[currentbits]

Part of #1475 (renderer finding 1 in docs/perf/2026-10-05-renderer.md). provenanceById was recomputed over the whole transcript on every streaming push: about 30% of renderer JS while streaming into a big open thread, and ~48 ms per big-thread switch.

Change

• New provenanceByMessageId(messages) in src/provenance.ts, used by useTranscriptAnnotations:
  • One pass carries each turn's tool refs forward, instead of rescanning the turn for every assistant message.
  • Each tool call is classified, and its input JSON-parsed, once per object.
  • An assistant message whose object and accumulated refs are unchanged returns its cached result.
• The renderer keeps unchanged message objects identical between pushes (the patch replaces only the changed tail), so the cache hits during streaming.

Checked independently on real data

• A read-only benchmark over the three largest real transcripts (1,019–1,062 messages, mostly tool rows) ran 200 simulated streaming pushes each, appending to a new last assistant message immutably, as the store does.
• Equivalence: at every push, the new map deep-equals the old per-message computation, with 0 mismatches over 600 pushes.

Transcript	Old per push (median)	New per push (median)	Speedup
1,062 messages	1.25 ms	0.03 ms	36.6×
1,019 messages	1.38 ms	0.04 ms	34.6×
1,039 messages	0.11 ms	0.03 ms	3.7×

These are plain-Node timings, without React rendering or the profiler; PM2's in-app figure was larger. The ratio is what carries over.

Tests: test/provenance.test.ts covers streaming, rewind and edit-and-resubmit equivalence. It fails on main, where the function doesn't exist. Renderer suite 2705/2705, typecheck clean.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.