Open Source Contributor#

Autonomous GitHub contribution agent using the Architect-Builder pattern. The main agent handles all git operations and I/O; subagents handle only focused cognitive work.

Quality Standards#

• Every contribution must be meaningful — no typo fixes, no placeholder PRs
• Must pass CI/tests where available
• Must include proper description and tests if applicable
• Must reference the issue it fixes
• Must disclose AI assistance
• No emojis in PR titles, descriptions, or comments. Plain text only. Professional tone throughout.

⚠️ GitHub Token#

Store your GitHub token securely (e.g., ~/.openclaw/.github_token). Never log or print token values. The agent should read it silently from disk.

Excluded Repos#

Maintain a list of repos you've already contributed to (to avoid duplicate PRs). Store in your config file.

---

Architecture: The Architect-Builder Pattern#

THE CORE RULE: Subagents do NOT clone repos, push code, or make GitHub API calls. The main agent does all I/O.

Why This Pattern?#

Subagents timed out when trying to do full git workflows (clone large repos + modify files + push + create PR). Root causes:

1. Each tool call through cloud model proxies takes 5-58 seconds — 50 tool calls = 12+ minutes of pure latency
2. Large repo clones eat into the time budget
3. Context window fills up after many tool calls, causing stalls

The fix: The main agent owns all I/O and git operations. Subagents only do focused cognitive work on pre-staged content.

How It Works#

text

┌─────────────────────────────────────────────────────┐
│  MAIN AGENT — The Architect & Builder                │
│                                                       │
│  1. Research: Find issues via GitHub API              │
│  2. Stage: Clone repo, identify files, read context   │
│  3. Delegate: Send focused edit task to subagent      │
│  4. Apply: Review subagent output, apply changes      │
│  5. Push: Commit, push, create PR via GitHub API      │
│                                                       │
│  Uses: exec, GitHub API, git CLI — all I/O tools     │
└─────────────────────────────────────────────────────┘
              │
              │ Focused task + file contents (pasted inline)
              ▼
┌─────────────────────────────────────────────────────┐
│  SUBAGENT — The Specialist                           │
│                                                       │
│  Receives: Specific files + the issue + instructions │
│  Returns: Exact code changes (diff/patch/new content)│
│  Does NOT: Clone repos, push code, create PRs        │
│  Does NOT: Make GitHub API calls                      │
│  Does NOT: Browse the web                             │
│                                                       │
│  Timeout: 600s max (focused work only)                │
└─────────────────────────────────────────────────────┘

---

Difficulty Levels#

Level 1: Easy — Warm-up Contributions#

CriteriaRequirementLabelsgood first issue, help wanted, bugStars500+Issue age3-30 daysComments< 10ScopeSingle file, < 30 lines changed

Typical fixes: UI text corrections, missing error handling, simple config fixes, accessibility improvements, missing validation.

Level 2: Intermediate — Real Feature Work (DEFAULT)#

CriteriaRequirementLabelsbug, feature, performance, enhancementStars1,000+Issue age7-60 daysComments< 5ScopeMulti-file, 30-150 lines changed

Typical fixes: Dependency bumps with tests, missing API parameters, edge case fixes, missing method implementations, performance improvements.

Level 3: Advanced — Architecture & Deep Fixes#

CriteriaRequirementLabelsbug, feature, performance, enhancement, designStars2,000+Issue age7-90 daysComments< 8ScopeMulti-module, 100-500+ lines changed

Typical fixes: Race conditions, memory leaks, API redesigns, query optimization, plugin systems, cross-platform fixes.

Optional: Spawn a Council of LLMs to evaluate approach before implementing.

---

Pipeline: Architect-Builder Pattern#

Phase 1: Research — Main Agent Finds the Issue#

The main agent searches GitHub for suitable issues:

1. Use GitHub Search API or an authenticated search script
2. Filter by difficulty level criteria (stars, age, labels, comment count)
3. Exclude already-contributed repos
4. Verify no open PRs already address the issue
5. Select the best candidate

Phase 2: Stage — Main Agent Prepares the Workspace#

The main agent does ALL the I/O work directly (no subagent):

1. Fork the repo via GitHub API
2. Shallow clone: git clone --depth 1
3. Create branch: fix/<issue-number>-<short-description>
4. Read CONTRIBUTING.md and relevant source files
5. Identify the files that need to change
6. Read the full content of those files
7. Prepare a focused task for the subagent with all file contents pasted inline

Phase 3: Implement — Subagent Does Focused Cognitive Work#

Spawn a specialist subagent with:

• runTimeoutSeconds: 600 (10 min max — focused work only)
• lightContext: true
• All file contents PASTED INLINE in the task description
• Clear instructions: "Return the EXACT modified file content" or "Return a unified diff"

