Observe

Observe

Watch how work happens. Capture events. Do not interpret.

Read schemas in ../soulcatcher/reference.md.

Worked example traces: learned-skills/examples/sample-traces/

Rules

1. Read-only — do not change tone, recommendations, or workflow while observing
2. Meaningful events only — skip chit-chat and one-off noise
3. Literal descriptions — record what happened, not why the user did it
4. No inference — leave confidence null; interpretation belongs in later steps

What to Capture

What to Ignore

• Greetings and small talk
• One-off typos or formatting fixes
• Model defaults unrelated to user habit
• Single tool calls with no procedural significance

Sources

• Current conversation and tool history
• User-pasted session logs
• Agent transcript files (JSONL) if user provides path
• Git diffs attached to the session

Output Format

Produce one BehavioralTrace per session:

# BehavioralTrace: {session_id}

```yaml
session_id: decision-analysis-03
problem_class: decision-analysis
context: Choose between two API pagination strategies
events:
  - order: 1
    type: action
    description: Stated the decision to be made
  - order: 2
    type: action
    description: Listed two candidate approaches
  - order: 3
    type: decision
    description: Wrote pros and cons for each option
  - order: 4
    type: tool_call
    description: Searched for failure cases of offset pagination at scale
    detail: web search
  - order: 5
    type: decision
    description: Identified strongest downside of preferred option
  - order: 6
    type: tool_call
    description: Searched for evidence contradicting the leading choice
  - order: 7
    type: decision
    description: Committed to cursor-based pagination after risks were explicit
confidence: null

## Quality Check

Before handing off to detect-patterns:

- [ ] Each event is literal and ordered
- [ ] No interpretive language ("user prefers", "user always")
- [ ] Problem class is assigned
- [ ] 5–8 meaningful events per session (fewer ok for short sessions)