fix: guard agents.models import in warm_up_llm against cross-thread partial-import race (#1248)

[not-knope]

Summary

• Fresh (non-resume) scans crash at LLM warm-up with KeyError: 'agents.models' because warm_up_llm imported agents.models.interface on the main thread while the perf: take heavy imports off the startup path and pre-warm them in the background #1141 daemon was still populating that package (#1248).
• Add an Event (_agents_models_ready) released after strix.core.runner finishes (it loads agents.models) or fails. wait_for_agents_models() blocks on that Event only — it does not join() the warm-up thread — then the original import runs. LiteLLM / Caido / Docker keep overlapping model preflight.
• The Event is set in a finally so a failed warm-up (missing optional deps, etc.) cannot deadlock waiters. _purge_orphaned_modules is unchanged. --resume shares this warm_up_llm path, so it is covered rather than regressed.

Test plan

• uv run pytest tests/test_import_warmup.py (8 passed), including: wait is a no-op without warm-up; wait does not join remaining modules; wait returns when the graph import fails; from agents.models.interface import ModelTracing after the wait succeeds.
• ruff check / ruff format --check / mypy on the touched files.
• Drive main() with Docker stubbed using --target http://example.com --instruction-file … --scan-mode quick --max-budget 2 --non-interactive: import survives; process reaches real model preflight (401 on a dummy key), not KeyError.
• wait_for_agents_models() returns with LLM_API_KEY unset (warm-up is import-only).
• Real CLI with Docker running and the sandbox image already cached (the [BUG] Fresh (non-resume) scans on v1.6.0 crash at warm-up: KeyError: 'agents.models' #1248 race window: pull_docker_image() is a local inspect, a few hundred milliseconds). 3/3 runs, no KeyError: 'agents.models'. The process gets past import warm-up and reaches live model preflight.

Real CLI logs (macOS arm64, source install, image cached)

uv run strix --target http://example.com \
  --instruction-file /tmp/strix-1248-instructions.txt \
  --scan-mode quick --max-budget 2 --non-interactive

STRIX_LLM=anthropic/claude-sonnet-5, dummy LLM_API_KEY (enough to prove we get past the import; the original crash happened before any provider call).

Run 1/2/3 — all exit 1 at preflight, none at agents.models:

╭─ STRIX ──────────────────────────────────────────────────────────────────────╮
│                                                                              │
│  LLM CONNECTION FAILED                                                       │
│                                                                              │
│  Could not establish connection to the language model.                       │
│  Please check your configuration and try again.                              │
│                                                                              │
│  Error: litellm.AuthenticationError: AnthropicException -                    │
│  {"type":"error","error":{"type":"authentication_error","message":"API key   │
│  is invalid."},"request_id":null}                                            │
│                                                                              │
╰──────────────────────────────────────────────────────────────────────────────╯

Exit 1 is expected (invalid key). Before this fix the same command died at warm_up_llm with KeyError: 'agents.models' and never sent a request.

Closes #1248.
Related: #1230 (same crash site on --resume; this guards that import without a full-thread join).

[greptile-apps]

Greptile Summary

This PR serializes the first cross-thread agents.models import against the background import warm-up without waiting for unrelated heavyweight modules.

• Adds a readiness Event around the strix.core.runner portion of import warm-up.
• Waits for that boundary in both headless and interactive model-preflight paths.
• Adds subprocess regression coverage for no-warm-up, failure, partial-wait, and successful-import cases.

Confidence Score: 5/5

The PR appears safe to merge, with no concrete blocking or independently actionable non-blocking issue identified.

The new wait covers both model-preflight entry paths, releases safely on failed warm-up attempts, and does not block completion of unrelated background imports.

Important Files Changed

Filename	Overview
strix/llm/warmup.py	Introduces the readiness Event and isolates individual warm-up imports while preserving background overlap for later modules.
strix/interface/main.py	Waits for the Agents model graph before the headless warm-up path imports ModelTracing.
strix/interface/scan_setup.py	Applies the same readiness guard to the interactive model-preflight path.
tests/test_import_warmup.py	Adds subprocess tests covering readiness signaling, failed graph imports, no-op behavior, and non-joining waits.

_{Reviews (1): Last reviewed commit: "fix: guard agents.models import in warm_..." | Re-trigger Greptile}