Add configurable, always-on execution-trace log sink (local|hf|s3)

perf(mcp): HTTP transport core — add_mcp_http on agent/tool_manager + adata cache + instrumentation

Surface per-dataset preprocessing + re-pin biodata-registry 0.1.5

feat: hide microarray from advertised inventory + plain-language Approach in solutions

Lint cleanup: remove unused imports, empty f-strings, split import lines

Stage: pinned deps, API-key fallback, apt probe, security scan

Add dataset selection heuristics to system prompt

Re-inject tool functions before each step to heal namespace shadowing

Enable Anthropic prompt caching on the system prompt

Fix history truncation dropping the agent's last response

Cap conversation history and code-output size to bound credit cost

Add dataset_filter_to_curated_samples tool and loading plan integration

Surface reporting_rules to agent system prompt; add DESeq2 explanation rule

Fix model prefill error on continue: convert trailing AIMessage to HumanMessage before model.invoke()

Inject full dataset manifests into system prompt to eliminate discovery calls

Fix message format for newer Claude models + upgrade to claude-sonnet-4-6

Ground agent in registered datasets; prevent hallucination

Fix __main__ CLI entry point in agent.py

Remove StateManager class and redundant PlanManager instantiations

Unify CLI and Gradio executor setup via _prepare_executor()

Clean up dead code, fix two bugs, remove unused abstractions

Align project layout with HF Spaces and FastMCP conventions