing live gh api search caught it.

The fact-check-frontier-models preset was built directly from this incident. Running the preset would have caught the fabrication automatically by hitting arxiv + GitHub for each (model, benchmark, %) triple in the diff.

→ F14: the preset existed by 2026-05-14, but a subsequent operator skipped it because "the task feels simple". The Phase D dogfood (49 files × 3 locales, 50k tokens inline vs 16k with gate = ~3× saving) proved the gate catches drift the operator misses. The cost of running the gate is much smaller than the cost of one missed drift.

Mandatory invocation triggers (no judgment call)

These triggers are mechanical — if any fire, invoke the matching preset before commit. Skipping is F14 territory.

Trigger condition	MUST invoke
Diff touches ≥ 2 locale variants of same file stem (e.g., 06-x.md + 06-x.en.md + 06-x.zh-Hans.md)	--preset=multi-locale-mirror-sync --stem=<stem>
Diff adds entries to any catalog file (project listings, framework comparisons)	--preset=catalog-entry-add --catalog-file=<path>
Diff touches a "frontier model" claim (model name within 3 lines of a benchmark %, e.g., "GPT-5.5 reaches 94% GPQA")	--preset=fact-check-frontier-models --file=<path> --models=<csv>

CLAUDE.md snippet to enforce the gate

## Acceptance gate rule (enforced)

Before any commit that:
- touches ≥ 2 locale variants of the same file stem → invoke
  `Skill("agent-collab-workspace:agent-acceptance-gate",
         args="--preset=multi-locale-mirror-sync --stem=<stem>")`
- adds catalog entries → invoke `--preset=catalog-entry-add ...`
- includes a frontier model + benchmark % claim → invoke
  `--preset=fact-check-frontier-models ...`

Skip only when the trigger does not apply (single-file diff, pure
mechanical rename, etc.). NEVER skip when the trigger fires —
this is the F14 anti-pattern that shipped the DeepSeek-R2
fabrication.

When the preset FAILS:
1. Read the FAIL reasons in the gate report.
2. Either fix-and-rerun (cheap drift) or re-delegate the task with
   tighter constraints (systemic drift).
3. Do NOT override the FAIL by hand. The whole point of the gate is
   that the operator's judgment failed earlier — adding more operator
   judgment on top defeats it.

When to use

Trigger phrases:

• "Run the acceptance gate."
• "Pre-commit check across multi-agent output."
• "Are we ready to commit this round?"
• "Verify all multi-agent output before I push."
• "Gate this round — go or no-go?"

Not for:

• Describing what each agent did → that's agent-output-reconciler.
• Running the agents → delegate skills.
• Per-task acceptance during agent execution → that's the agent skill itself (codex-delegate's wrapper checks its own task's acceptance).
• Single-agent runs without a .coord/plan.yml → just run pytest and call it a day.

Inputs (auto-discovered)

1. .coord/plan.yml — round, tasks, success_criteria per task, budget if declared, and context_policy if declared.
2. .ai/<agent>_log_<NNN>_<slug>.txt.result.json — token usage, risks, files_changed. Filename uses double extension by design (codex-delegate appends .result.json to the log path; see examples/codex_log_001_*.txt.result.json.sample).
3. .coord/reconciliation_<NNN>.md — reconciler's verdict; if reconciler said "retry", gate respects that.
4. .coord/context_<NNN>.md (optional, if agent-context-budget ran) — declared per-task context budgets for this round. Gate checks that actual summary sizes / log tail counts honored the declared budgets. Absence is OK (the round may not have used the context-budget skill); presence means the gate enforces it.
5. For prose changes: if any result.json shows files_changed matching *.md, *.tex, *.docx, the gate optionally invokes academic-writing-skills banned-word + claim-evidence audit. (Skipped silently if academic-writing-skills not installed.)

Presets (one-line invocation for common acceptance shapes)

Instead of hand-writing acceptance criteria every time, invoke a preset that codifies a tested set of checks:

Preset	When to use	Invocation
multi-locale-mirror-sync	After zh-TW → en + zh-Hans mirror sync (or any N-locale fan-out)	agent-acceptance-gate --preset=multi-locale-mirror-sync --stem=stages/06-foo --required-terms="A,B,C"
catalog-entry-add	Added entries to a catalog file	agent-acceptance-gate --preset=catalog-entry-add --catalog-file=resources/foo.md --new-entries="org/repo1,org/repo2"
fact-check-frontier-models	Touched a frontier-model table	agent-acceptance-gate --preset=fact-check-frontier-models --file=stages/06-foo.md --models="GPT-5.5,Claude Opus 4.7"

Preset YAMLs live in presets/. Each codifies failure modes observed in real dogfooding (see docs/observed-failure-modes.md). Don't hand-write checks if a preset covers your case.

