{"id":51551332,"url":"https://github.com/WenyuChiou/agent-collab-skills","last_synced_at":"2026-07-12T14:01:14.573Z","repository":{"id":354332726,"uuid":"1223161782","full_name":"WenyuChiou/agent-collab-skills","owner":"WenyuChiou","description":"Claude Code marketplace for multi-agent collaboration — task splitter, output reconciler, adversarial debate, shared memory, acceptance gate. Composes with codex-delegate / gemini-delegate.","archived":false,"fork":false,"pushed_at":"2026-05-20T19:42:05.000Z","size":1971,"stargazers_count":2,"open_issues_count":0,"forks_count":1,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-22T01:32:17.273Z","etag":null,"topics":["agent-collaboration","ai-agents","claude-code","claude-skills","codex","delegation","gemini","marketplace","multi-agent","orchestration"],"latest_commit_sha":null,"homepage":"https://github.com/WenyuChiou/agent-collab-skills","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/WenyuChiou.png","metadata":{"files":{"readme":"README.md","changelog":"CHANGELOG.md","contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-28T04:10:34.000Z","updated_at":"2026-05-25T16:31:17.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/WenyuChiou/agent-collab-skills","commit_stats":null,"previous_names":["wenyuchiou/agent-collab-skills"],"tags_count":1,"template":false,"template_full_name":null,"purl":"pkg:github/WenyuChiou/agent-collab-skills","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/WenyuChiou%2Fagent-collab-skills","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/WenyuChiou%2Fagent-collab-skills/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/WenyuChiou%2Fagent-collab-skills/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/WenyuChiou%2Fagent-collab-skills/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/WenyuChiou","download_url":"https://codeload.github.com/WenyuChiou/agent-collab-skills/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/WenyuChiou%2Fagent-collab-skills/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":35393398,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-07-12T02:00:06.386Z","response_time":87,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agent-collaboration","ai-agents","claude-code","claude-skills","codex","delegation","gemini","marketplace","multi-agent","orchestration"],"created_at":"2026-07-10T00:00:52.211Z","updated_at":"2026-07-12T14:01:14.558Z","avatar_url":"https://github.com/WenyuChiou.png","language":"Python","funding_links":[],"categories":["Multi-Agent Orchestration"],"sub_categories":[],"readme":"# Agent Collab Skills\n\n[![License](https://img.shields.io/badge/license-MIT-blue)](LICENSE)\n\n**English** ・ [繁體中文](README.zh-TW.md)\n\n![Pipeline overview: agent-task-splitter → codex-delegate / gemini-delegate / claude-in-session → agent-output-reconciler → agent-acceptance-gate, with agent-shared-memory and agent-debate as cross-cutting chips](docs/pipeline-overview.png)\n\n\u003e 6 Claude Code skills for context-safe multi-agent collaboration —\n\u003e task splitter, context budget, output reconciler, adversarial debate,\n\u003e shared memory, acceptance gate. Designed to compose with [`codex-delegate`](https://github.com/WenyuChiou/codex-delegate)\n\u003e and [`gemini-delegate-skill`](https://github.com/WenyuChiou/gemini-delegate-skill).\n\n\u003e 📚 Part of the [**agentic AI learning roadmap**](https://github.com/WenyuChiou/awesome-agentic-ai-zh) — a 7-stage curated path for building agentic AI, multilingual (zh-TW · zh-CN · English). Multi-agent orchestration is covered in Stage 7.\n\nA focused marketplace for the orchestration layer **above** delegation.\nExisting delegate skills solve \"Claude → Codex / Gemini handoff.\" This\ncatalog solves what comes next: how to split a goal across agents, how\nto reconcile their outputs, how to remember decisions across sessions,\nand how to gate merges.\n\nSister marketplace: [`ai-research-skills`](https://github.com/WenyuChiou/ai-research-skills)\n(13 skills for the research workflow).\n\n---\n\n## ⏱ Where the skills help most\n\nThree high-leverage scenarios where these skills clearly pay off:\n\n- **Multi-locale mirror sync** — delegate reads/writes the bulk, main session only consumes a short result summary. Largest saving in our usage.\n- **Mechanical sweeps in parallel** — Codex handles batch edits / renames / refactors; the reviewer pattern catches drift before merge.