feat(agent): apply per-reasoning-model stale-timeout floor in stream + non-stream detectors (#52217)

[teknium1]

Summary

Known reasoning models now tolerate their multi-minute thinking phase instead of having the upstream cloud gateway idle-kill the socket (BrokenPipeError / RemoteProtocolError) before the first token arrives.

The reasoning-model allowlist (agent/reasoning_timeouts.py) already landed on main via #52795. This PR wires its floor into the two stale detectors that previously used the chat-model defaults (180s stream / 90s non-stream), which are too short for reasoning models.

Salvage of #52238 (fixes #52217), cherry-picked onto current main with @DavidMetcalfe's authorship preserved.

Changes

• agent/chat_completion_helpers.py — stream stale detector consults get_reasoning_stale_timeout_floor() and applies it as max(stream_timeout, floor).
• run_agent.py — non-stream detector (_resolved_api_call_stale_timeout_base) returns the reasoning floor with uses_implicit_default=False, so the local-endpoint short-circuit doesn't disable stale detection for reasoning models on a local NIM endpoint.
• tests/agent/test_reasoning_stale_timeout_floor.py — 51 tests covering the floor table, anchoring, and wiring.

Dropped from the original PR branch (now redundant / stale): the duplicate agent/reasoning_timeouts.py (already on main), and the stale-base reverts of the MoA client path / _extract_landed_file_mutation_paths / moa_config that the original branch carried because it predated those landing on main.

Contract (verified)

• Never overrides explicit user config — providers.<id>.models.<model>.stale_timeout_seconds, HERMES_API_CALL_STALE_TIMEOUT, etc. resolve first.
• Never lowers an existing threshold — applied as max(default, floor).
• Zero effect on chat models — they return None from the allowlist.

Validation

Live E2E through the real AIAgent._resolved_api_call_stale_timeout_base() resolution chain (real imports, temp HERMES_HOME):

Model	Resolved stale timeout
nvidia/nemotron-3-ultra-550b-a55b	600s (floor), not implicit default
openai/o3-mini	300s (floor)
openai/gpt-4o	90s implicit default (unchanged)
nemotron + HERMES_API_CALL_STALE_TIMEOUT=45	45s (user config wins over floor)

Stream-detector path verified: max(180 base, 600 floor) == 600; chat model floor None → base unchanged; floor never lowers (max(300, 300) == 300).

Tests: tests/agent/test_reasoning_stale_timeout_floor.py — 51/51 passing.

Test plan

scripts/run_tests.sh tests/agent/test_reasoning_stale_timeout_floor.py

Infographic

[github-actions]

🔎 Lint report: salvage/52238-reasoning-floor-wiring 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: 11455 on HEAD, 11399 on base (🆕 +56)

🆕 New issues (12):

Rule	Count
invalid-argument-type	11
unresolved-import	1

First entries

tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `bool`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `list[str] | None`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `str`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `list[str]`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `int | float`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `list[dict[str, Any]]`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `dict[str, Any]`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `((str, dict[Unknown, Unknown], /) -> None) | None`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:38: [unresolved-import] unresolved-import: Cannot resolve imported module `pytest`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `int | float | None`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `int`, found `str | bool`
tests/agent/test_reasoning_stale_timeout_floor.py:154: [invalid-argument-type] invalid-argument-type: Argument to `AIAgent.__init__` is incorrect: Expected `IterationBudget`, found `str | bool`

✅ Fixed issues: none

Unchanged: 6003 pre-existing issues carried over.

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