Skip to content

openclaw-coder.md

Latest commit

 

History

History
51 lines (36 loc) · 5 KB

openclaw-coder.md

File metadata and controls

51 lines (36 loc) · 5 KB

OpenClaw Coder

Maintainer's map of the openclaw-coder/ subproject: a shareable packaging of OpenClaw as an autonomous AI programmer. This doc is the entry point for working on the subproject in this repo. For using it, see its own README.md.

Three layers

  1. Reference workspace — openclaw-coder/playbook-test/workspace/. The myclaw OpenClaw instance's bootstrap files (AGENTS.md, IDENTITY.md, SOUL.md, USER.md, TOOLS.md), auto-loaded into the system prompt every turn. It lives inside the harness (its only consumer) as the test fixture, and doubles as the worked example the consumer README points at. AGENTS.md is a thin pointer: on every user message it sends the agent into the openclaw-coder-playbook skill's dispatcher. The workspace carries no playbook copy — the skill is the single source.
  2. Operating-instructions playbook — the openclaw-coder-playbook skill. Its SKILL.md is the dispatcher: it routes by surface (thread → working-session.md; channel/DM → channel-handling.md) and carries the global rules (language, "tickets are labels", projects, chat_id). The procedures live in references/ (working-session.md, channel-handling.md, project-workspace-setup.md). Coding is delegated to the separate alignfirst-coaching coaching skill. Nothing here is auto-loaded — files are read on demand (see context engineering below).
  3. Regression-test harness — openclaw-coder/playbook-test/. A standalone Dockerised consumer of the published @paleo/openclaw-* packages that drives the workspace through synthetic Discord/Slack channels and judges the outcome. It bind-mounts both the workspace dir and the alignfirst-coaching skill into the gateway, so edits to layers 1 and 2 iterate live without rebuilding the image.

How a turn flows

user message
  → workspace AGENTS.md (auto-loaded)              layer 1
  → openclaw-coder-playbook/SKILL.md (read first)  layer 2  ← procedural dispatcher
  → references/working-session.md | channel-handling.md   layer 2
  → references/project-workspace-setup.md (if WORK)       layer 2
  → delegate coding to the alignfirst-coaching skill (coaching/CLI, read last)

Layer 1 is the only thing OpenClaw injects automatically; everything in layer 2 is pulled in by an explicit file read because nested workspace files and skill files are not auto-loaded. The dispatch skill is read first and is purely procedural; the coaching alignfirst-coaching/SKILL.md is read last, at delegation — keeping its protocol vocabulary out of the early user-facing acks (see writing-instructions-for-openclaw.md).

Reading order for maintainers

Running the suite

From openclaw-coder/playbook-test/:

cp .env.local.example .env.local   # fill ANTHROPIC_API_KEY
npm install
mkdir -p artifacts .gateway-logs   # create as your user so Docker doesn't make them root-owned
npm run env:build                  # only after image-affecting changes
npm run env:up
npm run e2e -- --channel discord-mock A1-new-work-to-be-done
npm run env:down

⚠️ Never rm -rf artifacts (or .gateway-logs). Runs are written to timestamped subdirs, so they accumulate without colliding — wiping the directory destroys prior runs for no reason. mkdir -p is enough to avoid root-owned dirs.

Scenario ids are the full filename stem (A1-new-work-to-be-done, not A1). Measure a flaky-looking assertion's true rate with --iterations N --max-failures N (raise --max-failures above its default of 1 so the matrix doesn't abort early). See writing-instructions-for-openclaw.md.

Deployment

Running a coder against real Discord/Slack (not the mock channels) is documented for consumers in openclaw-coder/README.md — the openclaw.json knobs and the AGENTS.md template. Creating the bot itself (tokens, scopes, Socket Mode) is standard OpenClaw; defer to OpenClaw's channel docs.