F020: Structured Reasoning Traces

Context

Current decision logging captures the decision (output) and context (input summary), plus high-level reasons (categorical). It completely misses the process—the step-by-step reasoning chain that led to the decision.

Recent research (GRPO, DAPO) shows that optimizing the reasoning process (step-level value) is critical for high-stakes agentic tasks. To support this, CSTP must capture the reasoning trace.

Requirement

Update the Decision model and recordDecision endpoint to accept a structured reasoning trace.

Specification

1. Schema Update

Add trace field to Decision model:

class ReasoningStep(BaseModel):
    step: int
    thought: str  # The reasoning content
    output: str | None = None  # Optional intermediate result/action
    confidence: float | None = None  # Step-level confidence (if available)
    tags: list[str] = []  # e.g. ["planning", "critique", "selection"]

class Decision(BaseModel):
    # ... existing fields ...
    trace: list[ReasoningStep] | None = None

2. API Update

Method: cstp.recordDecision

Params:

• trace (optional, list): List of reasoning steps.

Example Payload:

{
  "decision": "Use Postgres instead of SQLite",
  "trace": [
    {
      "step": 1,
      "thought": "Analyzing concurrency requirements...",
      "output": "High concurrency needed",
      "tags": ["analysis"]
    },
    {
      "step": 2,
      "thought": "Comparing SQLite vs Postgres limits...",
      "output": "SQLite locks on write",
      "tags": ["comparison"]
    }
  ]
}

3. Downstream Usage

• Offline Analysis: DAPO-style attribution of failure to specific reasoning steps.
• Pattern Detection: “Steps tagged ‘analysis’ often missing in failed decisions.”
• Debugging: Replay the agent’s thought process.

Migration

• Backward compatible (optional field).
• Existing decisions have trace=None.