Skip to main content

Reflection Engine

WASP has two reflection layers:

  1. Goal Meta-Reflection — post-mortem after every goal completion or failure.
  2. Consolidation reflection — narrative reflection during the background consolidation cycle.

Goal Meta-Reflection​

scheduler/goal_meta_reflection.py (or goal_orchestrator/reflection_job.py). Feature-flagged via goal_meta_reflection_enabled (default true).

When a goal completes or fails, the reflection job runs once and produces structured insights:

ExecutionReflection(
    timestamp,
    chat_id,
    intent,                 -- inferred from goal objective
    skills_used,
    duration_ms,
    success,
    efficiency_score,       -- 0..1
    issues,                 -- JSON array
    insight,                -- short LLM-generated note
    suggestion,             -- improvement proposal
    pattern_key,            -- normalized for SkillPattern detection
    recurring_pattern       -- True if this matches a known SkillPattern
)

Insights surfaced​

Stored in Redis with TTL 7 days; injected into context for similar future goals. Examples:

  • "Took 3 retries to find the right CSS selector. Consider using browser_smart_navigate instead of browser for sites with unpredictable DOM."
  • "Repeated the same web_search 4 times. Cache the result on first hit."
  • "Email send blocked by intent gate. The user did not explicitly request it; the agent was over-eager."

Maximum 3 reflections per goal.

Consolidation reflection​

scheduler/dream.py (gated). When the background consolidation cycle activates, it runs an LLM reflection over the day's activity and writes:

DreamLog(
    started_at, duration_seconds,
    memories_consolidated, kg_nodes_added,
    reflection,                 -- short LLM narrative
    improvements_proposed,
    improvements_json,          -- proposed self_improve diffs
    prefetch_done               -- crypto prices, news headlines
)

The reflection is informal: a paragraph of "what went well today, what could be better." It is NOT a behavioral rule and does NOT modify code on its own — improvements are proposed for operator review at /self-improve.

Dashboard​

/cognitive shows recent reflections (Goal Meta and Consolidation).

CPI gating​

Both reflection layers skip when agent:cpi_high is set.

See also​