[knqiufan]

Re-scoped onto current main (journal-v2 checkpoints).

Upstream now records turn changes through command frames and increments instead of scanning the whole workspace in beginTurn, so the incremental checkpoint cache this PR originally carried no longer applies. This PR now overlaps the remaining pre-request work — the file-artifact baseline — with model startup, and gates every write entry point behind that readiness.

Refs #687 (the turn-start checkpoint scan half is provided upstream).

Changes

• Start one shared preparation task after durable input acceptance: beginTurn still persists the turn's checkpoint record, and FileArtifactCollector.start prepares the artifact baseline.
• With no matching pre-request hooks, the model request proceeds immediately. Tools, custom validation, matching hooks, subagents, background tasks and plugins await readiness.
• Matching UserPromptSubmit / InstructionsLoaded / PreModelRequest / compaction hooks wait for protection; unmatched hooks never dispatch early.
• Preparation failure aborts an in-flight provider, blocks tools, records the failure and releases leases. Terminal success, artifact output and checkpoint finalization wait for preparation.
• Closing the stream or cancelling aborts preparation, closes the model iterator and joins cleanup; the artifact scan supports cooperative cancellation.
• Restore/undo formats, leases and conflict checks are untouched.

Validation

Node 22.23.2, Windows 11:

Check	Result
npm run build (tsc + plugins)	passed
node --test dist/tests/session/workspace-preparation.spec.js	29/29 passed, repeated runs
npm run test:checkpoints	remaining failures are the pre-existing Windows rename/timing flakes and reproduce on unchanged main

Not included

• The incremental checkpoint cache, its tests and its docs from the earlier revision, which targeted the removed turn-start scan.
• The old macOS benchmark numbers. They must be re-measured against the journal-v2 baseline; scripts/benchmark-turn-startup.ts is kept for that.

[xhd0728]

The re-scoped implementation looks sound and is worth landing as a startup-latency improvement. Overlapping the artifact baseline with model startup has a concrete benefit, while the shared readiness barrier keeps tools, matching hooks, subagents, background tasks, and terminal finalization behind workspace preparation.

The 29 preparation tests and 84 related tests passed locally, and Web and Desktop CI are green. In a synthetic 2,000-file workspace with a fixed 50 ms local model delay, the median first-text latency across the four warm runs fell from about 80 ms to 52 ms. This is a scoped benchmark, not a general end-to-end speedup claim.

Before merging, please update docs/parallel-workspace-preparation.md to follow CONTRIBUTING.md: maintained developer documentation must be in English, and implementation history or one-run validation narratives belong in the PR or CI artifacts. Keep the durable readiness/cancellation contract and benchmark usage in the canonical documentation.

I did not find a blocking implementation issue. Once that documentation cleanup is done, I recommend merging; its priority is lower than the correctness fixes in #702 and #703.

[knqiufan]

@xhd0728 Thanks for the review. docs/parallel-workspace-preparation.md is now the English contract: it keeps the readiness and cancellation rules and the benchmark usage, and the implementation history stays in this PR. I also linked it from docs/README.md.