agent-output-reconciler

Name: agent-output-reconciler
Author: WenyuChiou

16stars4forks

How to install agent-output-reconciler?

wenyuchiou/agent-collab-skills/agent-output-reconciler

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-output-reconciler

Installs into the current project.

Summary

Use when multiple agents have completed a round and the user asks to reconcile outputs, compare Codex and Gemini, synthesize run results, identify conflicts, or decide what should be retried.

Files · 1

View on GitHub

SKILL.md

1	# agent-output-reconciler
2
3	Cross-agent diff + synthesizer. After `agent-task-splitter` plans a
4	round and the delegate skills have run their tasks, this skill reads
5	everything they produced and reports:
6
7	- Did each task succeed?
8	- Where do agents agree (same files changed similarly, same
9	conclusions)?
10	- Where do they conflict (same files changed differently, contradictory
11	recommendations)?
12	- What's the recommended next move (merge / retry / escalate)?
13
14	## When to use
15
16	Trigger phrases:
17
18	- "Reconcile these N agent outputs."
19	- "Did Codex and Gemini agree on this round?"
20	- "Synthesize the multi-agent run results."
21	- "What did the agents do this round? Any conflicts?"
22	- "Round N is done — give me the reconciliation report."
23
24	Not for:
25
26	- Running the agents — that's `codex-delegate` (Claude lanes run via the Agent tool; `gemini-delegate` is deprecated, fails closed).
27	- Final accept-or-reject gate before merging — that's
28	`agent-acceptance-gate`. The reconciler describes; the
29	acceptance gate decides.
30	- Single-agent runs — if only one task in `.coord/plan.yml`, there's
31	nothing to reconcile.
32
33	## Inputs (auto-discovered)
34
35	1. `.coord/plan.yml` — the round's plan. Use the `round` field
36	to identify which task files are in scope.
37	2. `.ai/<agent>_log_<NNN>_<slug>.txt.result.json` — one per
38	non-Claude task. Schema (from `codex-delegate`'s contract):
39	```json
40	{
41	"status": "success\|fallback\|error",
42	"delegate": "codex\|gemini",
43	"model": "...",
44	"log_file": "...",
45	"output_file": "...",
46	"summary": "...",
47	"risks": [],
48	"files_changed": [],
49	"tests_run": [],
50	"timestamp_utc": "...",
51
52	// v0.2.2+ optional fields — see §2.6 promise/delivery contract:
53	"promised": [], // artifacts this task surfaces for downstream consumers
54	"consumed": [] // artifacts this task used from upstream tasks
55	}
56	```
57	Each entry in `risks` must be a single sentence (≤ 30 words).
58	Long-form analysis belongs in `output_file`, not here — verbose
59	`risks` entries are a context-contract violation and the
60	reconciler should flag them.
61	3. `.ai/<agent>_result_<NNN>_<slug>.md` — agent-written summary
62	(referenced from each task file's `Acceptance` section).
63	4. `.ai/<agent>_log_<NNN>_<slug>.txt` — full log path; read only
64	the configured tail for error context if `status: error`.
65	5. For `agent: claude` tasks — read the Claude session's
66	in-conversation output (whatever Claude said in the chat for that
67	task, treated as the equivalent of `result.md`).
68	6. `.coord/context_<NNN>.md` (optional, if `agent-context-budget`
69	ran) — declared per-task context budgets. Reconciler uses these
70	to flag oversized summaries / unbounded `risks` arrays at the
71	per-task granularity, not just the plan-wide default. Absence is OK.
72
73	If the user passes specific paths, use those instead of
74	auto-discovery.
75
76	## Workflow
77
78	### 1. Identify round + tasks
79
80	Read `.coord/plan.yml`. If multiple rounds exist, default to the
81	highest `round` number unless the user specifies. Collect the list
82	of `(task_id, agent, slug)`.
83
84	### 2. Read each task's outputs
85
86	For each task:
87	- Codex / Gemini: load `result.json` + `result_<NNN>_<slug>.md` +
88	log tail only when `status: error` (default max: last 50 lines).
89	- Claude: pull from current conversation history.
90
91	Flag any task where:
92	- `result.json` is missing → run never completed.
93	- `status: "error"` → run failed.
94	- `status: "fallback"` → run completed but in degraded mode.
95	- `result_<NNN>_<slug>.md` missing → agent didn't write the
96	required summary (acceptance criterion violated).
97	- `result_<NNN>_<slug>.md` exceeds `context_policy.result_summary_word_budget`
98	(default 250 words) → context contract violated.
99
100	### 2.4. Multi-locale lockstep check (catches Gemini drop / merge)
101
102	When the round's outputs include ≥ 2 locale variants of the same
103	file stem (e.g., `06-memory-rag.md` + `06-memory-rag.en.md` +
104	`06-memory-rag.zh-Hans.md`), verify they actually stayed in lockstep:
105
106	1. Line count parity — each mirror within ±3% of canonical's
107	line count. Larger delta usually means Gemini dropped or
108	duplicated a section.
109	2. H2 count parity — `grep -c '^## '` returns identical count
110	across all locales.
111	3. Per-table column count parity — for each markdown table at
112	position N in canonical, verify table N in each mirror has the
113	same column count. (F2 incident in `docs/observed-failure-modes.md`:
114	Gemini merged a 5-column Projects table's rows into a 3-column
115	Tools table.)
116	4. Required headline-term cross-presence — if `plan.yml` declares
117	`required_terms` for this round, each must appear in every locale
118	variant. Missing in one locale = mirror sync dropped content.
119	5.

