English | 日本語
- Handover: At moruku's request, Kurotchi has taken over maintenance and development from Muse. Moruku retains the final decision on PR submission.
- Running today: Phase 1 (issue collection and triage) runs on midorun (OCI) via a systemd user timer at 06:30 JST daily. The GitHub Actions collection workflow remains dry-run only and is not the production scheduler.
- Scope: Mattermost, Grafana, Prometheus, Brave, and OpenClaw — all five targets are
collect-only. Automated patch attempts and PR creation are not enabled. - Observation checkpoint complete (read-only coordinator verification): November 8–11 daily collections succeeded, with 0 new issues, 0 triage decisions and 0 pending decisions each run. The 363 historical candidates include 319 OpenClaw items last seen November 1; this is stored history, not fresh task selection or authority to act on an upstream issue.
- Development: PRs through #10 are owner-merged. Engine lifecycle and local review-report increments each passed three synthetic main-CI cases after owner merge. PR #10 fixed report-plan serialization; run
42964537214passed exactly three report cases (0.17s). These results do not verify real Docker isolation or repair the local approval-binding failure. The next commit-message compliance helper checks Conventional Commits syntax and required final DCO trailers; its three new cases are prepared but not executed. Aggregate compliance and upstream submission remain unimplemented. - Next: Real test-runner evidence/execution adapters, staged-generation integration, source/dependency acquisition and actual isolation acceptance remain unfinished. Trusted private/disjoint directory FDs are caller-supplied capabilities, not proof of container binding. No real provider, Docker backend or production CLI activation. Overall initial-version completion is provisionally 35% (collection → isolated correction → verification → human-review report → approval-gated draft PR). This is a functional estimate, not a file/test-count calculation. See filesystem staging / 日本語 and snapshot contract.
Digest filenames use UTC dates: the November 8 run at 06:30 JST produced digest-2026-11-07.md. The previous-day filename does not indicate a missing run.
Details (Japanese): Operations and handover / Phase 2 implementation plan.
The sections below describe the overall project goals and design. For the currently implemented scope, see this section and “Status”.
An agent that helps a human contribute to open-source projects — without ever becoming a source of low-quality "AI spam" pull requests.
It watches selected OSS repositories (for example Mattermost, Grafana, Prometheus), collects suitably scoped issues, tries to fix them in an isolated sandbox, and — only when the project's full test suite passes locally — prepares a draft pull request for a human to review and submit. The agent never opens a non-draft PR and never merges anything.
LLM-generated pull requests are flooding OSS projects. Many are unverified, out of scope, or ignore each project's contribution rules, and maintainers are closing them on sight. This project takes the opposite position:
- A patch that has not passed the project's own tests does not exist.
- A diff that violates the project's CONTRIBUTING.md does not ship.
- A human sees every diff before it becomes a pull request.
The goal is a small number of careful, verified contributions — not volume.
- Collector — polls the GitHub API for open issues carrying
watched labels (
good first issue,help wanted,bug,documentation) on the configured target repositories. - Triager — scores each issue. A deterministic rule-based scorer runs first; only the borderline band goes to an LLM for a second opinion. Broad, vague, or unreproducible issues are dropped.
- Dedupe & caps — a persistent ledger records every attempted issue so nothing is tried twice. Existing PRs (human or bot) for the same issue, per-project cooldowns, and weekly caps are checked before any work starts.
- Workspace Isolator — clones the repository into a throwaway Docker sandbox with no credentials inside, restricted network egress, and CPU/memory/time limits.
- Coder Loop — TDD: first reproduce the issue with a failing test, then apply the smallest fix, re-running tests until they pass or the attempt budget (default 5) runs out. If the issue cannot be reproduced, the attempt stops — no guessing.
- Compliance Guard — runs the project's linters and formatters, validates commit messages (e.g. Conventional Commits) and the DCO sign-off where the project requires it.
- Review Report — summarizes the change, the tests that ran, and the diff, and delivers it to the human operator (daily digest).
- Draft PR — only after the human approves the report, a draft PR is created from the project's PR template, including an AI-assistance disclosure line.
- Tests first. A draft PR is prepared only if the project's test suite passes 100% in the local sandbox.
- Human-in-the-loop, always. The agent creates draft PRs only, after explicit human approval. It never submits a ready-for-review PR and never merges.
- Per-project rules win. CONTRIBUTING.md, the PR template, CLA/DCO requirements, and lint/commit conventions of each target project are loaded per project and obeyed.
- Fail closed. On any uncertainty — unreproducible issue, flaky tests, unclear rules — the agent skips, logs the reason, and moves on.
