agent-oss-contribute

Original：🇨🇳 Chinese

Translated

2 scripts

End-to-end open-source contribution for Agent/LLM projects: discover GitHub repos (stars, downloads/usage, recent commits, issue close ratio, PR merge speed, release cadence) and rewrite each project description in the user's language (what it does, problem, scenarios), analyze an Agent/LLM repo, write a scoped PR plan, implement it, open a GitHub PR, then write a PR summary from saved artifacts. Use when the user wants to search for contributable projects, find open source projects, Agent open source, MCP/LangChain repositories, star 1000, GitHub repo search; pastes a GitHub URL for open source contribution analysis, PR direction, good first issue, Agent architecture flaws; asks for technical solution, implement the plan, submit PR with feat/fix/refactor/test/docs; or produce PR summary, summary.md, review submitted PR, pastes a pull request URL / previously mentioned project. Trigger even without the word skill. Do not use for private product feature work or non-Agent general OSS drive-bys.

14installs

Sourcemengma-dev/skills

Added on2026-09-25

NPX Install

npx skill4agent add mengma-dev/skills agent-oss-contribute

SKILL.md Content (Chinese)

View Translation Comparison →

Agent Open Source Contribution Full Workflow

A six-phase pipeline for Agent / LLM / MCP open-source repositories. When the user doesn't specify a phase, infer it from the current conversation and run only that phase; don't search a project and auto-edit code, don't open a PR without a plan, and don't auto-write a summary after a PR.

Response language follows the user; default to Chinese if not specified. Keep technical terms in English. Output only the structure required for the current phase.

Phase Routing

What the user is doing	Enter	Read first
Searching for contributable projects / looking for open-source repos / no `REPO_URL` but wants to contribute	Phase 0 Discovery	references/discover.md
Provided `REPO_URL` / looking for PR direction / issue triage / architecture flaws	Phase 1 Analysis	references/analysis.md and references/architecture-dimensions.md
Picked one from Top 3 / wants technical plan / problem scope / implementation steps	Phase 2 Design	references/design.md
Pasted the selected plan and wants implementation / technical solution implementation / define interfaces first	Phase 3 Implementation	references/implementation.md
Submitting PR / open PR / sending PR with feat/fix	Phase 4 Pull Request	references/pull-request.md
Producing PR summary / reviewing submitted PR / given PR URL or previously mentioned project	Phase 5 Summary	references/summary.md

Run only one phase at a time, then stop and wait for the user. When the user explicitly says "do it all in one go": without

REPO_URL

, run Phase 0 first and stop for the user to pick a repo; with

REPO_URL

, proceed from Phase 1 onward, and the Phase 2 plan must still be reviewed by the user first. Phase 5 is not auto-run with "do it all in one go" unless the user also asks for a summary.

Ask for missing key inputs; don't force-run with fabricated repos or plans.

Shared Inputs

Field	Description
`REPO_URL`	GitHub repo URL; not needed for Phase 0
`PR_URL`	GitHub PR URL ( `.../pull/N` ); Phase 5 uses it to locate the project first
`LOCAL_PATH`	Local repo root; default is the current task working directory (in this repo's evals, clone to repo root `test/` when unspecified)
`ARTIFACT_ROOT`	Artifact root directory; default is `<current task working directory>/artifacts` (in this repo's evals, `<repo root>/test/artifacts` )
Tech stack	If not given, use: Python, TypeScript, React, MCP integration, LangChain.js
Target cycle	If not given, use: 2 weeks

Current task working directory: the process working directory at the start of this phase (cwd / workspace root). Not tied to any specific editor or Agent product; if the user explicitly provides

ARTIFACT_ROOT

, use theirs.

When running skill tests /

evals/

in this repo: the working directory, fixtures, and clone target when

LOCAL_PATH

is not given are all under

test/

in the skill collection repo root (gitignored, don't commit its contents). Daily use follows user's cwd /

LOCAL_PATH

/

ARTIFACT_ROOT

.

