[zwanski-start]

Closes #146

What changed

• CLI: new --testing-model on autonomous_agent_demo.py and parallel_orchestrator.py. Testing agents get --model <testing_model>. Coding and initializer agents keep using --model. If you don't set it, testing agents use the coding model, so nothing changes by default.
• Settings: new global testing_model setting. "" clears it, which means "same as coding". It is validated against VALID_MODELS, and legacy IDs are migrated. registry.get_testing_model_setting() resolves it.
• API: SettingsResponse/SettingsUpdate/AgentStartRequest gain testing_model. ProcessManager.start() falls back to the global setting, so both manual and scheduled starts pick it up.
• UI: a "Testing Agent Model" selector (Same as coding / Opus / Sonnet) in both SettingsModal and SettingsView.

Assumptions

• The override applies only with the Claude provider. Alternative providers pin every model tier to one api_model, so a second model would have no effect. The selector is hidden for those providers.
• The testing model is global and is not stored per schedule.

Verification

• New test_testing_model.py (10 tests): checks the spawned commands for coding and testing agents, how the setting is resolved, and schema validation.
• ruff passes. mypy shows no new errors (the existing errors are in files this PR doesn't touch). The security, client, dependency-resolver and rate-limit tests pass.
• npm run lint has no errors (6 warnings that were already there). npm run build succeeds.
• Not done: manual browser check with Playwright (ran out of session budget).

🤖 Generated with Claude Code