The subagent task MUST include:

• The exact files to modify (full content pasted inline)
• The issue being fixed (full description)
• The project's style/lint requirements (from CONTRIBUTING.md)
• A request for EXACT code changes (not a plan, the actual code)

The subagent does NOT:

• Clone repos
• Run git commands
• Make GitHub API calls
• Browse the web
• Install packages

Phase 4: Apply & Push — Main Agent Finishes the Job#

The main agent takes the subagent's output and:

1. Applies the code changes to local files
2. Runs linters/tests if available
3. Fixes any remaining issues (lint errors, test failures)
4. Commits, pushes, and creates the PR via GitHub API
5. Returns the final summary

Phase 4.5: Council Review (Level 3 only — optional)#

For Level 3 fixes that are security-sensitive or architecturally complex, spawn a Council of LLMs BEFORE Phase 3. Use 3 different models as Strategos, Analyticos, and Creativos with distinct analytical perspectives. If consensus is "don't implement" → skip, move to next candidate.

---

API-First Approach (Preferred for Simple Fixes)#

For simple fixes (Level 1, some Level 2), skip cloning entirely and use the GitHub API directly:

1. Read files via GET /repos/{owner}/{repo}/contents/{path}
2. Send modified content via PUT /repos/{owner}/{repo}/contents/{path}
3. Create PR via POST /repos/{owner}/{repo}/pulls

This eliminates clone time entirely and completes in 2-5 minutes.

Use API-first when:

• Fix touches 1-3 files
• Files are under 500 lines each
• No test suite needs to run
• No complex branching required

Use clone when:

• Fix touches 4+ files or needs cross-file understanding
• Need to run tests/linters locally
• Need to understand project structure deeply

---

Configuration#

Store your settings in a config file (e.g., contrib-scout/config.json):

json

{
  "difficulty_level": 2,
  "github_token_path": "~/.openclaw/.github_token",
  "max_contributions_per_night": 1,
  "languages": ["python", "javascript", "typescript", "go", "rust"],
  "excluded_repos": ["your-org/your-repo"]
}

Council Models (Optional)#

For Level 3 Council review, configure your preferred models in ~/.openclaw/council-config.json:

json

{
  "council_models": [
    "your-first-model-here",
    "your-second-model-here",
    "your-third-model-here"
  ],
  "default_timeout": 900,
  "max_tokens": 8192
}

Example (using Ollama cloud models):

json

{
  "council_models": [
    "ollama/kimi-k2.6:cloud",
    "ollama/deepseek-v4-pro:cloud",
    "ollama/gemma4:31b-cloud"
  ]
}

Choose models with different strengths — one strategic, one analytical, one creative. The more diverse, the better the council output.

---

Critical Rules#

Timeout Rules#

• Subagents: max 600s (10 min) — they only do focused cognitive work, not I/O
• Main agent has no timeout — it handles all long-running I/O directly
• Never set runTimeoutSeconds above 600 for subagents — if they need more, the task is too broad

Subagent Constraints#

• Subagents do NOT clone repos, push code, or create PRs
• Subagents do NOT make GitHub API calls
• Subagents do NOT browse the web
• Subagents receive file contents inline and return code changes
• Keep subagent task descriptions under 1500 words
• Use lightContext: true on all subagent spawns

Quality Rules#

• Never force a contribution — if no good fit, report "no suitable issues found"
• Every PR must reference the issue it fixes
• Every PR must disclose AI assistance
• No emojis in PR titles, descriptions, or comments
• Professional, concise tone in all PR communications

Workflow Rules#

• API-first for simple fixes — no clone needed for 1-3 file changes
• Clone only when you need to run tests or understand complex structure
• Batch tool calls — read all files first, then send to subagent in one task
• Verify before push — run linters/tests locally after applying changes

---

Common Failure Modes#

SymptomCauseFixSubagent times outTask too broad (clone + edit + push)Use Architect-Builder pattern: main agent does I/OShallow code changesVague subagent promptPaste full file contents + specific instructionsLint failures after subagent editSubagent didn't know lint rulesInclude lint config in the prompt or run lint before subagentPR rejected by maintainerLow-quality contributionFollow CONTRIBUTING.md, write meaningful fixes onlyContext overflow in subagentToo much data pasted inlineSummarize to <1500 words, paste only relevant files

Companion Skills#

• subagent-orchestration — Required. Provides spawn patterns, timeout config, sandbox constraints.
• council-of-llms — Optional. For Level 3 complex decisions before implementing.

Storage#

text

contrib-scout/
├── repos/              # Cloned repositories (cleaned up after each run)
├── logs/               # Activity + audit trail (JSONL)
├── config.json         # Difficulty level + settings
└── nightly-report.json # Daily summary