\n- **Acceptance gate before commit** — structured verdict from a subagent replaces hand-rolled `grep × N` checks, and the YAML presets have caught real drift the human audit missed.\n\nToken measurements and incident logs are in [`docs/measured-benefits.md`](docs/measured-benefits.md) for the curious — but the practical takeaway is just the three scenarios above.\n\n\u003e 📋 **When should I actually invoke which skill / preset?** See [`docs/when-to-invoke.md`](docs/when-to-invoke.md) — has the trigger rules, YES / NO / judgment-call scenarios, decision flowchart, and a **copy-paste CLAUDE.md template** you can drop into your own project to codify the rules.\n\n---\n\n## Install\n\nPrerequisite: Claude Code (https://claude.ai/code). Recommended (but\nnot required): `codex-delegate` and `gemini-delegate` already\ninstalled via `ai-research-skills`, plus their CLI binaries on PATH.\n\n```bash\nclaude plugin marketplace add WenyuChiou/agent-collab-skills\nclaude plugin install agent-collab-workspace@agent-collab-skills\n```\n\nThis installs all 7 skills as one bundle. Verify:\n\n```bash\nclaude plugin list\nls ~/.claude/skills/   # should include agent-task-splitter, etc.\n```\n\nOr use the helper script:\n\n```bash\nbash scripts/install-all.sh        # macOS / Linux / git-bash\npwsh scripts/install-all.ps1       # Windows PowerShell\n```\n\n### Do I need to modify `CLAUDE.md`?\n\n**No.** Claude Code's built-in skill matching reads each `SKILL.md`'s `description` field and auto-routes user prompts. Plugin install is the only setup step — saying *\"split this across Claude, Codex, and Gemini\"* triggers `agent-task-splitter` without any extra configuration.\n\nYou may optionally add explicit routing rules to your `~/.claude/CLAUDE.md` if you:\n\n- already have a delegation protocol there that competes with these skills (e.g., a long-standing *\"always hand-write codex task files\"* rule that would otherwise win the routing race)\n- want to enforce hard behaviors (e.g., *\"always run `agent-acceptance-gate` before merging any multi-agent round\"*)\n\nOtherwise leave `CLAUDE.md` alone — the skills are designed to work via description-based discovery.\n\n---\n\n## The 7 skills\n\n| Skill | Triggered when you say... | Writes to `.coord/` |\n|---|---|---|\n| **`agent-task-splitter`** | \"Split this task across Claude / Codex / Gemini\" / \"Plan a multi-agent run for X\" | `plan.yml` + `.ai/codex_task_*.md` / `.ai/gemini_task_*.md` |\n| **`agent-context-budget`** | \"Context is getting too large\" / \"Prepare a fresh session primer\" / \"Bound this Codex + Gemini run\" | `context_\u003cNNN\u003e.md` + `session_primer.md` |\n| **`agent-output-reconciler`** | \"Reconcile these N agent outputs\" / \"Did Codex and Gemini agree?\" | `reconciliation_\u003cNNN\u003e.md` |\n| **`agent-debate`** | \"Have Claude and Codex debate this design\" / \"Adversarial review on X\" | `debate_\u003ctopic\u003e.md` |\n| **`agent-shared-memory`** | \"Update shared memory with X\" / \"What have agents decided so far?\" | `memory.yml` |\n| **`agent-acceptance-gate`** | \"Run the acceptance gate\" / \"Pre-commit check before merging\" | `acceptance_\u003cNNN\u003e.md` |\n| **`agent-plan-act-reflect`** (v0.2.2) | \"Iterate on X until it passes Y\" / \"Self-correct this draft\" | `par_\u003ctopic\u003e.yml` + `par_\u003ctopic\u003e_final.md` |\n\nNumbering `\u003cNNN\u003e` matches the `round` field in `plan.yml` so artifacts\ntrace back to the run that produced them.\n\n---\n\n## How they compose\n\n```\ngoal\n  ↓ agent-task-splitter\n.coord/plan.yml + .ai/codex_task_*.md / .ai/gemini_task_*.md\n  ↓ agent-context-budget\n.coord/context_\u003cNNN\u003e.md + .coord/session_primer.md\n  ↓ codex-delegate / gemini-delegate (existing)\n.ai/codex_log_*.txt + .result.json + codex_result_*.md\n  ↓ agent-output-reconciler\n.coord/reconciliation_\u003cNNN\u003e.md\n  ↓ agent-acceptance-gate\n.coord/acceptance_\u003cNNN\u003e.md → merge or retry\n```\n\n`agent-shared-memory` runs alongside the whole pipeline and stores\naccepted decisions, open questions, artifacts, and session outcomes.\n`agent-context-budget` keeps handoffs bounded so the main session does\nnot absorb raw logs or full memory. `agent-debate` is invoked on\nconsequential decisions (architecture, design choice), not in the main\nloop.