Module learning 

Learning loop — detect successful multi-step tool sequences from completed sessions and synthesize reusable skill documents.

Architecture

When a session closes (TTL expiry or rotation), the governor calls learn_on_close. This function:

1. Loads all tool calls for the session via get_tool_calls_for_session()
2. Flattens them chronologically and runs detect_candidate_procedures
3. For each candidate, either reinforces an existing learned skill or synthesises a new SKILL.md and persists it

No LLM call is involved — learning is pure template synthesis from observed tool-call data, keeping the hot path fast and deterministic.

Structs

CandidateProcedure

A candidate procedure detected from a session’s tool-call history.

ProcedureStep

A single step within a candidate procedure.

Functions

detect_candidate_procedures

Flatten tool calls from all turns into chronological order, then detect consecutive sequences of ≥ min_length tools where the success ratio meets the threshold.

learn_on_close

Main entry point: called by the governor when a session closes.

synthesize_skill_md

Generate a SKILL.md document from a candidate procedure.

write_learned_skill

Write a learned skill to disk under {skills_dir}/learned/.