People who liked this also viewed

6 picks

Type

agent-output-reconciler

16stars4forks

View on GitHub

How to install agent-output-reconciler?

wenyuchiou/agent-collab-skills/agent-output-reconciler

$npx -y skills add WenyuChiou/agent-collab-skills --skill agent-output-reconciler

Installs into the current project.

Summary

Use when multiple agents have completed a round and the user asks to reconcile outputs, compare Codex and Gemini, synthesize run results, identify conflicts, or decide what should be retried.

Files · 1

View on GitHub

SKILL.md

1	# agent-output-reconciler
2
3	Cross-agent diff + synthesizer. After `agent-task-splitter` plans a
4	round and the delegate skills have run their tasks, this skill reads
5	everything they produced and reports:
6
7	- Did each task succeed?
8	- Where do agents agree (same files changed similarly, same
9	conclusions)?
10	- Where do they conflict (same files changed differently, contradictory
11	recommendations)?
12	- What's the recommended next move (merge / retry / escalate)?
13
14	## When to use
15
16	Trigger phrases:
17
18	- "Reconcile these N agent outputs."
19	- "Did Codex and Gemini agree on this round?"
20	- "Synthesize the multi-agent run results."
21	- "What did the agents do this round? Any conflicts?"
22	- "Round N is done — give me the reconciliation report."
23
24	Not for:
25
26	- Running the agents — that's `codex-delegate` (Claude lanes run via the Agent tool; `gemini-delegate` is deprecated, fails closed).
27	- Final accept-or-reject gate before merging — that's
28	`agent-acceptance-gate`. The reconciler describes; the
29	acceptance gate decides.
30	- Single-agent runs — if only one task in `.coord/plan.yml`, there's
31	nothing to reconcile.
32
33	## Inputs (auto-discovered)
34
35	1. `.coord/plan.yml` — the round's plan. Use the `round` field
36	to identify which task files are in scope.
37	2. `.ai/<agent>_log_<NNN>_<slug>.txt.result.json` — one per
38	non-Claude task. Schema (from `codex-delegate`'s contract):
39	```json
40	{
41	"status": "success\|fallback\|error",
42	"delegate": "codex\|gemini",
43	"model": "...",
44	"log_file": "...",
45	"output_file": "...",
46	"summary": "...",
47	"risks": [],
48	"files_changed": [],
49	"tests_run": [],
50	"timestamp_utc": "...",
51
52	// v0.2.2+ optional fields — see §2.6 promise/delivery contract:
53	"promised": [], // artifacts this task surfaces for downstream consumers
54	"consumed": [] // artifacts this task used from upstream tasks
55	}
56	```
57	Each entry in `risks` must be a single sentence (≤ 30 words).
58	Long-form analysis belongs in `output_file`, not here — verbose
59	`risks` entries are a context-contract violation and the
60	reconciler should flag them.
61	3. `.ai/<agent>_result_<NNN>_<slug>.md` — agent-written summary
62	(referenced from each task file's `Acceptance` section).
63	4. `.ai/<agent>_log_<NNN>_<slug>.txt` — full log path; read only
64	the configured tail for error context if `status: error`.
65	5. For `agent: claude` tasks — read the Claude session's
66	in-conversation output (whatever Claude said in the chat for that
67	task, treated as the equivalent of `result.md`).
68	6. `.coord/context_<NNN>.md` (optional, if `agent-context-budget`
69	ran) — declared per-task context budgets. Reconciler uses these
70	to flag oversized summaries / unbounded `risks` arrays at the
71	per-task granularity, not just the plan-wide default. Absence is OK.
72
73	If the user passes specific paths, use those instead of
74	auto-discovery.
75
76	## Workflow
77
78	### 1. Identify round + tasks
79
80	Read `.coord/plan.yml`. If multiple rounds exist, default to the
81	highest `round` number unless the user specifies. Collect the list
82	of `(task_id, agent, slug)`.
83
84	### 2. Read each task's outputs
85
86	For each task:
87	- Codex / Gemini: load `result.json` + `result_<NNN>_<slug>.md` +
88	log tail only when `status: error` (default max: last 50 lines).
89	- Claude: pull from current conversation history.
90
91	Flag any task where:
92	- `result.json` is missing → run never completed.
93	- `status: "error"` → run failed.
94	- `status: "fallback"` → run completed but in degraded mode.
95	- `result_<NNN>_<slug>.md` missing → agent didn't write the
96	required summary (acceptance criterion violated).
97	- `result_<NNN>_<slug>.md` exceeds `context_policy.result_summary_word_budget`
98	(default 250 words) → context contract violated.
99
100	### 2.4. Multi-locale lockstep check (catches Gemini drop / merge)
101
102	When the round's outputs include ≥ 2 locale variants of the same
103	file stem (e.g., `06-memory-rag.md` + `06-memory-rag.en.md` +
104	`06-memory-rag.zh-Hans.md`), verify they actually stayed in lockstep:
105
106	1. Line count parity — each mirror within ±3% of canonical's
107	line count. Larger delta usually means Gemini dropped or
108	duplicated a section.
109	2. H2 count parity — `grep -c '^## '` returns identical count
110	across all locales.
111	3. Per-table column count parity — for each markdown table at
112	position N in canonical, verify table N in each mirror has the
113	same column count. (F2 incident in `docs/observed-failure-modes.md`:
114	Gemini merged a 5-column Projects table's rows into a 3-column
115	Tools table.)
116	4. Required headline-term cross-presence — if `plan.yml` declares
117	`required_terms` for this round, each must appear in every locale
118	variant. Missing in one locale = mirror sync dropped content.
119	5.

People who liked this also viewed

6 picks

Type