coder is a coding-focused OpenClaw sub-agent for background code execution, test-driven edits, bug fixing, small project scaffolding, and small-to-medium data-analysis tasks.
This repository now includes a ClawHub-ready wrapper skill at clawhub/coder-openclaw-agent/.
Use that folder as the publication entrypoint for ClawHub. The root repository remains the source package for:
- the OpenClaw prompt pack;
- the sandbox Docker image definition;
- the integration contract and runtime documentation.
ClawHub publishes skills under platform-wide MIT-0 terms. The repository itself keeps its own source license at the root.
For ClawHub-specific packaging and publication notes, see clawhub/README.md.
- A hardened
workspace-coderprompt pack for OpenClaw. - A Docker sandbox image definition for reproducible Linux-based execution.
- A practical document/data-processing runtime for HTML, PDF, spreadsheets, and office-style files.
- A minimal integration contract for wiring the agent into
openclaw.json.
Supported:
- code execution and verification in the OpenClaw sandbox;
- bug fixing and test-driven edits;
- small project scaffolding;
- small-to-medium data-analysis tasks;
- HTML parsing and common document-processing tasks;
- honest blocked-state reporting through
PARTIALorFAILURE.
Out of scope:
- large monorepo refactors without tuning;
- long-running background jobs;
- multi-agent recursion;
- unsupported language/toolchain stacks unless you extend the image yourself.
- OpenClaw 2026.3.x or later.
- Docker available on the host.
- An authenticated
openai-codexprovider profile in your OpenClaw installation.
- Clone or copy this repository to any local path.
- Copy
openclaw/workspace-coder/into your OpenClaw base directory, or point your agent config at that path directly. - Build the sandbox image:
docker build -f docker/coder-sandbox.dockerfile -t coder-sandbox:latest .
- Register the agent in
openclaw.jsonusingopenclaw/agent-config.template.jsonas a starting point. - Verify model access:
openclaw models status --agent coder --probe --probe-provider openai-codex --json
- Verify sandbox wiring:
openclaw sandbox explain --agent coder
- Run a first smoke task:
openclaw agent --agent coder --json --message "Return strictly valid JSON matching coder SOUL schema. GOAL: create /tmp/coder/smoke/main.py that prints hello. INPUTS: none. CONSTRAINTS: work only in /tmp/coder/smoke; use python3 and Linux/bash commands only; use PARTIAL if blocked. SUCCESS CRITERIA: python3 /tmp/coder/smoke/main.py prints hello. DELIVERABLES: codeblocks and sandbox_log."
The agent expects these injected prompt files:
AGENTS.mdSOUL.mdTOOLS.mdIDENTITY.mdMEMORY.mdUSER.mdHEARTBEAT.md
- Runtime source of truth:
openclaw/workspace-coder/SOUL.md - Sandbox image:
coder-sandbox:latest - Runtime style: Linux/bash only,
python3, no provider CLIs from shell - Working directory policy:
/tmp/coder/<task_name>/ - Output schema: strict JSON with
status,task_summary,iterations_used,steps_total,steps_completed,steps_blocked,codeblocks,sandbox_log,self_corrections, anderror_analysis - Main-agent prompt and orchestration rules: see
openclaw/main-coder-prompt.md
See docker/RUNTIME.md for the preinstalled tools and Python packages included in the default image.
Recommended checks before publishing changes:
openclaw models status --agent coder --probe --probe-provider openai-codex --jsonopenclaw sandbox explain --agent coder- at least one successful real coding smoke task through
openclaw agent --agent coder --json - one blocked-input task confirming honest
PARTIALorFAILUREbehavior - ClawHub structure check:
clawhub/coder-openclaw-agent/SKILL.mdexists, uses YAML frontmatter, and does not require Windows-specific local paths - ClawHub docs smoke: a new OpenClaw user can follow
clawhub/coder-openclaw-agent/SKILL.mdplus thisREADME.mdwithout reading private memory/session files
See docs/known-limits.md.
See docs/compatibility.md.
MIT