PR #68176: CLI: bake git commit at build time so --version reports built commit not live HEAD [AI-assisted]

• Repository: openclaw/openclaw
• Author: sparkeros
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/68176

Description (problem / solution / changelog)

Fixes #68148.

Summary

Before: node dist/index.js --version reports the current value of .git/HEAD, which can drift past the commit the binary was actually built from.

After: the short SHA is baked into dist/ at build time as a tsdown compile-time define, so --version reports what's executing — even if HEAD moves after the build.

Priority order

New: GIT_COMMIT/GIT_SHA env → injected define → live git → build-info.json → package.json.gitHead.

Env override still wins (CI use case). Live git becomes the fallback for dev runs from source (no define injected), which is where the existing "prefers live git metadata over stale build info in a real checkout" test case (unchanged, still passing) captures correct behavior.

Changes

• tsdown.config.ts: resolve short SHA at bundle time (env GIT_COMMIT/GIT_SHA → git rev-parse HEAD → undefined), pass as define.__OPENCLAW_GIT_COMMIT__ to every nodeBuildConfig output (covers both the unified core build and every buildBundledPluginConfigs entry).
• src/infra/git-commit.ts: add declare const __OPENCLAW_GIT_COMMIT__, a readCommitFromInjectedDefine helper, and a readInjectedCommit slot on CommitMetadataReaders. resolveCommitHash consults it after the env override and cache check, before live git.
• src/infra/git-commit.test.ts: 4 new cases covering injected-beats-live-git, env-still-overrides-injected, null-injected-falls-through, and cache-stability.

No change to scripts/write-build-info.ts — dist/build-info.json remains for external tooling.

Repro (on upstream/main before the fix)

git worktree add -b repro /tmp/repro upstream/main
cd /tmp/repro
pnpm install --frozen-lockfile && pnpm build
git log -1 --format=%h                    # be7a415
node dist/index.js --version              # OpenClaw 2026.4.16 (be7a415)
git commit --allow-empty -m "shift HEAD"
node dist/index.js --version              # OpenClaw 2026.4.16 (d2d0ede)  ← lie

Same sequence after the fix: second --version still reports be7a415.

Testing

• pnpm test src/infra/git-commit (via vitest.infra.config.ts) — 18/18 pass (14 existing + 4 new)
• pnpm build — green
• pnpm check — green
• pnpm test — one flaky 120s-timeout in extensions/memory-core/src/memory/index.test.ts > "indexes multimodal image and audio files from extra paths with Gemini structured inputs"; zero import-path coupling to this PR's changes; passed cleanly on isolated re-run. All other 60 shards green.
• End-to-end: built binary + empty commit + HEAD shift → --version stays stable at the built commit; env GIT_COMMIT=deadbee ... --version correctly reports deadbee (override preserved).

AI-assisted disclosure

• Degree of testing: fully tested — focused unit lane + full pnpm build && pnpm check && pnpm test + end-to-end HEAD-shift repro both before and after the fix.
• Tool: Claude Code (Opus 4.7). Sparkeros (PR author) reviewed and confirmed understanding of every change.
• Prompt / session: the session explored the issue collaboratively — identified root cause in resolveCommitHash priority order, confirmed no existing upstream PR addresses it (PR #56372 is cache-only), chose the bundler-define approach over alternatives (file-based marker, live-git-with-disclaimer), deliberately kept scope narrow by not also wiring the vestigial __OPENCLAW_VERSION__ define.
• I understand what this code does and why each line is there.

Changed files

• src/infra/git-commit.test.ts (modified, +66/-0)
• src/infra/git-commit.ts (modified, +14/-0)
• tsdown.config.ts (modified, +23/-0)