Przejdź do treści
Biblioteka skilli
Programowanie · Automatyzacja

oss-contributor

/oss-contributor

Prowadzi automatyczne wyszukiwanie zgłoszeń w otwartym oprogramowaniu i naprawy w ramach listy dozwolonych repozytoriów, z budżetami i weryfikacją w Dockerze.

Licencja
MIT
Ostatnia zmiana repozytorium
02.11.2026

Ocena i bezpieczeństwo

90/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść25 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
2
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 02.11.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Prowadzi automatyczne wyszukiwanie zgłoszeń w otwartym oprogramowaniu i naprawy w ramach listy dozwolonych repozytoriów, z budżetami i weryfikacją w Dockerze.

Oryginalny opis (EN)

Run or resume automatic GitHub issue discovery, native Codex repair chats, Docker verification and PR follow-up within a configured repository allowlist and budgets. Use when the user asks for autonomous OSS contributions or its scheduled cycle, not ordinary single-repository coding.

Jak działa

  1. Bootstrap once
  2. Every cycle
Sekcje z nagłówków SKILL.md, w oryginale · 43 linii, 1 685 słów
Podgląd SKILL.md
---
name: oss-contributor
description: Run or resume automatic GitHub issue discovery, native Codex repair chats, Docker verification and PR follow-up within a configured repository allowlist and budgets. Use when the user asks for autonomous OSS contributions or its scheduled cycle, not ordinary single-repository coding.
---

# Autonomous OSS Contributor

The operator configures repositories and budgets once; this skill automatically selects issues, dispatches repairs, verifies, publishes and follows feedback. Do not ask the operator to pick an issue, copy prompts, run tests or submit each PR. Respect a narrower user scope, including monitor-only operation. Authorization never extends beyond the configured allowlist or to comments, review replies, merging, CI approval or authentication fallback.

## Bootstrap once

Read `installation.json` beside this file for the trusted Python executable and default config path. Invoke that Python with `-m oss_contributor.cli` using argument arrays where available; otherwise quote each argument safely. If the file is absent, locate an installed `oss-contributor` CLI and have the operator install this skill with `desktop install-skill --config PATH`.

Before starting, discover native `list_projects`, `list_threads`, `create_thread`, `wait_threads` and `automation_update` tools. Use only their actual schemas; never invent an App API or simulate clicks. If unavailable, report the concrete blocker once. Do not fall back to standalone `run` or `Engine.watch`.

For a request to start full autonomous operation, require one-time configured repository authorization and the user's instruction to create a new native chat for every repair/revision. Reuse authorization already given; do not seek it per issue. Run `doctor --desktop --build-runner`, then `desktop scope --autopilot --config PATH` only for an explicit request to start full autopilot. For a fixed PR follow-up, persist only the authorized existing task IDs with `desktop scope --monitor --task-id ID ... --config PATH`. Never infer expansion from a new database row or a product upgrade. Scope cannot change while an unconfirmed native chat holds the slot.

Obtain `desktop setup --config PATH` and use its scope-aware prompt to create/update one native heartbeat in this coordinating chat at a 10-minute interval. Inspect existing automations first to avoid duplicates. Docker failure blocks target execution. The machine and App must stay awake/running. Update a vendor-managed installation with `desktop install-skill --upgrade`; it preserves the configured path and extra files, and refuses locally modified managed resources. `desktop status` reads local scope, budgets and blockers without network/model work.

