commit-push-pr

Files (1)

• SKILL.md

SKILL.md (first lines)

# Commit, Push, and Create PR

Follow these steps precisely. This skill chains to `/commit-push` (which itself chains to `/commit`).
Read `_shared/repo-config.md` for repo table, remote URLs, PR draft location/naming, and safety rules.

## PIPELINE_MODE (from `/kernel-fusion-pipeline`)

When invoked with a CONFIG object containing `"pipeline_mode": true`, run **fully autonomously** —
NEVER use `AskUserQuestion`, never wait for human review. Skip all interactive steps (2, 4-review,
6, 8, 11). Use CONFIG fields instead of asking. This is the SAME commit/PR contract as the
interactive flow; only the "ask the user" gates are replaced by CONFIG values.

**CONFIG fields:**
- `worktree` — run all git/gh commands from here (`cd "$worktree"` or `git -C "$worktree"`).
- `repo` — base repo for the PR (e.g. `sgl-project/sglang`).
- `base` — PR base branch (e.g. `main`).
- `commit_subject` — exact commit subject (already `[<Tag>] [AMD] <area>: …` formatted).
- `pr_title` — exact PR title.
- `pr_body_file` — path to a file with the full PR body (already composed; use VERBATIM).
- `update_pr` — optional PR number; if set, UPDATE that PR's body instead of creating a new one.
- `gh_env` — optional env prefix for gh, e.g. `GH_CONFIG_DIR=/home/yichiche/.gh GH_TOKEN=""`.
- `sglang_root` — repo root to return to for worktree cleanup.

**Flow:**

1. `cd "$worktree"`.

2. **Fix lint BEFORE committing** (CI enforces pre-commit; a dirty lint state fails CI):
   - `pip3 install -q -U pre-commit` (fine if already present)
   - `git add -A`
   - `pre-commit run --all-files` — let hooks auto-fix (isort, ruff F401/F821, black, codespell,
     clang-format, nbstripout).
   - If files changed: `git add -A && pre-commit run --all-files`; repeat up to 3×.
   - If a hook still fails (e.g. ruff F821 it cannot auto-fix), read the error, fix the code in the
     worktree, then retry from `git add -A`.
   - If `--all-files` reformats files UNRELATED to this change, reset them
     (`git checkout -- <file>`) so the commit stays focused on this change only.
   - Do not proceed until pre-commit exits clean.

3. Commit (only if there are staged changes): `git commit -m "<commit_subject>"`.
   **Never add `Co-Authored-By` trailers** (in subject or body).

3b. **SGLang triage gate, before push.** Run this only when `<repo>` is
    `sgl-project/sglang`. For every other repo, skip it. Do not read
    `/pr-merge-triage`, and do not run `path_cover.py` or `test_seam.py`.
    When the repo is SGLang, follow Step 3b of `/commit-push` on this worktree.
    Score `git diff <base>...HEAD`. Fix failing Affected Scope, AMD Guard,
    Unit Test Quality, and Flags rows, then commit each fix and score again,
    at most 3 times. `BLOCKED — SPLIT FIRST` is not auto-split. If a row still
    fails, do not push. Return `{ status: fail, error: <the failing row> }`.

4. Push the branch: `git push` (use `git push -u origin HEAD` if no upstream is set).
   Only after step 3b passes for SGLang.

5. PR — pick ONE path:
   - **Create** (no `update_pr`):
     `<gh_env> gh pr create --repo <repo> --base <base> --title "<pr_title>" --body-file <pr_body_file>`
     Use `--body-file` (never `--body`) so markdown tables survive shell quoting.
   - **Update existing** (`update_pr` set): `gh pr edit` currently fails on a projects-classic
     GraphQL deprecation, so update the body via the REST API:
     `<gh_env> gh api -X PATCH repos/<repo>/pulls/<update_pr> -f body="$(cat <pr_body_file>)"`
   - If `gh pr create` fails on auth/permission, fall through Step 9's Attempts 2-4
     (`--web`, `curl` REST POST, prefill URL). In pipeline_mode do NOT block on the user —
     record the failure and return `status: fail` with the reason.

5b. **Tag `run-ci` label** (when `<repo>` is `sgl-project/sglang` and the PR was created
    programmatically): follow Step 9b below. Non-fatal on failure — record
    `run_ci_tagged: true/false` in the return JSON.

6. Cleanup: `cd "$sglang_root"` then `git worktree remove <worktree>` (keep the local branch).

7. Return JSON: `{ slug, commit_hash, pr_url, run_ci_tagged, status }` (`status` = `pass`/`fail`,
   plus `error` if failed).

---

## Repo-specific fast paths

Detect the repo in Step 1. If it is **InferenceX** — also called **InferenceMax** —
(`SemiAnalysisAI/InferenceX`), read the InferenceX section of `_shared/repo-config.md` **before**
Step 2 and apply these deltas:

- **Step 2 (push target)**: push goes to `origin` = `SemiAnalysisAI/InferenceX` directly — there is
  no fork. Base `main`. The working branch must be named **`amd/<branch_name>`**; if the current
  branch is not, create one (`git checkout -b amd/<name>`) before pushing. Titles and commit
  subjects are **`[AMD][<model_name>] <title>`** (e.g. `[AMD][Qwen3.5] …`).
- **Step 3 (commit/push)**: no pre-commit config exists — skip the lint loop. Plain `git push`
  fails on a missing credential helper; use the `GIT_CONFIG_COUNT=1 … '!gh auth git-credential'`
  prefix from repo-config. A human `Co-Authored-By` trailer is allowed here when the user names one;
  a Claude co-author trailer never is.
- **Step 6 (data collection)**: for recipe/config/image-bump PRs there is no local benchmark to
  collect — the GPU sweep runs in CI. Skip 6a-6c and instead record the evidence that the change is
  valid (e.g. the Docker Hub tag check returning 200), then go straight to Step 7.
- **Step 7 (draft)**: use `_shared/pr-template-simple.md`. Write **English only — no Chinese
  characters anywhere** (this overrides the bilingual rule in the repo's `AGENTS.md`); the same
  applies to commit messages and PR comments. Every recipe or perf-affecting change
  also needs an appended `perf-changelog.yaml` entry **in the same commit** — see repo-config for the
  append-only / trailing-newline rules and the read-then-write footgun.
- **Step 9 (submit)**: `gh pr create --label full-sweep-fail-fast` works. Any later title/body edit
  must go through `gh api -X PATCH repos/.../pulls/<num>` — `gh pr edit` fails on this repo with a
  projects-classic GraphQL deprecation error.
- **Step 9b**: not applicable (that is sglang's `run-ci` label). The InferenceX equivalent is the
  `full-sweep-fail-fast` label, already applied at create time. Verify with
  `gh pr checks <num> --repo SemiAnalysisAI/InferenceX`.
- **After creating the PR**: patch the `pr-link:` placeholder in the `perf-changelog.yaml` entry with
  the real PR URL, amend, and force-push. Re-verify the trailing newline and `git diff origin/main
  --stat` before pushing — an amend is the step most likely to silently corrupt that file.
- **Bulk image bumps** across many recipes are a different job — use the repo's `/nuke` command
  (`.claude/commands/nuke.md`), which owns the one-PR-per-model+precision+SKU grouping rules.

## Step 0: Ensure `gh` CLI is available and authenticated

Check that the GitHub CLI is the real one (v2.x+, not the pip `gh` v0.0.4):
```bash
gh --version 2>&1
```
If missing or wrong version, follow the install steps in `_shared/repo-config.md`.

After `gh` is installed, check authentication:

From yichiche/agent-box, MIT.

More from yichiche/agent-box