Fix control_response request_id check — read from nested response (closes #975)

[FidoCanCode]

Fixes #975.

Real bug

Probed the actual claude-code subprocess (claude --version → 2.1.120). Its control_response shape is:

{"type": "control_response",
 "response": {"subtype": "success", "request_id": "..."}}

request_id is nested under response. Fido's _send_control_set_model checked obj.get("request_id") at the top level. Always None. Predicate never matched. Every switch_model call hung until the subprocess was killed.

Proof

$ grep -c 'switch_model: now on model=' ~/log/fido.log
0
$ grep -c 'switch_model:.*(control_request)' ~/log/fido.log
9

Zero successful returns. Nine attempts. Every one hung. Concrete trace:

• 15:40:23 [home] switch_model: opus → haiku
• (no [home] log lines for 5 minutes)
• 15:45:23 [home] worker started (fresh thread post-fido-restart)

The session was wedged the entire window. Same shape across every switch_model log line in the file.

Why fido seemed to work anyway

Many switch_model calls are no-ops (target == current model) and return at line 1102 before sending the control_request — so sessions that didn't need an actual model change kept working. The 15:40:30 webhook test that succeeded landed on a freshly-booted confusio session that was already on opus from spawn — no switch needed. Bug only fires on real model changes.

Fix

claude.py:1023-1028:

if obj.get("type") == "control_response":
    response = obj.get("response") or {}
    if response.get("request_id") == request_id:
        return

Test

Existing test fixtures had the wrong shape too — they put request_id at the top level, matching fido's broken expectation. Both helpers (TestClaudeSessionSwitchModel._make_response_line and TestClaudeSessionSendControlSetModel._make_response_line) updated to emit the real nested shape.

New regression test test_ignores_top_level_request_id_in_control_response explicitly emits a malformed top-level placement first, then the correct nested one — asserts the malformed one is skipped and the wait completes only on the nested response. This catches any future regression that re-introduces top-level lookup.

All 264 claude tests pass.

Related

• Bug: switch_model hangs when called with _in_turn=True after preempt-cancelled drain #975 — this issue
• Use control_request set_model for mid-session model switch instead of session restart (closes #852) #971 — landed control_request set_model with the broken predicate
• Bug: webhook handler's first prompt aborted by leftover cancel flag from worker preempt #973 / Bug: webhook handler's first prompt aborted by leftover cancel flag from worker preempt (closes #973) #974 — cancel-leak fix; uncovered the hang underneath
• Lint: ban silent .get(literal) and 'k in t and t[k]' patterns at non-dynamic call sites #976 — silent .get linter (would have caught this at write time)