swarmspace

investigator-swarm.md

Use for parallel read-only investigation across a narrowed scope (repos, files, or service interactions). Peer agents explore in parallel, produce structured findings with cited evidence, and do NOT modify anything. The right tool when you need to diagnose a bug, locate a code path, or gather evidence — not when you need to apply changes. Expects an orchestrator (or human) to have written memo keys `goal` (what is being investigated) and `scope` (which repos/files/services are in play) before launch. Triggers when: diagnosing cross-repo bugs with a known scope, locating the source of an incident, gathering evidence before a refactor, producing a triage report. Examples: "find which service emits the malformed X-Request-ID header across these 6 repos", "locate the code path that handles stale sessions in auth-service + user-service", "identify which repos still use the deprecated v1 auth middleware".

swarm-agent.md

Use this agent for complex multi-repo tasks that benefit from peer coordination, shared state, and self-organized work splitting. Multiple instances of this agent can run simultaneously on the same workspace — they coordinate as equals via memo and journal. They observe independently, post proposals, reach quorum before acting, then split work by claiming. No hierarchy, no designated leader. Triggers when: working across multiple repos simultaneously, running parallel agents on a workspace, needing emergent work distribution. Examples: "refactor auth across all services", "coordinate API and frontend changes", "run a migration across the workspace".

synthesizer.md

Use when the user is working with a multi-repo **workspace directory** (a parent dir holding several cloned repos side by side) and wants routine multi-repo git or shell operations — cloning a set of repos together, pulling them all, checking branch/dirty/ahead-behind status across everything, running the same command in all or some of them, or doing an advisory pre-action check before touching shared branches. Triggers on phrases like "clone these N repos", "pull all my repos", "status across all the services", "run X in every repo", "run this in just auth and billing", "which repos are dirty", "check before I push". This skill teaches the orchestrator (you) the `ws workspace|status|preflight|pull-all|each|some` commands, their quoting rules, and the safety boundaries between read-only polling and fan-out execution. Use this skill for routine multi-repo ops; for parallel *investigation* or coordinated *refactors*, hand off to `orchestrating-swarms`.

Use when a task spans multiple repos or can be split into N independent scopes that an agent could chase in parallel — diagnosing a cross-service bug, locating a code path across several services, auditing who still uses a deprecated API, running a coordinated refactor or migration across a workspace, or any question that starts with "across these repos…". Triggers on phrases like "across all our services", "which repo emits…", "find where X happens in any of…", "refactor Y everywhere", "who still uses…", "parallel investigation", "swarm". This skill teaches the orchestrator (you) how to launch `ws investigate` (read-only evidence-gathering) or `ws swarm` (executor peers) with the right pre-launch contract, peer count, backend CLI, and post-run synthesis — instead of sequentially grepping repos one at a time or spawning a single Task agent for work that is embarrassingly parallel.

2 hooks across 2 events