\n\nSee [docs/example-walkthrough.md](docs/example-walkthrough.md) for\na worked example with real `.coord/` sample artifacts produced by\nrunning the pipeline end-to-end against actual Codex + Gemini CLI\ninvocations.\n\n---\n\n## Why these 6 specifically\n\nThe pain points each one solves, in order:\n\n1. **Task splitting is mental load.** You currently classify \"is\n   this Codex-shaped or Gemini-shaped?\" in your head every time. The\n   splitter encodes the heuristics.\n2. **Context explodes during large runs.** The context-budget skill\n   turns memory, logs, and agent outputs into bounded packets and a\n   session primer.\n3. **Multi-agent output is hard to compare.** When 3 Codex jobs come\n   back in parallel, you read 3 result.json files and merge them\n   manually. The reconciler does the diff.\n4. **Consensus-driven LLM output hides trade-offs.** When you ask one\n   agent for a design, you get one answer. The debate skill forces\n   two agents to argue opposing positions.\n5. **Agent sessions don't share memory.** Codex resume works\n   per-session; nothing persists across Claude session A → Codex\n   session B → Gemini session C. Shared-memory makes `.coord/memory.yml`\n   the cross-session blackboard.\n6. **No standardized merge gate.** You currently eyeball the diff +\n   run `pytest` manually. The gate runs all `success_criteria` from\n   `plan.yml` + cost budget + cross-agent consistency check.\n\n---\n\n## Composes with\n\n- [`codex-delegate`](https://github.com/WenyuChiou/codex-delegate) —\n  consumes splitter output, produces input for reconciler.\n- [`gemini-delegate-skill`](https://github.com/WenyuChiou/gemini-delegate-skill)\n  — same.\n- [`academic-writing-skills`](https://github.com/WenyuChiou/academic-writing-skills)\n  — acceptance gate optionally calls its banned-word audit on prose\n  changes.\n- [`agentmemory`](https://github.com/rohitg00/agentmemory) —\n  optional recall cache only. `.coord/memory.yml` remains canonical;\n  see [docs/agentmemory-integration.md](docs/agentmemory-integration.md).\n\n---\n\n## Known issues\n\n- **Gemini-cli refuses to read gitignored files.** The `.ai/`\n  directory is conventionally gitignored to keep transient task\n  files out of commits, but `gemini -p \"Read .ai/gemini_task_*.md\"`\n  fails with `file ignored by configured ignore patterns`.\n  **Workaround**: invoke gemini with the task content inlined into\n  the prompt:\n  ```bash\n  TASK=$(cat .ai/gemini_task_\u003cNNN\u003e_\u003cslug\u003e.md)\n  gemini -p \"$TASK\" --yolo \\\n    \u003c /dev/null \u003e .ai/gemini_log_\u003cNNN\u003e_\u003cslug\u003e.txt 2\u003e\u00261\n  ```\n  Side effect: gemini doesn't have file-system context for paths\n  the task file references — make sure the prompt body itself\n  contains all critical context, not just paths to read. The\n  splitter skill's step 6b documents this.\n- **Both `codex` and `gemini` hang at startup if stdin is open.**\n  When launching from a script or non-interactive shell, codex-cli\n  ≥ 0.121.0 prints \"Reading additional input from stdin...\" and\n  waits forever. Same for gemini-cli. **Workaround**: redirect\n  stdin to `/dev/null` on every direct invocation:\n  ```bash\n  codex exec --full-auto -m \u003cmodel\u003e \"\u003cprompt\u003e\" \\\n    \u003c /dev/null \u003e .ai/codex_log_\u003cNNN\u003e_\u003cslug\u003e.txt 2\u003e\u00261\n  ```\n  The `codex-delegate` wrapper script handles this internally; only\n  direct `codex exec` / `gemini -p` calls need the explicit\n  redirect.\n- **Codex reads gitignored files normally** — only gemini has the\n  gitignore conflict.\n- **Worked example** with sample `.coord/` artifacts and honest\n  documentation of what real multi-agent runs look like:\n  [docs/example-walkthrough.md](docs/example-walkthrough.md).\n\n---\n\n## Status \u0026 License\n\nMIT. Early-stage — the SKILL.md prompt scaffolding is complete and\ntested in real workflows; please file issues if a skill misfires or\nthe `.coord/` schema breaks under your use case.\n\nContributions welcome — see [CONTRIBUTING.md](CONTRIBUTING.md) for\nthe catalog ↔ delegate-skill interop rules.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2FWenyuChiou%2Fagent-collab-skills","html_url":"https://awesome.ecosyste.ms/projects/github.com%2FWenyuChiou%2Fagent-collab-skills","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2FWenyuChiou%2Fagent-collab-skills/lists"}