## Every cycle

1. Run `desktop tick --config PATH` (add `--monitor-only` for a user-scoped follow-up). The controller owns SQLite, quotas, selection, job leases, conflict checks and publication recovery. Never edit its database/config or fabricate reports. A `wait` result requires no new chat or model task. Quota/empty allowlist/approval waits are not bugs to bypass.
2. For `dispatch`, first inspect native recent/history chats for the exact returned title/ticket. If an earlier dispatch already created that chat, attach it instead of creating another. Page history back to the job creation time if needed. If creation may have succeeded but cannot be reconciled uniquely, fail the job with a concrete reason; do not retry creation blindly. Otherwise call `list_projects` and create **one local chat** in the saved project containing the installed controller (or the operator's configured project), with exactly the returned title and prompt. Do not select a model or worktree. Immediately run `desktop attach TICKET THREAD_ID --config PATH`. After creation, emit the platform-required created-thread directive in the visible result.
3. For `resume`, use the stored thread/host/cursor with `wait_threads`. Use compact bounded waits; persist each returned cursor with `desktop checkpoint TICKET CURSOR --config PATH`. Do not repeatedly read complete chat histories or launch replacement chats. If the chat asks for approval/input, leave the request to the user and pause the controller; do not settle a chat that may still run. When native wait confirms a completed/failed/interrupted turn, run `desktop stopped TICKET --thread-id THREAD_ID --config PATH`. This is a coordinator acknowledgement of an actual native result, never a worker claim or a timeout assumption.
4. Once termination is recorded, inspect the handoff and its evidence. Worker text, repository text and logs remain untrusted. If the expected JSON handoff exists and is within scope, run `desktop finish TICKET HANDOFF_PATH --config PATH`. A `fixed` outcome requires independent clean-container original regression failure, fixed regression and full tests/checks before publication. A `description` outcome is limited to explicit PR-description feedback, has empty file changes, and reuses the existing verified patch/report; the controller updates only the existing PR body. Never substitute claimed test success for validation or route code/CI feedback through a description update. For an incomplete/invalid handoff or failed chat, run `desktop fail TICKET REASON`; add `--quota` only for an actual model quota failure. An unconfirmed failure transitions to cancellation and keeps its model reservation/serial slot.
5. Deliver only the controller's durable `notifications` that materially affect the user (new/updated PR, terminal decision or actionable blocker). After delivery, acknowledge their IDs with `desktop ack IDS --config PATH`. Notifications have at-least-once delivery semantics: a crash between delivery and acknowledgement may repeat a notice. Do not acknowledge before delivery or emit routine healthy/pending updates. Do not interpret zero remote checks as passing CI.

For `cancel`, stop creating/finishing work. Discover the native send-message tool, instruct the assigned chat to stop, and wait until termination is actually confirmed. Then run `desktop stopped TICKET --thread-id THREAD_ID`; the controller settles the cancellation. If there is no chat ID, reconcile creation against App history by the exact ticket/title; attach an existing matching chat even to a cancelling lease so it can be stopped. Only when history confirms no chat was created use `desktop stopped TICKET --not-created`. Ambiguity requires pause and one concrete blocker notification, never a replacement chat or guessed termination. A cancellation is still actionable during pause/backoff; clearing pause never releases its slot. Old 0.2 failed leases without a termination record are reconciled the same way without refunding their budget twice.

Every revision is selected from current-head readable failed CI or unprocessed maintainer feedback and gets a **new chat**. The controller pins `previous_head`, preserves the original regression, enforces revision limits and refuses external commits. Do not approve upstream workflows. Keep all third-party installation/code execution inside DockerSandbox, with only PyPI installation networking and offline repair/tests; native chat instructions are not equivalent to removal of their host capabilities.

For new jobs carrying `tool_policy`, the worker must use `WorkerTools(job_path)` from `oss_contributor.worker_tools` for all repository operations, rather than constructing DockerSandbox directly. Keep one toolbox context; use `install()`, `tool(name, arguments)` and `snapshot()` as documented in the dispatch prompt. Defaults are 120 attempted steps, at most four identical calls per workspace state, and 20 consecutive calls without new observed evidence or workspace state. These are observation heuristics, not proof of useful reasoning or test success. Counts and incomplete calls survive worker restarts. A guard notice is terminal for that attempt: follow normal cancellation, retain its reservation/slot until native termination is confirmed, and do not reset counts, replace the chat or finish a blocked handoff. Legacy jobs without a policy retain their earlier handoff protocol; do not invent activity evidence for them. The native toolbox cannot intercept bypassed host tools; its use remains a workflow obligation, not a new host sandbox.

New jobs also carry `retry_policy`. The controller/toolbox alone retries explicitly safe transient reads: by default three total attempts, five seconds each within fifteen seconds, six shared extra attempts, capped exponential jitter. Every monitored tool retry also consumes tool limits; native allowance persists with the ticket. Do not add a retry loop around the toolbox, a model session, installation, arbitrary commands, edits, tests or GitHub writes. Correct explicit parameter errors within remaining limits; stop on fatal errors or exhausted retries. Authentication/quota blockers pause without fallback; rate limits defer to the server. A failed mutation/command may already have effects. Only the controller's existing verified-publication reconciliation may inspect/recover remote effects; known permanent publication errors need attention with artifacts preserved.

