vm-osil-dataflow

OSIL Data Flow

How a single task flows through the system and learning propagates.

sequenceDiagram
    participant T as Task
    participant A as Agent
    participant TC as tool_calls
    participant SK as L1 SI Skill
    participant REF as L2 Reflexion
    participant DSPY as L3 DSPy/GEPA
    participant SKI as L4 Skill Induction
    participant SDB as Agent SQLite

    T->>A: incoming work
    A->>A: execute current prompt + skills
    A->>TC: write tool_call CHOKEPOINT-1
    A->>T: result

    alt Task succeeded
        SKI->>TC: nightly scan recent successes
        SKI->>SKI: detect repeated pattern
        SKI->>A: propose new skill PR with auto-test gate
    else Task failed
        SK->>SK: capture learning to learnings dir
        REF->>TC: hourly scan failures
        REF->>REF: generate reflection via Claude
        REF->>SDB: write reflection w/ vec0 embed
        Note over A,SDB: Next similar task<br/>retrieves top-3 reflections<br/>injects into prompt
    end

    Note over DSPY: Weekly batch
    DSPY->>SDB: pull eval set
    DSPY->>DSPY: GEPA optimize prompts
    DSPY->>A: deploy via CHOKEPOINT-2

Flow types

• Success path → Skill Induction Worker scans for repeated patterns, proposes new skills via PR (auto-test gate + Henry signoff).
• Failure path → Self-Improvement skill captures session learning; Reflexion runner generates reflections, stores w/ vec0 embed; next similar task retrieves and injects.
• Weekly batch → DSPy/GEPA pulls eval set from agent SQLite; optimizes prompts via reflective Pareto evolution; deploys via CHOKEPOINT-2 (writes infra_config_changes).

Cross-links

• Parent plan: openclaw-self-improvement-layer-2026-05-03
• Related maps: vm-osil-overview · vm-osil-stack · vm-osil-decision-tree · vm-osil-vendor-map
• Chokepoint policy: feedback_chokepoint_principle