Mandatory preset trigger conditions (gate auto-suggests if you don't specify):

• Diff touches ≥ 2 locale variants of the same file stem → multi-locale-mirror-sync MUST be invoked.
• Diff adds entries to any file under resources/ matching catalog shape → catalog-entry-add MUST be invoked.
• Diff touches a frontier-model claim (regex: model name within 3 lines of a benchmark %) → fact-check-frontier-models MUST be invoked.

Preset is mandatory when trigger fires (F14, 2026-05-14)

The presets above are not "consider running" — they are must run before commit when their trigger condition matches. The F14 incident (docs/observed-failure-modes.md) is the cautionary tale: a Phase D run on awesome-agentic-ai-zh touched 49 files across 3 locale variants (textbook multi-locale-mirror-sync trigger), skipped the preset, used a code-reviewer subagent instead, and shipped a drift the preset's cross_document_link_text_parity check was designed to catch.

Why skipping is tempting: when the work feels "just a title sweep, surely nothing can go wrong", the operator short-circuits the mandatory invocation. The presets exist precisely because that intuition is wrong — drift hides in the "obvious" cases.

Anti-patterns:

1. Replacing the preset with an ad-hoc code-reviewer subagent. The subagent is a reasonable backup but cannot substitute for the codified checks, which encode observed failure modes. Run both, not one-instead-of-the-other.
2. "I'll run the preset later". The preset's diff-size check is tied to the commit-staged diff; later means re-staging or running against history. Just run it before commit.
3. Manually grep'ing for the same patterns the preset already knows. You'll miss one. The preset won't.

Enforcement options (in increasing strength):

• Documentation (current): this section + CLAUDE.md rule. Held in 5 of 6 Phase B rounds, failed in Phase D.
• Pre-commit hook (recommended): mechanical check that any mirror-diff commit prompts for preset invocation. Recipe in docs/observed-failure-modes.md F14.
• Block on missing preset run (strict): pre-commit hook fails unless a .coord/acceptance_<NNN>.md file exists in the commit. Use this in repos where the cost of drift is high (curriculum, public docs, anything user-facing).

Workflow

1. Identify round

Read .coord/plan.yml. Default to highest round. User can override.

2. Run each task's success_criteria

For each task with agent: codex or agent: claude-cheap (or, in historical pre-0.3.0 rounds, agent: gemini):

• Each success_criteria is either a runnable command or a checkable assertion.
• Runnable command (pytest tests/auth, mypy src/, npm test):
  • Execute it. Record exit code + last 20 lines of output.
  • PASS = exit 0; FAIL otherwise.
• Checkable assertion ("src/auth/interfaces.py exists and defines AuthProvider ABC"):
  • Translate to a verification (file existence, grep, AST check).
  • Run it.
  • PASS = assertion holds; FAIL otherwise.

For agent: claude tasks:

• success_criteria is usually "explicit YES/NO verdict in chat".
• Read the current Claude conversation for the most recent statement matching the criterion. Mark PASS / FAIL based on whether Claude actually delivered the verdict.

3. Check the reconciler's recommendation

Read .coord/reconciliation_<NNN>.md. If the reconciler's "Recommended action" section says anything other than "merge all" (e.g., "retry T2", "escalate", "manual merge needed"), the gate's verdict is at most CONDITIONAL PASS — the user is responsible for resolving the reconciler's flagged issue.

4. Aggregate risks

Concat all risks arrays from result.json files. Group by severity (gate makes its own call if not labeled — failed test = high; legacy compat concern noted but not breaking = medium).

5. Optional: prose audit

If any task changed *.md / *.tex files AND academic-writing-skills is installed:

• Invoke its banned-word audit on the changed files.
• Invoke its claim-evidence audit if .paper/claims.yml exists in the project.
• Add results to the gate report.

If academic-writing-skills isn't installed, skip silently — don't fail the gate just because prose audit isn't available.

6. Cost / budget check

If .coord/plan.yml declared a budget.tokens:

• Sum tokens field across all result.json files for this round (if present; some delegate wrappers don't write tokens — handle missing gracefully).
• PASS if under budget; FAIL with clear "you exceeded budget by X tokens" if over.

If no budget declared, skip — don't invent one.

