oss-radar

Third sibling in the orchestrator family alongside autodev and agent-radar.

Different mission. autodev and agent-radar create new repos under the user's account. oss-radar contributes PRs to OTHER PEOPLE'S open-source repos — small, high-quality contributions to active AI startup projects, with the goal of building a track record of merged PRs that recruiters can verify.

Status: working MVP. Discovers candidate repos, runs an Opus-free Sonnet-only agent chain to research → plan → implement → review → submit. Cap: 2 contribution attempts per week. AI involvement is always disclosed in the PR body.

---

Why this is harder than autodev

In autodev, Claude builds something from scratch in an empty workspace. Here, Claude has to:

• Read someone else's codebase and understand it
• Match their style, conventions, and patterns
• Make a surgical change without scope creep
• Run their test suite and pass
• Write a PR body the maintainer actually opens
• Disclose AI involvement honestly

The merge-vs-reject ratio is fundamentally outside our control: maintainers are humans with opinions and time constraints. Quality matters more here than anywhere else in the orchestrator family.

Architecture

  ┌─────────┐    ┌──────────┐    ┌────────┐    ┌──────────┐
  │  scout  │───▶│researcher│───▶│ planner│───▶│implementer│
  └─────────┘    └──────────┘    └────────┘    └─────┬────┘
                                                      │
       ┌──────────┐    ┌──────────┐    ┌──────────┐  ▼
       │ improver │    │ pr_writer│◀───│ reviewer │
       │ (weekly) │    └────┬─────┘    └──────────┘
       └──────────┘         │
                             ▼
                  ┌──────────────────┐
                  │  publisher       │
                  │  fork → clone →  │
                  │  branch → push → │
                  │  open upstream PR│
                  └──────────────────┘

Components

File / dir	Role	Model
scout.py	GitHub Search API for AI startup-tier repos with good-first-issue	none (HTTP)
agents/researcher.py	Read CONTRIBUTING, detect anti-AI policy / CLA, pick ONE issue	Sonnet 4.6
agents/planner.py	Plan technical approach (read-only)	Sonnet 4.6
agents/implementer.py	Make actual code changes in the forked clone	Sonnet 4.6
agents/reviewer.py	Pre-PR self-review (scope, AI-tells, tests pass)	Sonnet 4.6
agents/pr_writer.py	Compose PR title + body + AI disclosure	Sonnet 4.6
agents/improver.py	Weekly self-improvement, edits agent prompts	Sonnet 4.6
publisher.py	Fork + clone + branch + push + open upstream PR	none (API)
contributor.py	Stitch agents into the contribution lifecycle	none
orchestrator.py	Tick loop, pacing, self-heal, retries	none
state.py	SQLite state at %LOCALAPPDATA%\oss-radar\	none
claude_cli.py	claude -p headless wrapper	none

Sonnet-only by design

User reserved Opus quota for primary work. Every LLM call in oss-radar runs on Sonnet 4.6 — including the improver self-improvement layer (autodev uses Opus there; agent-radar and oss-radar both use Sonnet for cost discipline).

Pacing

• 2 contribution attempts per week max (rolling 7-day cap)
• 36-hour minimum gap between PR-open events
• Skip-day probability 20%
• Commit windows: weekday 19:00-23:00, weekend 10:00-22:00
• One open PR per upstream repo at a time

Quality gates

Before opening a PR, all must hold:

1. Researcher verdict is proceed (not skipped for anti-AI / CLA / etc.)
2. Planner's estimated diff is ≤ 80 lines
3. Implementer's tests pass
4. Reviewer's verdict is pass (or fix resolved within MAX_FIX_ROUNDS)
5. PR body includes AI disclosure (non-negotiable)

If any gate fails the attempt becomes skipped with a reason logged, and the orchestrator picks the next candidate target on the next tick.

Setup + running

python orchestrator.py preflight    # token + claude check
python orchestrator.py scout        # discover candidate repos
python orchestrator.py status       # active attempt + counts
python orchestrator.py health       # full system check
python orchestrator.py improver     # force a self-improvement run
python orchestrator.py run          # loop forever

For unattended operation: install_autostart.ps1 registers a Task Scheduler entry that auto-launches on login with a 30-min heartbeat (offset 25 min from autodev so all three orchestrators alternate).

License

MIT — see LICENSE.