autoclaw

Local-first intelligence layer that learns from past AI coding sessions, does RAG over your codebase, and cuts token waste. Trigger on "learn from my sessions", "index my code", "retrieve context", "/learn", "/index-code", "/retrieve", or "intelligence layer".

Multi-agent parallel development orchestrator. Reads task manifests, builds dependency DAGs, generates sprint plans, assigns scoped work to parallel agents, and coordinates review gates. Trigger on "/orchestrate", "plan sprints", "parallel agents", "assign sprint", or "orchestrate tasks".

Multi-agent parallel development orchestrator. Reads task manifests, builds dependency DAGs, generates sprint plans, assigns scoped work to parallel agents, and coordinates review gates. Trigger on "/orchestrate", "plan sprints", "parallel agents", "assign sprint", or "orchestrate tasks".

Autonomous scheduled build workflows and pipelines. Trigger on "/autobuild schedule", "run workflow", "automate build", or "schedule task".

Keeps user-facing docs in sync with public-API changes. Triggered by /persona doc-writer and auto-dispatched on a task_complete whose diff touches a public API (exported types, command contributions, MCP tools, CLI flags). Writes only docs + CHANGELOG; never code. Reads its persona memory so doc conventions accumulate. Local-first provider with cloud fallback.

Persistent always-on background agent with automatic memory consolidation. Trigger on "start background agent", "enable kdream", "/kdream start", "persistent daemon", or "auto-dream memory".

Audits a target module for security defects against a seeded pattern set, writes a structured finding report with severity + disposition, and gates security-tier merges. Triggered by /persona security-auditor, by /sprint when a task touches auth/crypto/network/secrets, and before any GA flip. Reads prior findings from its persona memory so a re-audit picks up where the last left off. Local-first provider with cloud fallback.

Onboards the Hermes personal-assistant platform as a typed fabric worker (agent_type assistant, human-in-the-loop) and directs assistant tasks to it through the runner-hermes adapter. Hermes is a long-lived REST task service (HERMES_ENDPOINT) — AutoClaw submits a task with POST /tasks, polls its status, and reads the result; nothing Hermes drafts or schedules takes effect without a human confirming it. Default trust is off (Hermes autonomy "manual"). No direct LLM calls — the Hermes service is the agent host.

Onboards the OpenClaw platform as a typed fabric worker (agent_type coder) and directs repo work to it through the runner-openclaw adapter. OpenClaw is a hybrid host — REST endpoint when OPENCLAW_ENDPOINT is set, otherwise the openclaw CLI — that runs coding tasks under its own trust levels (gated/supervised/unattended) and mints its own job IDs, which the adapter maps back to AutoClaw task IDs. Triggered when the orchestrator routes a coder-typed task to the openclaw runner. Local-first; no direct LLM calls — OpenClaw is the agent host.

Owns the canonical RFC index and writes new RFCs against the existing decision record. Triggered by /persona architect, by /sprint when an unresolved cross-cutting design question is detected, and after any merge that touches docs/rfc/. Inputs prior RFCs from skills/architect/bibliography.md and writes specs to docs/specs/<feature>/. Defaults to local Ollama (per v3.1 governance) with cloud fallback.

Six short rules for running an autonomous loop well. Loaded at the start of any perpetual loop, watch-mode tick, or /loop dispatch. Keeps the loop's signals clean and prevents over-supervision.

Multi-agent coordinator that spawns specialized sub-agents working in parallel. Trigger on "/mateam launch", "spawn agents", "multi-agent", or "coordinate team".