agent-learning-consumer

name	agent-learning-consumer
description	Set up and operate voku/agent-learning in a consuming repository, including learning-root config.json, findings, proposals, consolidation prompts, validation, and lifecycle commands. Use for end-user learning-loop workflows, not package source development.

Agent Learning Consumer

Use this skill when a project wants to run voku/agent-learning locally. The goal is reviewable evidence, not automatic long-term guidance edits.

Fast Path

1. Locate or create the learning root, commonly infra/doc/agent-learning, .agent-learning, docs/agent-learning, or agent-learning.
2. Add config.json when the project layout is not obvious or when generated constraints need project-relative files:

{
  "schema_version": "1.0",
  "project_root": "../../..",
  "constraint_generation_dir": "constraint-generation",
  "active_constraints_dir": "constraints/active"
}

3. Capture raw experience as findings first. Do not skip straight to a durable rule unless evidence is already validated.
4. Classify reusable learning explicitly:
   • ADD_LEARNING_NOTE is the default for useful raw observations.
   • UPDATE_SKILL is preferred when an existing skill owns the behavior.
   • CREATE_SKILL is rare and requires existing-skill overlap evidence.
   • IGNORE is valid for praise, vague reflection, and one-off noise.
5. Every non-IGNORE classification needs pattern_key and a concrete validation_case with given, when, and then.
6. Prefer NO_DURABLE_LEARNING when the evidence does not support a bounded reusable change.
7. Use proposals for reviewable mutations to skills, docs, memory, constraints, or other durable guidance.
8. Keep approval human-gated: candidate proposals can be generated by agents; durable activation should be explicit.
9. Use guidance-evaluate only with immutable recall-selection and per-guidance outcome events from voku/agent-recall-compiler; it may write candidate proposals, never approved/applied guidance.
10. Run validation before handing off the output.

Commands

vendor/bin/agent-learning validate --root infra/doc/agent-learning
vendor/bin/agent-learning prepare --root infra/doc/agent-learning --task PROJECT-123
vendor/bin/agent-learning proposal-validate --root infra/doc/agent-learning --proposal proposal.2026-06-15.001.json
vendor/bin/agent-learning guidance-evaluate --root infra/doc/agent-learning --selection-history history/recall-selections.jsonl --outcome-history history/outcomes.jsonl

Use --task-id-pattern when the consuming project does not use the default ticket format.

Output Expectations

• Findings stay under findings/<status>/.
• Proposals stay under proposals/<status>/.
• Decision history stays in history/*.jsonl.
• Recall selection and guidance outcome histories stay append-only; do not replace them with mutable counters.
• Consolidation input should be deterministic and based on validated findings plus rejected proposal context.
• Generated long-term guidance should name its boundary and validation command.
• Do not store praise, success summaries, self-justification, or generic advice as learning notes.