Agentic Architecture Audit

CSPilot now has an iterative agent path in run and graph-run. The model chooses one structured action, observes actual tool output and verification feedback, and then chooses its next action or stops. This follows the dynamic tool-use pattern in the LangGraph workflows and agents guide.

The scientific tools, verifier, and reporter remain deterministic. Fixed workflows are reusable tools; the runtime sequence between tool calls is model-selected. The profile router is still keyword-based, and legacy --agent-mode multi means specialist routing. The new --multiagent flag instead runs a model-backed decomposer, isolated child agent loops, structured summaries, deterministic aggregation, verification and reporting. Optional bounded concurrency is available.

Implemented: domain Markdown guidance, one-action decisions, observation history, failure feedback, bounded execution, saved artifacts, and full terminal reports. The reporter loads field selections from skills and reports only returned values.

Limits: no checkpoint/resume across processes, no human clarification node, no automatic proof that arbitrary scientific requests were fully satisfied, and no dependent task-DAG scheduling or inter-agent debate. Reaction aggregation has explicit physical/data compatibility checks; generic collection completeness still depends partly on model decomposition. Historical failed attempts keep final verification failed even after a successful recovery. A model finish decision is not itself proof of scientific correctness. Default tests mock the model boundary.