6.5. Context contract check

If .coord/plan.yml declares context_policy OR .coord/context_<NNN>.md exists, enforce both:

From context_policy (plan-wide defaults):

• Each result summary must be at or below result_summary_word_budget (default 250 words).
• Reconciliation and acceptance reports must reference raw logs by path, not paste them.
• Failure diagnostics may include only the configured log tail (default 50 lines).
• .coord/memory.yml entries must be promoted facts: decisions, open questions, artifact pointers, or session outcomes. Long analysis in memory is a context violation.

From context_<NNN>.md (per-task overrides, if present):

• Read the per-task task_packet_token_budget declarations.
• For each task, verify <task-id> packet (the actual .ai/<agent>_task_<NNN>_<slug>.md) did not exceed declared budget (rough char/word count, no need for exact tokenizer — flag at >120% of declared).
• Verify result_summary_word_budget per-task (overrides plan-wide default if specified).
• Verify raw_logs_inline: path-only was honored — any log file pasted inline in reconciliation / acceptance is a violation.

Debate caps (if .coord/debate_*.md files exist and are linked from plan.yml):

• Each per-turn Pro / Con argument ≤ 400 words.
• Total debate rounds ≤ 3 (unless plan.yml declares debate_rounds: N override).
• Final synthesis section ≤ 250 words.
• Total debate file size ≤ 8 KB.
• Violations: gate at most CONDITIONAL PASS; recommend compressing the transcript before promoting any decision to memory.

Violations make the verdict at most CONDITIONAL PASS. If the violation hides acceptance evidence, mark FAIL and require a bounded summary rewrite.

6.6. Scope diff check (W1 — work boundary enforcement, file-level)

Compare git diff --name-only against each task's declared files_in_scope (from .coord/plan.yml task entries).

For each modified file F in the diff:

• If F appears in any task's files_in_scope → in-scope ✅
• If F is a "transitive fix" (e.g., anchor heading sync, mentioned in task's result.md as intentional spillover with justification) → accept with WARN
• Otherwise → FAIL with message: Scope violation: <file> not in any task's files_in_scope; agent went outside brief

This is the file-level enforcement for the W1 work boundary discipline. Without this check, brief writing "files in scope: [a, b, c]" is just guidance — the agent may still touch [d, e, f] and only manual diff review would catch it.

What this check does NOT verify (be honest):

• It does NOT verify the agent actually emitted the Confirmed scope: echo block before editing. git diff --name-only only tells you which files got modified — it can't reconstruct whether the echo was the agent's first action. If you need echo verification, add an optional secondary check that greps the agent's result.md for the Confirmed scope: sentinel string.
• It does NOT catch scope violations WITHIN a permitted file (e.g., agent was told to edit foo.md section §3 but also edited §1). Use finer-grained acceptance criteria (per-section grep / line-range check) for that.

F11 + F12 specific catches:

• If diff touches resources/style-guide*.md AND the change replaces a literal term in a contrast table → FAIL (F11 violation)
• If diff inserts a line matching pattern ^>?\s*(Attribution|Source| Credits|Citation)s?:\s* that wasn't requested in brief → FAIL (F12 violation)

7. Compose verdict

Condition	Verdict
All success_criteria PASS, no risks, reconciler says merge, prose audit clean, budget ok	✅ PASS
All success_criteria PASS but reconciler flagged something	⚠ CONDITIONAL PASS — user resolves reconciler's issue, then re-run gate
Context contract violated but evidence is still checkable	⚠ CONDITIONAL PASS — rewrite bounded summaries before next round
Any success_criterion FAIL	❌ FAIL — list which task / criterion
Risks include unresolved blockers	❌ FAIL — must address before merge
Budget exceeded	❌ FAIL — over budget (user explicitly OK can override by editing plan.yml)

8. Write .coord/acceptance_<NNN>.md

# Acceptance gate — round 1

**Verdict:** ⚠ CONDITIONAL PASS
**Run:** 2026-04-28T11:50:00Z
**Tasks gated:** 4
**Reconciliation report:** .coord/reconciliation_001.md

## Per-task results

### T1 — codex — extract-interfaces
- ✅ "src/auth/interfaces.py exists and defines AuthProvider ABC" — file present, grep matches.
- ✅ "no other source files modified" — git diff scope confirmed.

### T2 — codex — refactor-providers
- ❌ "pytest tests/auth/test_providers.py passes" — test_legacy_compat FAILED.
- ✅ "no imports of src.auth.legacy from other modules" — grep clean.

### T3 — gemini — review-doc-coverage
- ✅ "every public symbol in src/auth has a docstring" — gemini's report confirms.
- ✅ "report flags any docstring still mentioning the legacy class" — 12 flagged.

