Agent Collab Skills
English ・ 繁體中文
6 Claude Code skills for context-safe multi-agent collaboration —
task splitter, context budget, output reconciler, adversarial debate,
shared memory, acceptance gate. Designed to compose with codex-delegate
and gemini-delegate-skill.
📚 Part of the agentic AI learning roadmap — a 7-stage curated path for building agentic AI, multilingual (zh-TW · zh-CN · English). Multi-agent orchestration is covered in Stage 7.
A focused marketplace for the orchestration layer above delegation.
Existing delegate skills solve "Claude → Codex / Gemini handoff." This
catalog solves what comes next: how to split a goal across agents, how
to reconcile their outputs, how to remember decisions across sessions,
and how to gate merges.
Sister marketplace: ai-research-skills
(13 skills for the research workflow).
⏱ Where the skills help most
Three high-leverage scenarios where these skills clearly pay off:
- Multi-locale mirror sync — delegate reads/writes the bulk, main session only consumes a short result summary. Largest saving in our usage.
- Mechanical sweeps in parallel — Codex handles batch edits / renames / refactors; the reviewer pattern catches drift before merge.
- 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.
Token measurements and incident logs are in docs/measured-benefits.md for the curious — but the practical takeaway is just the three scenarios above.
📋 When should I actually invoke which skill / preset? See 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.
Install
Prerequisite: Claude Code (https://claude.ai/code). Recommended (but
not required): codex-delegate and gemini-delegate already
installed via ai-research-skills, plus their CLI binaries on PATH.
claude plugin marketplace add WenyuChiou/agent-collab-skills
claude plugin install agent-collab-workspace@agent-collab-skills
This installs all 7 skills as one bundle. Verify:
claude plugin list
ls ~/.claude/skills/ # should include agent-task-splitter, etc.
Or use the helper script:
bash scripts/install-all.sh # macOS / Linux / git-bash
pwsh scripts/install-all.ps1 # Windows PowerShell
Do I need to modify CLAUDE.md?
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.
You may optionally add explicit routing rules to your ~/.claude/CLAUDE.md if you:
- 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)
- want to enforce hard behaviors (e.g., "always run
agent-acceptance-gate before merging any multi-agent round")
Otherwise leave CLAUDE.md alone — the skills are designed to work via description-based discovery.
The 7 skills
| Skill | Triggered when you say... | Writes to .coord/ |
|---|
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 |
agent-context-budget | "Context is getting too large" / "Prepare a fresh session primer" / "Bound this Codex + Gemini run" | context_<NNN>.md + session_primer.md |
agent-output-reconciler | "Reconcile these N agent outputs" / "Did Codex and Gemini agree?" | reconciliation_<NNN>.md |
agent-debate | "Have Claude and Codex debate this design" / "Adversarial review on X" | debate_<topic>.md |
agent-shared-memory | "Update shared memory with X" / "What have agents decided so far?" | memory.yml |
agent-acceptance-gate | "Run the acceptance gate" / "Pre-commit check before merging" | acceptance_<NNN>.md |
agent-plan-act-reflect (v0.2.2) | "Iterate on X until it passes Y" / "Self-correct this draft" | par_<topic>.yml + par_<topic>_final.md |
Numbering <NNN> matches the round field in plan.yml so artifacts
trace back to the run that produced them.
How they compose
goal
↓ agent-task-splitter
.coord/plan.yml + .ai/codex_task_*.md / .ai/gemini_task_*.md
↓ agent-context-budget
.coord/context_<NNN>.md + .coord/session_primer.md
↓ codex-delegate / gemini-delegate (existing)
.ai/codex_log_*.txt + .result.json + codex_result_*.md
↓ agent-output-reconciler
.coord/reconciliation_<NNN>.md
↓ agent-acceptance-gate
.coord/acceptance_<NNN>.md → merge or retry
agent-shared-memory runs alongside the whole pipeline and stores
accepted decisions, open questions, artifacts, and session outcomes.
agent-context-budget keeps handoffs bounded so the main session does
not absorb raw logs or full memory. agent-debate is invoked on
consequential decisions (architecture, design choice), not in the main
loop.
See docs/example-walkthrough.md for
a worked example with real .coord/ sample artifacts produced by
running the pipeline end-to-end against actual Codex + Gemini CLI
invocations.
Why these 6 specifically
The pain points each one solves, in order:
- Task splitting is mental load. You currently classify "is
this Codex-shaped or Gemini-shaped?" in your head every time. The
splitter encodes the heuristics.
- Context explodes during large runs. The context-budget skill
turns memory, logs, and agent outputs into bounded packets and a
session primer.
- Multi-agent output is hard to compare. When 3 Codex jobs come
back in parallel, you read 3 result.json files and merge them
manually. The reconciler does the diff.
- Consensus-driven LLM output hides trade-offs. When you ask one
agent for a design, you get one answer. The debate skill forces
two agents to argue opposing positions.
- Agent sessions don't share memory. Codex resume works
per-session; nothing persists across Claude session A → Codex
session B → Gemini session C. Shared-memory makes
.coord/memory.yml
the cross-session blackboard.
- No standardized merge gate. You currently eyeball the diff +
run
pytest manually. The gate runs all success_criteria from
plan.yml + cost budget + cross-agent consistency check.
Composes with
Known issues
- Gemini-cli refuses to read gitignored files. The
.ai/
directory is conventionally gitignored to keep transient task
files out of commits, but gemini -p "Read .ai/gemini_task_*.md"
fails with file ignored by configured ignore patterns.
Workaround: invoke gemini with the task content inlined into
the prompt:
TASK=$(cat .ai/gemini_task_<NNN>_<slug>.md)
gemini -p "$TASK" --yolo \
< /dev/null > .ai/gemini_log_<NNN>_<slug>.txt 2>&1
Side effect: gemini doesn't have file-system context for paths
the task file references — make sure the prompt body itself
contains all critical context, not just paths to read. The
splitter skill's step 6b documents this.
- Both
codex and gemini hang at startup if stdin is open.
When launching from a script or non-interactive shell, codex-cli
≥ 0.121.0 prints "Reading additional input from stdin..." and
waits forever. Same for gemini-cli. Workaround: redirect
stdin to /dev/null on every direct invocation:
codex exec --full-auto -m <model> "<prompt>" \
< /dev/null > .ai/codex_log_<NNN>_<slug>.txt 2>&1
The codex-delegate wrapper script handles this internally; only
direct codex exec / gemini -p calls need the explicit
redirect.
- Codex reads gitignored files normally — only gemini has the
gitignore conflict.
- Worked example with sample
.coord/ artifacts and honest
documentation of what real multi-agent runs look like:
docs/example-walkthrough.md.
Status & License
MIT. Early-stage — the SKILL.md prompt scaffolding is complete and
tested in real workflows; please file issues if a skill misfires or
the .coord/ schema breaks under your use case.
Contributions welcome — see CONTRIBUTING.md for
the catalog ↔ delegate-skill interop rules.
