Reliable Long-Session Recovery and Turn Cost Tracking / 可靠的长会话恢复与本轮费用统计

[SivanCola]

Summary

• Use AppendOnlyLog.totalLength when marking the start of a turn so discardCurrentTurn does not truncate long sessions after the log window is applied.
• Add a regression test that aborts a turn after the in-memory window is exceeded and verifies prior history is preserved.
• Split status bar cost display into current turn cost and cumulative session cost for desktop/dashboard.

Root-cause notes for #2287

The concrete correctness bug found in this area is the abort/discard index: after the windowed AppendOnlyLog change, this.log.length is only the retained window size, not the full persisted history length. If an explicit abort with discardCurrentTurn happens in a long session, the rewrite can slice at the wrong index and drop persisted history outside the current window. That changes future prompt prefixes and can damage cache behavior.

I did not find the proposed stripDroppableReasoningContent explanation to be the primary cause of every tool-call iteration becoming a full cache miss: the strip path skips assistant messages with tool_calls, and ordinary assistant messages stripped before their first inclusion in a later request do not invalidate a previously sent prefix that already contained them. Missing required reasoning_content for historical thinking-mode assistant messages would also be more likely to surface as a request-shape error than as a silent full-cache-miss loop.

Verification

• npm run test -- tests/loop-user-persist.test.ts desktop/src/App.test.ts desktop/src/ui/context-panel.test.tsx tests/desktop-btw-status.test.ts tests/desktop-user-message.test.ts
• npm run typecheck
• npm --prefix desktop run build
• npm --prefix dashboard run build
• npm run lint (passes with existing PlanPanel.tsx import-type warning)
• pre-push npm run verify: build, lint, typecheck, and full test suite passed (314 test files, 4042 tests passed, 9 skipped); the first push attempt failed only at the final GitHub TLS send-pack step and was retried with --no-verify after verification had completed.

Related to #2287.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d7a591aba9

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".