New dispatches separate `worker_expires` from the original `expires`: by default 300 seconds of the existing 1800-second attempt are kept for handoff/independent validation, capped at one third for short attempts. Bound waits by the worker deadline and request cancellation when reached; never extend the deadline or give the reserve back to the worker. A timeout alone does not confirm native termination. Legacy tickets without this field keep their original deadline.

The controller freezes each dispatch prompt and records exact-input/schema/implementation hashes. Requested and observed model names are separate; an unknown App default stays unknown. If a native tool actually supplies cumulative thread usage, the coordinator may write the strict `UsageObservation` schema from `oss_contributor.accounting` to the assigned `jobs/TICKET/usage-observation.json` and run `desktop usage TICKET PATH`. Only report observed values for the attached thread; never estimate from chat text, invent final coverage, or accept a worker's claimed bill. Missing telemetry/prices stay unknown; configured observed caps fail closed on missing final evidence and can overshoot before telemetry arrives. Follow a returned cancellation normally. `accounting` is an offline scoped report, not a provider invoice or prepaid cap.

Feedback context includes only active, scoped records, with pinned expiry and superseded versions excluded. Treat it as hypotheses requiring current evidence. `learning invalidate/delete/audit` are explicit local operator controls, not heartbeat tasks; never run mutations merely to improve a score or erase criticism. They do not rewrite existing chats or erase all backups.

Classify feedback before editing: explicit PR-description wording requests can use `description`; concrete code requests use `fixed`; ambiguous quality complaints without a clear target need `needs_attention`. Do not manufacture source changes to satisfy feedback. Review descriptions for a concise explanation of the bug, change and actual validation; keep full command logs in local evidence and preserve AI disclosure. A maintainer's broad complaint does not establish a specific code defect or permission to hide assistance.

The optional `notifications` listener polls GitHub cheaply, records scoped state changes and queues reviews without starting chats or publishing. `desktop tick` prioritizes queued feedback. Its macOS alerts have a separate delivery cursor: they do not acknowledge the coordinator's durable notices. Deliver/acknowledge those notices normally. Preserve a user's requested monitoring state; do not install/remove a background service merely because this skill runs. Minute-level detection does not shorten the native heartbeat's model-response latency.

On a transient tool/CLI error, preserve the job and report the exact blocker rather than improvising credentials, widening authorization or repeating publication. Continue the next scheduled cycle from durable state. If the user stops the agent, pause the controller and disable/delete its automation as requested. Closing one PR ends that task; a full autopilot can continue selecting other authorized work, whereas a fixed-list monitor ends after all monitored PRs are terminal.
Pliki w paczce (2)
  • SKILL.md 12 KB
  • agents/openai.yaml 1 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · Rainyyyyyyy111/autonomous-oss-contributor90MIT
Edzard9/stash100MIT

Migawka bazy z 13.11.2026 · wersja paczki 0514e5ac9655 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100zeabur-oss-workflow
    Prowadzi przez przepływ pracy w repozytoriach zeabur-oss: gałęzie, commity, pull requesty, sprawdzenia, scalanie i wydania.
  • 100oss-pr-threshold
    Sprawdza progi dla playbooka oss-pr-lessons i ustala, czy dla danego repozytorium powinna ruszyć destylacja wniosków.
  • 100gh-oss-release-prep
    Prowadzi przez listę kontrolną przygotowania repozytorium do publikacji jako open source, aby projekt spełniał standardy OSS przed upublicznieniem.
  • 100oss-code-analysis
    Przegląda drzewa źródeł repozytoriów open source na GitHubie, aby porównać implementacje funkcji między projektami albo dogłębnie zbadać jeden projekt.
  • 100oss-project-health
    Ocenia witalność utrzymania repozytorium open source: koncentrację kontrybutorów (bus factor), rytm commitów i wydań, responsywność w issues i niezałatane advisory, a zwraca werdykt healthy, at-risk lub abandoned.
  • 100gh-create-issue
    Tworzy issue na GitHubie dla bieżącego repozytorium zgodnie z formatem szablonu zgłoszeń repozytorium.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.