Phases 1–4 must run in the target repo working directory. Phases 3 and 4 must not modify other repos. Phase 5 doesn't require switching to the target repo and must not modify any repo code.

Artifact Persistence

Each phase, while replying to the user, must write the full output of this phase to

ARTIFACT_ROOT

(default

<current task working directory>/artifacts

; in this repo's evals

<repo root>/test/artifacts

). Chat output and saved content must match; write the file before ending the phase. All paths are written and saved relative to

ARTIFACT_ROOT

; you may include absolute paths in replies for convenience.

Phase	Path relative to `ARTIFACT_ROOT`	Rule
0 Discovery	`search/<YYYYMMDD-HHMMSS>.md`	Create a new file per search; filename uses local time at execution
1 Analysis	`project/<owner>-<repo>/analysis.md`	Create project directory if missing; overwrite if exists
2 Design	`project/<owner>-<repo>/design.md`	Same as above
3 Implementation	`project/<owner>-<repo>/implementation.md`	Same (change summary + commit plan + verification results)
4 Pull Request	`project/<owner>-<repo>/pull-request.md`	Same (title, draft body, final PR URL)
5 Summary	`summary/<owner>-<repo>/<pr-name>.md`	Create project directory if missing; overwrite if same PR rerun

<owner>-<repo>

is derived from

REPO_URL

(e.g.

https://github.com/langchain-ai/langchainjs

→

langchain-ai-langchainjs

).

```
<pr-name>
```
rules see references/summary.md (prefer
```
{number}-{slug}
```
).
Artifacts for Phases 1–4 are written only in the corresponding
```
project/
```
directory, not in other project folders.
Phase 0 writes only to
```
search/
```
, don't create directories under
```
project/
```
.
Phase 5 writes only to
```
summary/
```
, don't overwrite intermediate artifacts in
```
project/
```
.
Don't hardcode machine-specific absolute paths (e.g.
```
/Users/...
```
); don't hardcode product-specific directories.
If persistence fails (permission/path missing): create missing directories first then write; if it still fails, tell the user the error, don't pretend it was saved.

Phase Gates

0 → 1: The user selected a repo from discovery results (or provided their own
```
REPO_URL
```
).
1 → 2: The user picked one from the analysis Top 3, or gave an equivalent direction.
2 → 3: The selected plan, problem scope, success criteria, file list, and implementation steps are complete, and the user confirmed implementation.
3 → 4: Code has been implemented and verified according to the plan; the user explicitly says commit / open PR. Don't
```
git commit
```
/
```
push
```
/
```
gh pr create
```
unless asked.
4 → 5: The user explicitly asks for a summary, or provides the submitted PR project /
```
PR_URL
```
. Don't auto-write a summary after Phase 4. Phase 5 can be entered directly from a new session; it doesn't require Phases 0–4 in the same session.

Hard Constraints

Phase 0 must run
```
scripts/discover_repos.py
```
, and is forbidden from fabricating repo rankings from memory; must pass
```
--artifact-root
```
and
```
--user-language
```
; user verbal filters must be translated into script parameters. Project descriptions must be in the user's language, covering what it does / problems solved / applicable scenarios; downloads, usage scores, and rankings come from the script, and rankings or usage must not be fabricated.
Each phase must persist artifacts (see "Artifact Persistence"); no replies without writing files.
Read the repo's real files and Issues first, then use training memory. Use repo-relative paths.
Contribution scope must be completable in about 2 weeks; directions like "rewrite the orchestration layer" are forbidden.
Phases 3, 4: minimal intrusion, backward compatible, consistent with repo style; new third-party dependencies must be asked first. Phase 5 doesn't change code.
Phase 4 commit types only:
```
feat
```
/
```
fix
```
/
```
refactor
```
/
```
test
```
/
```
docs
```
.
Phase 4 PR body: forbidden to fabricate checklist /
```
Test plan
```
, forbidden to append any AI/editor attribution footer.
Phase 5: star counts, PR body, and artifacts must come from
```
gh
```
or files already stored in
```
ARTIFACT_ROOT
```
; don't fabricate analysis sources or repo data.