fix(core): preserve session state on abort and error / session expiration (replaces #1055)

[Utkarsh-X]

Fixes #1054
Reopens #1055 (rebased onto the new main history following @victorxheng's note).

What this fixes

When a user interrupts a run with Esc or when a session expires, sending a follow-up message immediately was reading a stale/null in-memory previousRunStateRef. This caused the SDK to construct a fresh empty session state (messageHistory = []), wiping conversation context.

What's included (addressing the #1055 review feedback)

1. Regression tests (send-message.test.ts): Added 3 dedicated tests verifying:
   • User abort (Esc) preserves turn history + tool calls in createRunConfig.previousRun.
   • Session expiration / gate errors preserve history on "continue".
   • Mid-run chat switches (runChatIsCurrent === false) never leak state into other chats.
2. DRY state sync (use-send-message.ts): Consolidated the assignment into a simple 5-line syncRunState closure used uniformly across abort, completion, and error handlers.
3. Clean scope: Reverted the previous prompt text tweaks so this PR is strictly isolated to the session state bug (3 files total).
4. Re-render safety: Checked store subscriptions — zero React components subscribe to state.runState, so updating it alongside streamStatus = 'idle' causes no visual layout flash.

I intentionally kept this as minimal and safe as possible to solve the demonstrated bug without bloating the change surface or touching unrelated client/storage APIs. That said, if you'd prefer any tweaks or want me to look into anything further here, just let me know and I'm happy to research and adjust!

Verification

• Test Suites (202 / 202 passed across touched areas):
  • bun test cli/src/hooks/helpers/__tests__/send-message.test.ts (49 passed, including the 3 new regression tests)
  • bun test sdk/src/__tests__/run-cancellation.test.ts (16 passed)
  • bun test cli/src/utils/__tests__/run-state-storage.test.ts (37 passed)
  • bun test packages/agent-runtime/src/__tests__/compact-history.test.ts (31 passed)
  • bun test agents/__tests__/context-pruner.test.ts (69 passed)
• Builds:
  • bun run build:sdk (ESM, CJS, and TypeScript declarations bundled with 0 errors)
  • bun run build:freebuff (Native executable built with exit code 0)

[codebuff-team]

The core insight is reasonable: previousRunStateRef (and the store's runState) weren't being refreshed in the abort callback (registerActiveRun) or in the error/catch path, so an immediate follow-up message could pick up a stale/null snapshot and start a fresh session. Introducing a single syncRunState closure and calling it uniformly in the abort handler, the completion path, and the error path (use-send-message.ts lines ~332-353, ~700, ~752) is a sensible, minimal way to fix that, and it's a legitimate in-scope change (cli/src/hooks).

The problem is the tests in send-message.test.ts. All three new tests hand-roll a local copy of the syncRunState logic (const syncRunState = (state) => { if (currentChatDir !== '/chat-1') return; ...}) inside the test body rather than importing and exercising the actual closure from use-send-message.ts. They then assert against createRunConfig using a manually-set previousRunStateRef.current. This proves the concept works, but it does not verify that the real hook wires latestRunStateSnapshot into syncRunState correctly, or that latestRunStateSnapshot is actually fresh (not stale/null) at the abort callback call site — which is the crux of the original bug. Since use-send-message.ts isn't unit-testable in isolation here, this would need either an integration test that drives the actual hook (e.g. via renderHook) through an abort/error cycle, or at minimum a comment/justification for why the current test strategy is an acceptable proxy.

Worth also double-checking that latestRunStateSnapshot is guaranteed to be populated with the latest partial state at the moment the abort callback fires (before client.run settles) — the diff assumes this but the surrounding declaration isn't shown.

Right direction, but please add a test that actually calls into the hook/module under test rather than reimplementing its logic in the spec file.

[Utkarsh-X]

Addressed the feedback. Replaced the 3 manual tests with 3 hook-level integration tests in use-send-message.test.tsx. Each test mounts the real useSendMessage hook via createTestRenderer + createRoot, drives it through the real codepath, and asserts against the actual previousRun passed into the real createRunConfig call — no reimplemented logic in the test body.

The 3 tests cover:

1. Abort with streaming progress — calls onStateSnapshot(snapshot) then stopActiveRun('user-interrupt'), proving the real abort callback at line ~355 commits the live snapshot to previousRunStateRef before client.run() settles.

2. Error / session expiry — calls onStateSnapshot(snapshot) then rejects client.run(), proving the real catch block at line ~752 commits the snapshot.

3. Abort with no prior snapshot (directly addresses your follow-up question about latestRunStateSnapshot freshness) — aborts immediately before any onStateSnapshot fires, proving latestRunStateSnapshot initializes from previousRunStateRef.current (the prior completed run's state) and is therefore never null at the abort callsite.

All 4 tests in the file pass (3 new + 1 existing upstream test).