fix(agent): close tool-call sequence on all interrupt aborts (#48879 follow-up)

[OutThisLife]

Summary

Bug report: "in profiles, if I send a message, it starts working, I hit stop, and send a new one — it seems to lose context frequently." The attached screenshot shows the agent replying "i dont have context on what 'they' refers 2 or what timing 2 increase" and falling back to a Session Search to recover what it was just doing — classic lost-context-after-stop.

Root cause

This is the #48879 failure mode on sibling code paths. #48879 ("close tool-call sequence on interrupt") added a synthetic assistant turn in finalize_turn so a /stop after a tool no longer persists a raw tool tail. Without it, the next user message lands as … tool → user, a role-alternation violation that strict providers (Gemini, Claude) react to by hallucinating a continuation of the user's message and ignoring prior context.

But finalize_turn is not the only interrupt exit. The retry-wait, error-handling, and post-error retry-wait interrupt aborts in agent/conversation_loop.py return early and never reach finalize_turn, so they still persisted and returned an unclosed tool tail. Interrupting during provider backoff / rate-limiting (common under heavy work like the CI-timing session in the report) hits exactly these paths — hence "frequently".

Fix

• Extract the close into a single shared helper close_interrupted_tool_sequence(messages, final_response) in agent/message_sanitization.py.
• Apply it at every interrupt abort: finalize_turn (now delegates to the helper) and the three early returns in conversation_loop.

This fixes the whole bug class rather than the one site, per the contribution rubric ("fix the whole bug class — sibling call paths included").

Test plan

• New tests/agent/test_close_interrupted_tool_sequence.py — pins the helper contract (tool tail closed with placeholder / preserved interrupt text / blank falls back) and the alternation-safety guarantee (appending a next user turn never yields tool → user); no-ops on assistant/user/empty tails.
• Existing tests/agent/test_turn_finalizer_interrupt_alternation.py still green (finalize_turn now routes through the helper).
• Broader interrupt suites green: test_stream_interrupt_retry, test_interrupt_propagation, test_turn_finalizer_cleanup_guard, test_message_sequence_repair, test_empty_response_recovery_persistence.

[github-actions]

🔎 Lint report: bb/close-interrupt-tool-seq-sibling-paths vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11268 on HEAD, 11270 on base (✅ -2)

🆕 New issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

✅ Fixed issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:2989: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

Unchanged: 5941 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.