LeapFlow 0.2.0 — Teacher/Student World-Model Self-Evolution

[wangxingjun778]

A two-agent architecture: the world model (cold path) reviews each session and distils what it learned into the student's next-turn context. It never acts — it changes what the agent knows. Self-evolution (writing new capabilities) is a separate, off-by-default switch.

Feature

• World-model teacher grades every session and returns four-value adaptation verdicts — absorb / rebind / acquire / escalate — each carrying knowledge for the acting agent.
• C1 — distilled knowledge: the teacher writes environment facts the student reads as context, enabling zero-code adaptation; entries retire by supersession, TTL, and explicit retract.
• C2 — rebind preference: the teacher's recommended provider becomes a resolver selection preference, never a gate (an inadmissible candidate still loses to declarations).
• C3 — gated acquisition: an acquire verdict becomes a queued plugin proposal, behind the new switch; queued still passes generation, validation, approval, sandbox, and trust.
• evolution.enabled switch (default off) in its own Self-Evolution config category, announced at startup in both states.
• Degradation feedback loop: plugin health becomes capability-scoped evidence; the teacher sees prior verdicts and available alternatives; recovery retires stale knowledge.

Fix

• Lifecycle governor, durable trust ledger, and proposal sink were never wired in production (read via getattr against unassigned attributes) — Progressive Trust and acquisition were inert; now wired, with a startup audit to prevent regressions.
• Single risk-clamp point for model-authored proposals, preserving the original requested_max_risk_level for audit.
• rebind guard no longer rejects a provider the teacher was itself shown; retry-owned failure classes are filtered from the teacher; failure_class persisted.

Refactor

• Removed the unused Thompson/UCB1/Bucketed selection policies and their state store (measured: consulted zero times, one candidate per capability) — kept the SelectionPolicy seam and GreedyPolicy.
• intents are derived from acquire verdicts (single source of truth); one shared capability-name rule; one switch admits world-model evidence (was two knobs).

Enhance

• Environment fingerprint on degradation facts, so one upstream change reads as a single root cause rather than N.
• Inert-wiring audit tool (tools/audit_inert_wiring.py) added to CI.

Documentation

• Third-party plugin development guide updated for the evolution contracts; repository-wide copyright headers.