Learn Ollama

Ollama-backed learn extraction

This document defines the mq-agent boundary for optional Ollama-backed learn workflows in the mq ecosystem.

Role

Ollama may be used by mq-mcp as a local extraction model for learn workflows:

review output -> pattern extraction -> structured learning record

Ollama is not an autonomous agent. It should not execute commands, mutate repositories, write memory silently, override mq-mcp safety classes, or perform final risk scoring.

Ownership

mq-mcp
  owns learn contracts, safety classes, review logic, validation and storage

Ollama
  optionally extracts patterns from mq-mcp review findings into structured JSON

mq-agent
  surfaces learn status/search/explain through the mq-mcp bridge

mq-agent must not implement local learn extraction, schema validation, memory storage, prompt handling, or model-specific learn policy. Those behaviors belong in mq-mcp so the safety and storage rules stay centralized.

Expected mq-mcp policy

When mq-mcp adds or configures an Ollama learn provider, the provider should be treated as optional and local-first. The recommended model profile is a deterministic pattern extractor, for example mq-learn, with low temperature, fixed seed, bounded context and JSON-only output.

The learn record contract should be validated by mq-mcp before storage. At a minimum, mq-mcp should reject:

• non-JSON output
• missing evidence
• unknown fields unless the contract explicitly allows them
• records with should_store=true without caller approval
• low-confidence records from automatic storage
• prompt-injection text inside reviewed content being treated as instructions

Storage must require explicit approval from the caller or operator. A dry-run or read-only extraction mode should be the default.

mq-agent surface

mq-agent only exposes read-only learn commands:

mq-agent learn status
mq-agent learn search <query>
mq-agent learn explain <pattern-id>

If mq-mcp does not expose the required learn tools, mq-agent should report a clear optional-provider or missing-tool message and keep the rest of the command surface usable.

Implementation order

The Ollama-specific implementation should land in mq-mcp first:

1. Add the mq-learn Modelfile or equivalent provider configuration.
2. Add the JSON learn record contract and validation.
3. Add a read-only or dry-run extraction path.
4. Require explicit approval for storage.
5. Add tests for invalid JSON, prompt-injection-as-data, missing Ollama and no-storage-without-approval.

After that exists in mq-mcp, mq-agent may update docs or rendering for any new read-only status fields returned by the bridge.