- Respect maintainers' time. Conservative caps: at most 1 draft PR per day globally and 2 per project per week by default.
| Project | Languages | Notes |
|---|---|---|
| mattermost/mattermost | Go, TypeScript/React | Collection only; docs and small server issues first |
| grafana/grafana | Go, TypeScript | Collection only; narrowly scoped packages only |
| prometheus/prometheus | Go | Collection only; single-language, Makefile-driven |
| brave/brave-browser | C++, JavaScript, TypeScript | Collection only; source code lives in brave/brave-core |
| openclaw/openclaw | TypeScript | Collection only; watches good first issue only |
Targets live in configs/targets.yaml. Each target carries a
mode: with collect-only, patch attempts and PR creation are
impossible for that target regardless of any other setting. For
Phase 1 the five targets above are enabled: true with
mode: collect-only. Moving a target to mode: full is a human
decision, made only after verifying that project's rules.
configs/ Target projects and global agent rules (YAML)
docs/ Architecture and design documents
scripts/ CLI entry points (single-issue run, scheduled watch)
src/collector/ GitHub client and triage scoring
src/sandbox/ Docker sandbox and git workspace management
src/agent/ Coder loop (TDD), prompts, tools
src/review/ Compliance guard, review report, draft PR builder
tests/ pytest suite
.github/workflows/ CI and the daily triage schedule
Setup (the only runtime dependency is PyYAML; pytest for the tests):
pip install -e . # or: pip install pyyaml
pip install pytest # only when preparing an authorized test runLocal verification passed exactly three new filesystem import/staging cases once. Prior lifecycle, CoderLoop and snapshot trials are each sealed at 9/10 executions, not re-run. The distinct filesystem feature trial currently uses local3/10; one PR3 and eventual main3 would bring it to 9/10. PR/main CI selects the same new three cases; feature pushes do not duplicate CI. Historical sources stay, but no historical suite, matrix, scheduled tests or retries run automatically. The owner cap is 10 total case executions per development trial, counting every local/CI/re-run. Selection and receipts are in the Phase 2 plan.
Run one collection pass over all enabled targets:
# Optional: relaxes rate limits. Public repos are readable without it.
export GITHUB_TOKEN=<your token>
python3 scripts/watch_repos.py --onceResults land in the ledger at state/ledger.sqlite3 (git-ignored)
and in the day's digest at reports/digest-YYYY-MM-DD.md
(git-ignored). Re-running never re-triages an issue that already has
a recorded decision.
LLM triage for borderline issues is enabled through environment
variables. When unconfigured, borderline issues wait in the explicit
pending_llm state — never guessed at — and are judged by the first
run that has a provider configured:
export LLM_TRIAGE_PROVIDER=openai # or: anthropic
export LLM_TRIAGE_MODEL=gpt-4o-mini # optional (provider default)
export OPENAI_API_KEY=<your key> # the provider's key
python3 scripts/watch_repos.py --oncepython3 scripts/watch_repos.py --daemon polls periodically
(60 minutes by default, --interval-minutes). While a target is
collect-only, the single-issue entry point
(scripts/run_agent.py) refuses patch attempts for it.
The current scheduler is a systemd user timer on midorun, running daily at 06:30 JST to collect and triage issues and save a digest. It does not attempt patches or create PRs. The GitHub Actions workflow remains dry-run only and is not used for production collection. An optional watch daemon mode polls more frequently. See docs/architecture.md for the full design. Background research notes are collected in docs/references/index.md.
Phase 1 implemented (v0.2) — collector + triage are working.
What exists today: the full design documents, configuration
schemas, issue collection over the GitHub API, the SQLite ledger, a
working deterministic triage scorer with tests, a provider-pluggable
LLM triage stage for borderline issues (borderline issues stay
pending_llm when no provider is configured), and the daily digest.
The five targets are enabled for collection only: patch attempts
and PR creation are impossible, both by configuration and by
construction. The sandbox lifecycle contract is merged; the injection-only CoderLoop is implemented and locally mock-tested. Unconfigured execution refuses explicitly. The caller supplies protected reproduction/full relevant-suite commands and trusted model/workspace adapters; automatic exploration and test authoring remain unfinished. Unknown usage/application, infrastructure failures and unconfirmed cleanup cannot produce success or automatic retry. CPU/memory, offline non-root execution, bounded capture and in-flight deadlines remain adapter obligations, not verified isolation. The bounded memory-only SnapshotWorkspace now implements validated read/literal-search/context/atomic patch/diff operations with explicit editable source and immutable test/runner roles. A synthetic integration verifies source changes drive reproduction outcomes through the real control classes. Directory-FD import and verified new-directory staging now support explicit, trusted filesystem capabilities and are locally tested. No Docker, real provider or host-shell fallback is supplied; arbitrary Path/Git helpers, actual runner adapters, staged-generation orchestration and production CLI integration remain unfinished.
- Phase 0 — Design and scaffold (done).
- Phase 1 — Collector + triage: issue collection, ledger, LLM triage for borderline issues, daily digest (done; targets enabled for collection only).
- Phase 2 — Sandbox + coder loop, exercised on documentation and typo issues only.
- Phase 3 — Compliance guard + draft PR submission behind the human approval gate.
- Phase 4 — Scheduled daily operation; targets enabled one at a time, caps raised only on evidence of maintainer acceptance.
MIT — see LICENSE. Note that contributions produced with this tool must additionally comply with each target project's own license and contribution rules.