fix(esc): cancel streaming API call when abort signal trips

[ericleepi314]

Summary

• ESC during a multi-tool_use streaming response was still taking 20+ seconds to actually stop, despite the prior abort plumbing landing in PRs fix(esc): propagate cancel signal into tool_context for subagents #135/refactor(context): make ToolContext.abort_controller non-optional #140/feat(headless): SIGINT mid-tool cancels via AbortController #141/feat(grep,glob): abort ripgrep mid-search instead of waiting out timeout #143
• Root cause: provider.chat_stream_response had no way to observe a tripped abort controller — the only interrupt point was on_text_chunk, which never fires for tool_use-only turns
• Thread AbortSignal through chat_stream_response; the Anthropic provider registers a listener that calls stream.response.close() (same pattern as the existing idle-timeout StreamWatchdog), translating the resulting iterator raise to AbortError

The user-reported symptom

User pressed ESC during a turn where the model emitted 8 parallel Write blocks. The abort controller tripped immediately, but client.messages.stream(...) happily read to EOF (~20s of model generation). Only then did query.py's post-API abort check fire and yield "Interrupted by user" for all 8 tool_use blocks. With this fix, the listener forcibly closes the streaming HTTP response within ~50ms of the abort firing.

Changes

• src/providers/base.py — new abort_signal: AbortSignal | None = None parameter on chat_stream_response; docstring explains why text-chunk-based interruption isn't sufficient (tool_use-only turns produce no text chunks to observe).
• src/providers/anthropic_provider.py — three abort defenses:
  1. Pre-call fast-path: pre-tripped signal raises AbortError before any network work (skips the round-trip on turn-boundary aborts).
  2. Stream-close listener: registered via abort_signal.add_listener(_close_stream_on_abort, once=True) immediately after entering the stream context. Listener calls stream.response.close(). Register-then-recheck ordering closes the sub-microsecond race where _fire could snapshot the listener list and drop a newly-appended listener.
  3. Post-with-block recheck: catches a signal that fires between __exit__ and return.
  4. Exception translation: except Exception block consults abort_signal.aborted (not exception class — SDK versions vary) and re-raises AbortError when true.
  5. Cleanup: finally block detaches the listener so long-lived controllers (the REPL engine's, reused across many turns) don't accumulate dead listeners.
• src/providers/minimax_provider.py + src/providers/openai_compatible.py — pre-call fast-path only. Mid-stream listener for these providers is a future PR (Minimax shares the anthropic SDK, so it should be a small lift).
• src/query/query.py:
  • _call_model_sync grows abort_signal keyword and forwards to the provider.
  • New except AbortError: raise in _call_model_sync prevents the cancel from routing through the substring-classification error-message gauntlet (also stops the misleading [DIAG] EXCEPTION log).
  • New except AbortError: pass in the query loop falls through to the existing if params.abort_controller.signal.aborted: block — cancellation processing happens in one place.
• src/tool_system/agent_loop.py — _call_provider_for_turn grows cancel_signal keyword and forwards as abort_signal=; run_agent_loop passes through. Same fix benefits the TUI path (which uses run_agent_loop via AgentBridge).
• tests/test_provider_abort_signal.py (new) — 5 regression tests:
  • test_pre_aborted_signal_short_circuits_before_request — _ensure_client is never called for a pre-tripped signal
  • test_mid_stream_abort_closes_stream_and_raises_abort_error — synthetic 500ms stream + mid-stream abort returns in <1s (vs the 20s symptom this fixes); stream.response.close() is called
  • test_uncancelled_stream_returns_normally — no regression for never-tripped signals
  • test_no_abort_signal_param_preserves_legacy_callers — abort_signal=None default keeps SDK consumers working
  • test_listener_detached_after_normal_completion — pins that the listener is removed after normal exit

Test plan

• 5 new regression tests pass
• 85 directly related tests pass (abort, query, streaming executor, subagent, esc-cancel propagation, tool context default)
• 4409 broader related tests pass
• Critic subagent review: APPROVE (after applying race-window fix, pre-call fast-path on other providers, explicit AbortError handler in _call_model_sync)
• Manual: kick off a prompt that triggers a multi-Write parallel response, press ESC during the model's generation — should unwind in ~50ms, not 20+s

Out of scope (deferred to follow-ups)

• Mid-stream cancellation for Minimax / GLM / OpenAI providers — pre-call fast-path is in place; the response-close listener pattern needs to be ported (Minimax uses the same anthropic SDK so it's a small lift; OpenAI/GLM need their own listener around the SDK's iterator).
• MCP tools — out-of-process JSON-RPC; transport-level cancellation is a cross-cutting change across the MCP client.

🤖 Generated with Claude Code