Fix agent model cache identity

[dubh3124]

Summary

Fixes a runtime cache bug where agent-level model overrides could reuse a stale compiled Deep Agents graph.

Root Cause

_model_cache_key() collapsed plain string model IDs to str, so different string IDs such as us.amazon.nova-lite-v1:0 and global.anthropic.claude-sonnet-5 could share the same graph cache entry. The cache also lacked a resolved provider/config identity, leaving same-model/different-provider or same-model/different-base-url cases vulnerable to stale graph reuse.

Changes

• Include string model IDs in _model_cache_key().
• Add a secret-free resolved model cache identity including provider, model ID, base URL, Bedrock region/role, generation settings, retries, and timeout.
• Pass that identity from model resolution through DeepAgentStreamingService into CognitionAgentParams / RuntimeContext.
• Preserve legacy resolve_model_for_session() tuple behavior while adding richer resolve_runtime_model_for_session() metadata.
• Add regression tests for string model IDs, same-model/provider-config changes, resolver identity, and service pass-through.
• Record both bug-fix entries in ROADMAP.md.

Validation

• PYTHONPATH=. uv run pytest tests/unit/test_pluggability.py tests/unit/test_provider_resolution.py tests/unit/test_agent_def_field_wiring.py -q passed, 45 tests
• PYTHONPATH=. uv run pytest tests/unit -q passed, 797 passed, 4 skipped
• uv run ruff check . passed
• PYTHONPATH=. uv run mypy server/app/agent/cognition_agent.py server/app/agent/resolver.py server/app/llm/deep_agent_service.py --ignore-missing-imports passed

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
cognition	Ready	Preview, Comment	Jul 9, 2026 1:35am