Reflection

WASP runs reflection at two levels: per-goal post-mortems and the dream-cycle narrative. Both produce structured artifacts the operator can review.

Per-goal post-mortem

scheduler/goal_meta_reflection.py. Runs once when a goal completes or fails.

Field	Meaning
`intent`	Inferred from the goal objective
`skills_used`	Ordered list of skills the executor ran
`duration_ms`	Total wall-clock time
`success`	True/False
`efficiency_score`	0–1, weighted by skills used vs minimum needed, retries, replans
`issues`	JSON array of issues encountered (timeouts, replans, gate hits)
`insight`	Short LLM-generated note
`suggestion`	Improvement proposal
`pattern_key`	Normalized for SkillPattern detection
`recurring_pattern`	True if this matches a known SkillPattern

Operator reads it at /goals (per-goal detail page) and /cognitive (recent reflections).
Context Builder injects recent reflection insights into similar future goals.
Skill Evolution uses pattern_key and recurring_pattern to detect candidates for composite-skill synthesis.

Maximum 3 reflection rows per goal. Older reflections are pruned by execution_reflection_pruner (daily).

scheduler/dream.py. Runs during the dream cycle (idle-time, gated by inactivity + CPI).

Field	Meaning
`reflection`	Short LLM narrative on the day's activity
`improvements_proposed`	Count of proposed self-improvements
`improvements_json`	The actual proposals (each is a diff)
`memories_consolidated`	Episodic entries promoted to semantic
`kg_nodes_added`	New entities discovered
`prefetch_done`	Whether crypto prices were prefetched

Operator reads the narrative at /cognitive.
Self-Improve dashboard shows proposals from improvements_json for review at /self-improve.
Failure-pattern analysis updates self_model["known_failures"].

The dream cycle does NOT modify code on its own — every improvement is operator-gated.

Where	What
`/goals` (per-goal page)	Goal Meta-Reflection for that goal
`/cognitive` (Reflections tab)	Recent goal reflections + dream narratives
`/self-improve`	Pending dream-cycle improvements
`dream_log` table	Raw history of dream cycles
`execution_reflection` table	Raw history of goal post-mortems

Trigger	What runs
Goal completes	Goal Meta-Reflection (one-shot)
Goal fails	Goal Meta-Reflection (one-shot)
Dream cycle activates	Dream reflection + memory consolidation + improvement proposals

To turn off reflection entirely:

GOAL_META_REFLECTION_ENABLED=false
DREAM_ENABLED=false

This stops the LLM-cost of reflection. Operator loses post-mortem visibility; self-model accumulates failures less precisely.