agent-collab

Agent skills for coordinating work between agent sessions - distributed via the Agent Skills open standard.

Skills

Auto-applied when relevant to the current task, and directly callable where noted.

agent-thread

Let two already-running agent sessions (any mix of Claude Code and Codex, same machine) collaborate by taking turns through a shared on-disk thread, looping until resolved - so you stop copy-pasting messages between two agents. The channel is pattern-agnostic; it covers review, debate/decide, consult (a peer with different tools/model/repo), delegate a sub-task, pair (driver/navigator), and brainstorm. Triggers whenever you want a second agent to weigh in until you converge, or want two open sessions to talk without you relaying. Either side can also drop an out-of-band note at any time without taking the turn. Invoke explicitly as $agent-thread in Codex or /agent-thread in Claude. For a persistent session that may span compaction, a pause, or ownership transfer, compose it with maintaining-continuous-handoffs.

• SKILL.md
• Design notes
• CLI: scripts/athread.mjs (zero deps, Node >= 18)

maintaining-continuous-handoffs

Keep a repository-local HANDOFF.md/LOG.md pair current while work spans compaction, pauses, active operations, or ownership transfers. HANDOFF.md contains mutable resume state; LOG.md contains append-only history. The skill reconciles recorded claims against real state and keeps exactly one next action visible. Small same-turn tasks skip the pair entirely. Invoke explicitly as $maintaining-continuous-handoffs in Codex or /maintaining-continuous-handoffs in Claude.

• SKILL.md
• Concrete handoff pair example

coordinating-parallel-sessions

Coordinate N parallel agent sessions (Claude Code or Codex) from a single coordinator session - one session per independent, PR-sized workstream, each in its own git worktree and on its own agent-thread channel. The coordinator never writes the code: it scopes the streams, gates every step (plan, implement, review, UX-verify, PR, merge), adds a parallel-stream overlay to the canonical HANDOFF.md/LOG.md pair, and sequences merges to avoid shared-file conflicts. Triggers when you fan a large effort across several long-lived sessions and want each gated and merged safely - the layer on top of agent-thread. Invoke explicitly as $coordinating-parallel-sessions in Codex or /coordinating-parallel-sessions in Claude.

• SKILL.md
• Parallel handoff example

Installation

# Install everything
npx skills add gourmetpro/agent-collab

# Or install a specific skill
npx skills add gourmetpro/agent-collab --skill agent-thread

For local development, symlink a skill into your agent's skills directory, e.g. ~/.claude/skills/agent-thread -> skills/agent-thread.

Usage

Background skills auto-trigger from natural-language asks. For example:

I just wrote this spec. Have another agent review it until there are no blocking issues left, and come back to me when it's resolved.

or

Ask my other session (it has the backend repo open) what shape this API should be, and settle it with them.

The agent-thread skill activates, opens a thread, posts the opening message, and hands you a one-paste launcher for your second session. From there the two agents loop on their own until resolved.

Contributing

Each skill follows the Agent Skills open standard. See AGENTS.md for authoring conventions, frontmatter requirements, and style rules.