fix(web): make mid-turn delta offsets step-relative

[wbxl2000]

Related Issue

No linked issue — focused bug fix for mid-turn resync on the v2 (kap-server) path. Same class of bug as the earlier v1 mid-turn resync blob fix.

Problem

During a multi-step turn, in_flight_turn text and the wire delta offset were turn-scoped. After a mid-turn snapshot resync (reconnect / refresh / resync_required), the client seeded all prior steps' text again on top of the structured transcript, producing a duplicated wall of Markdown. Live deltas from step 2+ were also aligned against the wrong local length, so they were skipped or treated as gaps.

What changed

Make in-flight accumulation and client delta alignment step-relative:

• kap-server InFlightTurnTracker: clear assistantText / thinkingText on turn.step.started (keep running_tools); stamped offset restarts at 0 each step.
• kimi-web agentEventProjector: reset turnTextLen / turnThinkLen on step boundaries so live deltas align against the current step.
• protocol: document that in_flight_turn text fields are current-step only.
• Tests cover step-boundary reset, step-relative offsets, and seed-then-continue alignment.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue, or explained the problem above.
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update. (N/A — bug fix; no user-facing docs/config change.)

Test plan

• pnpm --filter @moonshot-ai/kimi-web test -- test/agent-event-projector.test.ts
• pnpm --filter @moonshot-ai/kap-server exec vitest run test/inFlightTurnTracker.test.ts
• Manual: start a multi-step turn on v2 (pnpm dev:v2 + web), refresh or reconnect mid-turn, confirm prior steps stay structured and only the current step is seeded/streamed

[changeset-bot]

🦋 Changeset detected

Latest commit: 86be5e3

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

pnpm dlx https://pkg.pr.new/@moonshot-ai/kimi-code@86be5e3

npx https://pkg.pr.new/@moonshot-ai/kimi-code@86be5e3

commit: 86be5e3

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 86be5e384e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve v1 turn-scoped offsets

When the web UI is connected to the v1 backend, these resets make every post-step delta look like a gap: packages/server/src/services/gateway/inFlightTurnTracker.ts still computes offset from the full turn text and never handles turn.step.started, so after step 1 the next v1 assistant.delta arrives with offset > 0 while turnTextLen has just been reset to 0 here. That routes normal streaming through historyCompacted/resync instead of appending, so multi-step turns on v1 can repeatedly resync or stop streaming. Either keep v1 counters turn-relative or update the v1 tracker in the same change.

Useful? React with 👍 / 👎.

[wbxl2000]

Closing as superseded by #1609. #1609 contains this PR's exact commit 86be5e3 and adds the normalized resync-seed dedupe fix/tests in 51db588; it has merged. The legacy v1 compatibility concern also no longer applies after #1617 removed packages/server and made kap-server the only /api/v1 server path.