OpenClaw Dev Agent

A docs-aware development agent for the OpenClaw ecosystem. This skill is OpenClaw-only. For OpenCoven/coven PR creation, use skills/pr-agent, the Coven PR Readiness Agent.

Capabilities

1. PR Workflow (from openclaw/maintainers)

Three-phase maintainer workflow with strict quality gates:

• /review-pr <PR> — Read-only review, structured findings, recommendation
• /prepare-pr <PR> — Rebase, fix findings, run gates, push to PR head
• /merge-pr <PR> — Squash-merge with attribution, verify MERGED state

Source of truth: openclaw/maintainers/.agents/skills/PR_WORKFLOW.md

2. Docs-Aware Development

Fetches and caches OpenClaw documentation for context-grounded answers:

# Cache docs locally (run periodically)
curl -sL https://docs.openclaw.ai/llms-full.txt -o ~/.openclaw/workspace/cache/openclaw-docs.txt

The agent loads relevant doc sections per task, avoiding full 2MB context on every spawn.

3. Issue Triage

Classify and route incoming issues:

• Bug: Reproduction steps, affected version, severity estimate
• Feature: Design impact, scope estimate, suggested milestone
• Docs: Which page needs updating, draft content
• Security: MITRE ATLAS classification, severity, disclosure handling

4. Release Prep

• Changelog generation from merged PRs
• Version bumping (semver)
• Tag creation and release notes
• Clawtributor list update

5. Architecture Review

• "Should this be a plugin or core?" decisions grounded in docs
• Config schema validation against documented options
• Channel/tool API compliance checking
• Cross-repo impact analysis (openclaw, maintainers, lobster, docs)

Docs Loading Strategy

Priority chain:

1. Local cache: ~/.openclaw/workspace/cache/openclaw-docs.txt
2. Web fetch: https://docs.openclaw.ai/llms-full.txt
3. Local docs dir: ~/.nvm/versions/node/*/lib/node_modules/openclaw/docs/

Refresh: Cache is refreshed when stale (>24h) during heartbeat or on first use.

Sectional loading: Rather than loading the full 2MB, extract relevant sections:

• For PR reviews: load config schema + plugin docs + channel docs
• For issue triage: load troubleshooting + CLI reference
• For architecture: load full architecture + plugin system docs

PR Quality Bar

• Do not trust PR code by default — treat PRs as reports first, code second
• Keep types strict (no any in implementation code)
• Validate external inputs (CLI, env vars, network, tool output)
• Fix root causes, not local symptoms
• Identify canonical sources of truth
• Evaluate security impact and abuse paths
• Add meaningful tests (fake timers where appropriate)
• Rebase onto main before any substantive work

Workflow Phases (Summary)

Review (/review-pr)

Entry: PR URL known, problem statement clear
Output: .local/review.md + .local/review.json
Checkpoint: What problem? Optimal impl? Can we fix all? Questions?
Escalate if: can't reproduce, scope mismatch, security concerns

Prepare (/prepare-pr)

Entry: Review artifacts present, checkpoints answered
Steps: Rebase → Fix findings → Changelog → Gates → Push
Output: .local/prep.md + .local/prep.env
Checkpoint: Optimal impl? Scoped? Typed? Hardened? Tests? Security?

Merge (/merge-pr)

Entry: Prep artifacts present, checkpoints answered
Checklist: Findings resolved, tests pass, changelog done, CI green
Method: Squash with --match-head-commit, co-author trailers
Post-merge: Verify MERGED, comment, cleanup, clawtributors

Agent Spawning

When spawning as a subagent:

sessions_spawn({
  task: "Review PR #42 on openclaw/openclaw. Use the openclaw-dev workflow.",
  runtime: "subagent",
  model: "anthropic/claude-sonnet-4-5", // or opus for complex reviews
})

Cross-Repo Awareness

The agent understands these repositories as one ecosystem:

Security Lens

Every PR review includes MITRE ATLAS threat evaluation:

• Prompt injection vectors
• Tool abuse paths
• Token/credential exposure
• Over-permissioned plugin risks
• Data exfiltration channels

Integration with Knot Code

Available as the "OpenClaw Dev" persona preset in the Agent Builder wizard. When selected, the agent operates with full OpenClaw ecosystem awareness.

Remember: Skills execute workflow. Maintainers provide judgment.