[albin-zhu]

Summary

Repeated Stop hooks currently load and parse the entire main rollout, including already acknowledged turns. This change checkpoints acknowledged byte boundaries and parser context so ordinary main-thread hooks read only the newly appended suffix. It adds no daemon, SQLite dependency, model routing change, or exporter replacement.

• Store a private, atomic <rollout>.langfuse.cursor checkpoint with parser context, bounded conversation suffix, original turn count and source guards.
• Preserve generation input, developer/system instructions, discovered tool definitions, token usage, UTF-8 boundaries and seeded turn numbering across hook processes.
• Defer incomplete trailing lines; retain pre-completion Stop-turn bytes until completion or a following turn.
• Advance only after the export callback succeeds and the existing turn delivery receipt is present. Removed receipts, corrupt checkpoints, changed history limit, replaced/truncated sources rebuild.
• Preserve the full-read ownership path for subagent/ancestor traversal and max_history_chars: 0; directory discovery remains unchanged. This is explicitly a main-rollout optimization, not a claim that all plugin I/O is incremental.

Why

#119 / #122 address generation-history allocation/flush pressure. Reading and parsing the acknowledged rollout prefix on each subsequent hook remains a separate cost. This PR reduces that repeated logical input; it does not assert physical disk bytes, an across-the-board machine-lag root cause, or lossless provider-wire history.

Verification

• pnpm run lint: formatting and TypeScript passed.
• pnpm test: 11 files / 136 tests passed.
• New behavioral regressions cover warm/cold generation equality, omission counts, historical tool definitions, late subagent ownership, UTF-8 chunks/partial tails, final Stop before completion, failed acknowledgement, corrupted/replaced/truncated sources, removed receipts and seeded numbering.
• Ran the actual bundled hook in separate Node processes against synthetic loopback OTLP data with fake credentials. No real model turns, production uploads or user configuration changes.

Synthetic acknowledged prefix	Cold hook	Warm hook	Warm rollout suffix read	Upload payload	Cursor file
985,889 bytes	73 ms	61 ms	846 bytes	3,843 bytes	690 bytes
33,552,754 bytes	135 ms	58 ms	846 bytes	3,846 bytes	693 bytes

Both warm runs emitted one new root and preserved the previous turn's generation context. Reads also include fixed source guards, metadata discovery and checkpoint/receipt files; the measured suffix is not total physical disk I/O. Published 0.4.3 warm runs on equivalent fixtures took 83 ms and 92 ms; these single-run timings are descriptive, not a performance guarantee.

Review points

The checkpoint contains transcript content and is documented as private (POSIX 0600; Windows inherited directory ACLs). First use/cold rebuild still reads the history. Concurrent writers can roll the optimization back to an older acknowledged position (extra reads) rather than risking skipped delivery. Subagents deliberately remain on the existing fidelity path.

Also included: opt-in LANGFUSE_CODEX_DIAGNOSTICS hook progress records (no transcript content, credentials, raw errors or full paths), and bounded subagent discovery — the day pruning never matched on Windows because path.relative returns backslashes, so every hook walked the entire session history.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.