[ci, scripts] feat: add agent skills and workflow

[Zachary-wW]

Summary

Set up two-pillar agent automation for LoongForge contributors:

1. Reusable agent skills under skills/ — code-review (framework-agnostic PR review prompt) and submit-pr (PR submission workflow), discoverable by Claude Code via project-level config.
2. Two @//-comment-triggered GitHub Actions workflows — claude-review.yml for AI PR review, internal-ci.yml for bridging public PR review with the internal build pipeline.

End-to-end verified on a fork PR: the Claude reviewer reads modified files, follows skills/code-review/SKILL.md, and posts severity-tagged inline comments on file:line plus a top-level verdict / summary / checklist.

Changes

Agent skills and Claude project config (commit 1)

• skills/code-review/SKILL.md — framework-agnostic review prompt: 10 review areas (correctness, security, API design, performance, reliability/obs, testing, readability, docs, conventions, repo hygiene), 4-tier severity (🔴 Critical / 🟠 Major / 🟡 Minor / 🟢 Nit), structured verdict + checklist output.
• skills/submit-pr/SKILL.md — PR submission workflow: validates branch state, derives a CI-conformant title from the diff, generates the description, gets user approval, then pushes and opens the PR via gh.
• .claude/settings.json — allow-list a small set of read-only git and pre-commit commands so they run without per-call prompts.
• .claude/skills symlink → ../skills makes the skills discoverable by Claude Code's project-level skill loader.
• .gitignore — keep .claude/projects/ and .claude/credentials/ out of git while allowing settings.json and the skills symlink to be committed.
• CLAUDE.md — new "Code Review" section directs the reviewer agent to follow skills/code-review/SKILL.md and to post each finding as an inline comment via the create_inline_comment MCP tool, reserving the top-level summary for verdict + checklist only.

Claude PR review workflow (commit 2)

• .github/workflows/claude-review.yml — @claude comment-triggered review using the upstream anthropics/claude-code-action@v1.
• Trigger: @claude mention on issue_comment, pull_request_review_comment, pull_request_review, or issues (default phrase per upstream action).
• Authorization gate: only OWNER / MEMBER / COLLABORATOR commenters can fire the action, with a self-author bypass for the original PR/issue author. Without this gate, drive-by commenters on a public repo could trigger paid agent runs and execute tools on the self-hosted runner.
• Permissions: contents: read (agent reads files only), pull-requests: write + issues: write (agent posts comments), id-token: write (the action's OIDC token request), actions: read.
• API routing: env.ANTHROPIC_BASE_URL routes Claude API traffic through the internal gateway; verified via the Claude Code CLI that the gateway supports tool use + streaming.
• Model: Claude Sonnet 4.6 — the gateway exposes Claude models with display-name IDs (verified via GET /v1/models); the canonical hyphenated form is rejected.
• Tools: --allowedTools enables the inline review-comment MCP tool plus gh / Read / Grep / Glob.
• Behavior: prompt: is intentionally empty so the action runs in default agent mode — PR context and GitHub MCP tools are auto-injected. Project-wide review guidance lives in CLAUDE.md and skills/code-review/SKILL.md.
• Progress: track_progress: true shows live checkbox progress in the bot's PR comment.

Internal CI workflow (commit 3)

• .github/workflows/internal-ci.yml — /loongforge-ci PR-comment-triggered (or manual workflow_dispatch) workflow that packages PR code together with the Loong-Megatron submodule, uploads to BOS using runner-local config, writes a CI trigger signal file, and notifies the configured Ruliu group.
• All internal config (BOS bucket, Ruliu webhook, build trigger path) is read from the self-hosted runner's local filesystem rather than from repo secrets.

Test Plan

Skills

• code-review and submit-pr SKILL.md files are loaded by Claude Code via the .claude/skills -> ../skills symlink and discoverable as project skills.

Claude review workflow

End-to-end verified on fork PR (Zachary-wW/LoongForge#8):

• Triggered with @claude review this pr.
• 14 turns, ~3 min runtime, ~$3 in API cost.
• Agent posted 6 inline review comments on specific file:line (severity-tagged: 🔴 / 🟠 / 🟡) plus 1 top-level verdict + summary + checklist comment, exactly matching the skills/code-review/SKILL.md output contract.
• Two earlier dry-runs uncovered (and the fixes are included in this PR):
  • prompt: must be left empty for the action to run in agent mode and inject GitHub tools.
  • id-token: write is required for the action's OIDC token request.
  • Internal gateway uses display-name model IDs, not the hyphenated canonical form.
  • mcp__github_inline_comment__create_inline_comment must be explicitly allow-listed via --allowedTools.
• The agent's own self-review caught two security regressions (missing author_association gate, over-broad contents: write); both are fixed in commit 2 before merging.

Required Setup (admin) Before Running on This Repo

1. Add repo secrets (Settings → Secrets and variables → Actions):
   • ANTHROPIC_API_KEY — internal oneapi-comate gateway key.
   • ANTHROPIC_BASE_URL — internal gateway base URL.
2. Self-hosted runner with labels [self-hosted, macOS, ARM64] registered to this repo (or shared from the org). The internal API gateway is on the corp network and is not reachable from GitHub-hosted runners.
3. No GitHub App / OAuth setup is required — the action authenticates with anthropic_api_key + the workflow's built-in GITHUB_TOKEN and OIDC.

Notes

• Per-review cost is roughly $1–$3 (multi-turn agent vs. one-shot). Acceptable for the depth gain (full-file reading, cross-reference analysis, true inline comments) but worth flagging.
• Workflow uses the self-hosted macOS ARM64 runner already standard for this org; internal CI similarly relies on the runner's local config.

[Zachary-wW]

/loongforge-ci