### T4 — claude — design-review
- ✅ "explicit YES/NO verdict + rationale in chat" — said YES with conditional concerns.

## Risks

- **High:** test_legacy_compat failing in T2. Means backwards compat
  is broken under the refactor.

## Budget

Declared: 200,000 tokens. Used: 142,000. ✅ Under budget.

## Reconciler verdict

The reconciler flagged a cross-agent contradiction: T2's fallback
status conflicts with T4's "design is sound" verdict. The reconciler
recommended either retrying T2 with a deprecation shim, or accepting
the breaking change.

## Decision

⚠ **CONDITIONAL PASS.** Don't merge T2 in its current state.

**To unblock:**

Path 1 — Retry T2 with deprecation shim:
  1. Edit `.coord/plan.yml` round 1, T2: add to constraints
     "preserve test_legacy_compat backward compatibility via shim".
  2. Re-run T2 (`bash .claude/skills/codex-delegate/scripts/run_codex.sh ...`).
  3. Re-run reconciler + gate.

Path 2 — Accept breaking change:
  1. Update `tests/auth/test_legacy_compat.py` to reflect new
     architecture.
  2. Re-run pytest manually to confirm green.
  3. Manually mark this gate PASS by replacing the verdict above
     with ✅ PASS + your override rationale.

T1, T3, T4 are individually mergeable; only T2 is blocked.

9. Hand off

[agent-acceptance-gate]
  Round: 1
  Verdict: ⚠ CONDITIONAL PASS
  Tasks gated: 4 (3 pass, 1 fail)
  Risks: 1 high (test_legacy_compat)
  Budget: 142k / 200k tokens — under
  Report: .coord/acceptance_001.md

  Don't push yet. Resolve T2 (see report for two paths).
  Then re-invoke this skill.

What NOT to do

• Don't merge or commit. The gate decides; the user merges.
• Don't override the reconciler. If reconciler said retry, the gate's max verdict is CONDITIONAL PASS.
• Don't invent success criteria. If .coord/plan.yml doesn't declare any for a task, that's a planning bug — flag it, don't invent assertions.
• Don't skip the prose audit silently if it found issues — only skip silently if the audit skill isn't available. If it ran and flagged banned words, those go in the report.
• Don't fail the gate on things the user can override. Budget overrun is a FAIL but the user can edit plan.yml's budget and re-run. Test failures are a real FAIL.
• Don't run inside an agent's task. This skill runs as a final step after all delegate tasks completed and the reconciler has been read. It's a Claude-in-session skill, not delegated.

Subagent review (keep main session lean)

When: ≥ 5 success_criteria checks to run, OR ≥ 4 result.json files to aggregate, OR prose audit on ≥ 3 changed .md files.

Why: The acceptance gate naturally pulls a lot of data into the main session (test outputs, all result.json files, reconciliation report, optional banned-word audits). Delegating the mechanical checks to a subagent and keeping only the structured verdict in main session cuts context cost roughly 3-5×.

Pattern:

Spawn `code-reviewer` subagent with:
  - Read .coord/plan.yml + .coord/reconciliation_<NNN>.md + every
    .ai/*.result.json + .coord/context_<NNN>.md (if present)
  - For each success_criterion in plan.yml: execute the command
    or check the assertion (file existence, grep, etc.)
  - Aggregate risks; sum tokens against budget
  - Verify context contract (summary word budgets, raw-log-paths-only,
    memory promotion rules, debate caps)
  - Return: structured verdict
    { verdict: PASS/CONDITIONAL_PASS/FAIL,
      per_task: [...], risks: [...], budget_used: N,
      context_violations: [...], next_actions: [...] }

Main session reads the structured verdict and writes
.coord/acceptance_<NNN>.md from it without re-reading the result.json /
test output files.

This makes the gate auditable AND keeps the gating session itself under the context_policy main_session_token_budget.

Commit Boundary

Every agent boundary is a commit boundary (see global rule: ~/.claude/CLAUDE.md → "Commit Discipline for Multi-Agent Work"). This makes multi-agent work auditable (commit log = agent log) and enables surgical rollback via git revert <hash> of just one agent's commit.

Specific to this skill: this gate IS the final pre-merge commit check. It reads the per-agent commits between the round's plan commit and HEAD, verifies they collectively satisfy success_criteria, and writes its PASS/FAIL verdict to .coord/acceptance_<NNN>.md as a final commit. Only after PASS does the round get merged to main.