mesh_skill

Overview

This skill coordinates a guarded swarm of subagents to complete tasks from a persisted $st plan, activated only when the user includes the literal token $mesh. It optimizes solution quality via propose/critique/synthesize/vote cycles, enforces consensus and validation before applying changes, and persists learnings with $learnings. The orchestrator never implements task code directly—workers do the work and produce patches as diffs.

How this skill works

When invoked with $mesh, the orchestrator resolves the plan file and runtime adapter, selects runnable tasks, and emits a one-line preflight. It fans out role workers (proposer, critics, skeptic, synthesizer) using a runtime adapter, mediates artifact passing between rounds, collects votes, and applies accepted unified diffs. Integration only occurs after consensus, validation commands pass, and the orchestrator persists task state and learning artifacts.

When to use it

• You need structured, review-driven code changes from an automated multi-agent swarm.
• You want enforced consensus and automated validation before mutating a plan or repository.
• You must coordinate multiple concurrent workers while avoiding merge/branch risks.
• You want durable lessons recorded for future runs via $learnings.
• You need to drain or batch $st plan tasks with controlled parallelism.

Best practices

• Always include the literal $mesh token and any overrides (ids, plan_file, max_tasks, parallel_tasks).
• Provide well-scoped mesh metadata in task notes: scope, acceptance, validation commands. Hydrate metadata when missing.
• Prefer adapter=auto or explicit adapter choice and confirm adapter capabilities (fanout, collect, retry, follow_up, close).
• Avoid merging to protected branches; open PRs or provide explicit merge instructions only when asked.
• Use parallel_tasks=auto with the provided capacity math and reserve slots to avoid worker saturation.

Example use cases

• Run $mesh to complete one ready $st task with a full 5-role swarm and then integrate the patch after validation.
• Drain all ready tasks: $mesh max_tasks=auto parallel_tasks=auto (the skill explains chosen parallelism).
• Target specific tasks: $mesh ids=st-003,st-007 to operate on a transitive dependency closure.
• Headless CI: $mesh adapter=auto headless=true to run non-interactively; it exits with one actionable line if prerequisites are missing.
• Scale-out mode example: run a fleet with one integrator and N workers for heavy parallel workloads (requires mailbox+leases coordination).