agent-acceptance-gate

Name: agent-acceptance-gate
Author: WenyuChiou

16stars4forks

How to install agent-acceptance-gate?

wenyuchiou/agent-collab-skills/agent-acceptance-gate

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-acceptance-gate

Installs into the current project.

Summary

Use when a multi-agent round needs a pre-merge gate, pre-commit check, verification before push, or a PASS/FAIL decision after reconciliation.

Files · 1

View on GitHub

SKILL.md

1	# agent-acceptance-gate
2
3	The last gate before merging multi-agent output. The reconciler
4	describes what each agent did; the acceptance-gate decides
5	whether the round is mergeable.
6
7	It runs the `success_criteria` declared in `.coord/plan.yml`,
8	aggregates risks, optionally audits prose, checks budget, and
9	produces a single PASS / FAIL / RETRY verdict per task and overall.
10
11	## Why use this instead of "I'll grep it myself"
12
13	The gate replaces ad-hoc inline shell verification with one structured
14	verdict. Concrete differences observed in production (`measured-benefits.md`
15	R5 + Phase D):
16
17	\| Inline shell verification \| This skill (R5 measured) \|
18	\|---\|---\|
19	\| 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 \|
20	\| Easy to forget a check ("did I grep for banned phrases this time?") \| Preset YAML encodes the full check set — no missed checks \|
21	\| 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 \|
22
23	### The DeepSeek-R2 fabrication that almost shipped
24
25	2026-05-13 session: Codex generated a paragraph claiming "DeepSeek-R2
26	reaches 94.2% on GPQA Diamond" — pure fabrication. Manual grep on the
27	diff missed it because the claim was internally consistent. Only a
28	third-party reviewer agent doing live `gh api` search caught it.
29
30	The `fact-check-frontier-models` preset was built directly from this
31	incident. Running the preset would have caught the fabrication
32	automatically by hitting arxiv + GitHub for each (model, benchmark, %)
33	triple in the diff.
34
35	→ F14: the preset existed by 2026-05-14, but a subsequent operator
36	skipped it because "the task feels simple". The Phase D dogfood
37	(49 files × 3 locales, 50k tokens inline vs 16k with gate = ~3× saving)
38	proved the gate catches drift the operator misses. **The cost of running
39	the gate is much smaller than the cost of one missed drift.**
40
41	### Mandatory invocation triggers (no judgment call)
42
43	These triggers are mechanical — if any fire, invoke the matching preset
44	before commit. Skipping is F14 territory.
45
46	\| Trigger condition \| MUST invoke \|
47	\|---\|---\|
48	\| 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>` \|
49	\| Diff adds entries to any catalog file (project listings, framework comparisons) \| `--preset=catalog-entry-add --catalog-file=<path>` \|
50	\| 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>` \|
51
52	### CLAUDE.md snippet to enforce the gate
53
54	```markdown
55	## Acceptance gate rule (enforced)
56
57	Before any commit that:
58	- touches ≥ 2 locale variants of the same file stem → invoke
59	`Skill("agent-collab-workspace:agent-acceptance-gate",
60	args="--preset=multi-locale-mirror-sync --stem=<stem>")`
61	- adds catalog entries → invoke `--preset=catalog-entry-add ...`
62	- includes a frontier model + benchmark % claim → invoke
63	`--preset=fact-check-frontier-models ...`
64
65	Skip only when the trigger does not apply (single-file diff, pure
66	mechanical rename, etc.). NEVER skip when the trigger fires —
67	this is the F14 anti-pattern that shipped the DeepSeek-R2
68	fabrication.
69
70	When the preset FAILS:
71	1. Read the FAIL reasons in the gate report.
72	2. Either fix-and-rerun (cheap drift) or re-delegate the task with
73	tighter constraints (systemic drift).
74	3. Do NOT override the FAIL by hand. The whole point of the gate is
75	that the operator's judgment failed earlier — adding more operator
76	judgment on top defeats it.
77	```
78
79	## When to use
80
81	Trigger phrases:
82
83	- "Run the acceptance gate."
84	- "Pre-commit check across multi-agent output."
85	- "Are we ready to commit this round?"
86	- "Verify all multi-agent output before I push."
87	- "Gate this round — go or no-go?"
88
89	Not for:
90
91	- Describing what each agent did → that's
92	`agent-output-reconciler`.
93	- Running the agents → delegate skills.
94	- Per-task acceptance during agent execution → that's the agent
95	skill itself (codex-delegate's wrapper checks its own task's
96	acceptance).
97	- Single-agent runs without a `.coord/plan.yml` → just run
98	`pytest` and call it a day.
99
100	## Inputs (auto-discovered)
101
102	1. `.coord/plan.yml` — round, tasks, `success_criteria` per
103	task, `budget` if declared, and `context_policy` if declared.
104	2. `.ai/<agent>_log_<NNN>_<slug>.txt.result.json` — token usage,
105	risks, files_changed. Filename uses double extension by design
106	(codex-delegate appends `.result.json` to the log path; see
107	`examples/codex_log_001_*.txt.result.json.sample`).
108	3.

