Contentrain/ai

Minor Changes

• f1a2dce: Three additive GitHub-provider optimizations that cut API calls per write (no breaking changes)

  • applyPlan blob inlining — applyPlanToGitHub now inlines file content directly in the Git tree entries instead of POSTing a blob per file. A write drops from N+3 GitHub mutations to a fixed 3 (tree + commit + ref) regardless of file count, easing the mutation-rate secondary limit. This brings the GitHub provider to parity with the GitLab provider, which already inlines content in its commit actions. Deletions are unchanged (null-sha entries); content is passed through byte-for-byte.
  • GitHubReader opt-in read memoization — new GitHubReader(client, repo, { memoize: true }) deduplicates repeated (path, ref) reads within one operation across readFile / listDirectory / fileExists. Opt-in and off by default; failed reads are evicted so transient errors are retried, not cached. The provider's own long-lived reader does not enable it (a reader that outlives a write must not serve stale results).
  • buildContextChange stats injection — buildContextChange(reader, operation, source, { stats }) lets a caller that already knows the model/entry counts skip the O(models·locales) reader walk; only readConfig remains. The emitted context.json is byte-identical to the scanned variant. Parameterless callers are unaffected.

Patch Changes

• cfface5: Speed up local writes by batching git subprocess spawns in the worktree transaction

  Every local write (content_save, model_save, bulk, normalize apply, …) runs through createTransaction, which spawned ~40 git processes for a single content save. Two behavior-preserving batches cut that:

  • Commit identity via environment — the worktree git instance now carries GIT_AUTHOR_* / GIT_COMMITTER_* in its env instead of running two git config user.* spawns per transaction (and per contentrain_merge). Git honors these for both the sync merges and the feature-branch commit.
  • Batched selective sync — selectiveSync replaced its per-file git cat-file -e existence loop with a single git ls-tree over the changed paths, and its per-file git checkout HEAD -- <file> loop with a single batched checkout (per-file fallback on error preserves precise skip accounting). Working-tree deletions are parallel fs removals.

  Measured: a 2-locale content_save drops from ~40 git spawns to ~30 (config 2→0, cat-file 5→0, checkout 7→3), roughly halving the operation's wall-clock. The selective-sync win scales with file count, so bulk writes benefit the most. No behavior change — the full worktree-transaction test suite passes unchanged.