People who liked this also viewed

6 picks

Type

lark-openapi-explorer飞书/Lark 原生 OpenAPI 探索：从官方文档库中挖掘未经 CLI 封装的原生 OpenAPI 接口。当用户的需求无法被现有 lark-* skill 或 lark-cli 已注册命令满足，需要查找并调用原生飞书 OpenAPI 时使用。SkillsJul 2026378k16k
lark-skill-maker创建 lark-cli 的自定义 Skill。当用户需要把飞书 API 操作封装成可复用的 Skill（包装原子 API 或编排多步流程）时使用。SkillsJul 2026377k16k
azure-cost-optimizationOfficial agent plugin providing skills and MCP server configurations for Azure scenarios.SkillsJul 2026209k1.3k
vercel-react-native-skillsVercel's official collection of agent skillsSkillsJul 2026172k29k
mcp-builderGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate externaSkillsJul 202693k163k
gws-sharedgws CLI: Shared patterns for authentication, global flags, and output formatting.SkillsJul 202653k30k

agent-acceptance-gate

16stars4forks

View on GitHub

How to install agent-acceptance-gate?

wenyuchiou/agent-collab-skills/agent-acceptance-gate

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-acceptance-gate

Installs into the current project.

Summary

Use when a multi-agent round needs a pre-merge gate, pre-commit check, verification before push, or a PASS/FAIL decision after reconciliation.

Files · 1

View on GitHub

SKILL.md

1	# agent-acceptance-gate
2
3	The last gate before merging multi-agent output. The reconciler
4	describes what each agent did; the acceptance-gate decides
5	whether the round is mergeable.
6
7	It runs the `success_criteria` declared in `.coord/plan.yml`,
8	aggregates risks, optionally audits prose, checks budget, and
9	produces a single PASS / FAIL / RETRY verdict per task and overall.
10
11	## Why use this instead of "I'll grep it myself"
12
13	The gate replaces ad-hoc inline shell verification with one structured
14	verdict. Concrete differences observed in production (`measured-benefits.md`
15	R5 + Phase D):
16
17	\| Inline shell verification \| This skill (R5 measured) \|
18	\|---\|---\|
19	\| 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 \|
20	\| Easy to forget a check ("did I grep for banned phrases this time?") \| Preset YAML encodes the full check set — no missed checks \|
21	\| 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 \|
22
23	### The DeepSeek-R2 fabrication that almost shipped
24
25	2026-05-13 session: Codex generated a paragraph claiming "DeepSeek-R2
26	reaches 94.2% on GPQA Diamond" — pure fabrication. Manual grep on the
27	diff missed it because the claim was internally consistent. Only a
28	third-party reviewer agent doing live `gh api` search caught it.
29
30	The `fact-check-frontier-models` preset was built directly from this
31	incident. Running the preset would have caught the fabrication
32	automatically by hitting arxiv + GitHub for each (model, benchmark, %)
33	triple in the diff.
34
35	→ F14: the preset existed by 2026-05-14, but a subsequent operator
36	skipped it because "the task feels simple". The Phase D dogfood
37	(49 files × 3 locales, 50k tokens inline vs 16k with gate = ~3× saving)
38	proved the gate catches drift the operator misses. **The cost of running
39	the gate is much smaller than the cost of one missed drift.**
40
41	### Mandatory invocation triggers (no judgment call)
42
43	These triggers are mechanical — if any fire, invoke the matching preset
44	before commit. Skipping is F14 territory.
45
46	\| Trigger condition \| MUST invoke \|
47	\|---\|---\|
48	\| 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>` \|
49	\| Diff adds entries to any catalog file (project listings, framework comparisons) \| `--preset=catalog-entry-add --catalog-file=<path>` \|
50	\| 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>` \|
51
52	### CLAUDE.md snippet to enforce the gate
53
54	```markdown
55	## Acceptance gate rule (enforced)
56
57	Before any commit that:
58	- touches ≥ 2 locale variants of the same file stem → invoke
59	`Skill("agent-collab-workspace:agent-acceptance-gate",
60	args="--preset=multi-locale-mirror-sync --stem=<stem>")`
61	- adds catalog entries → invoke `--preset=catalog-entry-add ...`
62	- includes a frontier model + benchmark % claim → invoke
63	`--preset=fact-check-frontier-models ...`
64
65	Skip only when the trigger does not apply (single-file diff, pure
66	mechanical rename, etc.). NEVER skip when the trigger fires —
67	this is the F14 anti-pattern that shipped the DeepSeek-R2
68	fabrication.
69
70	When the preset FAILS:
71	1. Read the FAIL reasons in the gate report.
72	2. Either fix-and-rerun (cheap drift) or re-delegate the task with
73	tighter constraints (systemic drift).
74	3. Do NOT override the FAIL by hand. The whole point of the gate is
75	that the operator's judgment failed earlier — adding more operator
76	judgment on top defeats it.
77	```
78
79	## When to use
80
81	Trigger phrases:
82
83	- "Run the acceptance gate."
84	- "Pre-commit check across multi-agent output."
85	- "Are we ready to commit this round?"
86	- "Verify all multi-agent output before I push."
87	- "Gate this round — go or no-go?"
88
89	Not for:
90
91	- Describing what each agent did → that's
92	`agent-output-reconciler`.
93	- Running the agents → delegate skills.
94	- Per-task acceptance during agent execution → that's the agent
95	skill itself (codex-delegate's wrapper checks its own task's
96	acceptance).
97	- Single-agent runs without a `.coord/plan.yml` → just run
98	`pytest` and call it a day.
99
100	## Inputs (auto-discovered)
101
102	1. `.coord/plan.yml` — round, tasks, `success_criteria` per
103	task, `budget` if declared, and `context_policy` if declared.
104	2. `.ai/<agent>_log_<NNN>_<slug>.txt.result.json` — token usage,
105	risks, files_changed. Filename uses double extension by design
106	(codex-delegate appends `.result.json` to the log path; see
107	`examples/codex_log_001_*.txt.result.json.sample`).
108	3.

People who liked this also viewed

6 picks

Type