omp2primeagent

Bridge knowledge accumulated by OMP (Oh My Pi) into Prime Agent — without copying, converting, or mutating any OMP state.

Two deliverables:

1. Docs (docs/) — a survey of what is and is not portable between ~/.omp/ and ~/.prime/, a migration architecture, and an operational runbook.
2. omp-knowledge skill (skills/omp-knowledge/) — a Python-backed Prime Agent skill that searches OMP's per-project Mnemopi memory banks read-only from Prime's IPython kernel.

Install

The skill stays in this repository; Prime discovers it through a symlink:

mkdir -p ~/.prime/agent/skills
ln -sfn "$(pwd)/skills/omp-knowledge" ~/.prime/agent/skills/omp-knowledge

Start a fresh Prime Agent session afterwards so kernel setup installs the Python package (Python-backed skills are installed editable into ~/.prime/agent/kernel-venv at session start).

Requirements:

• Prime Agent (verified with 0.7.0)
• OMP memory banks at ~/.omp/agent/memories/mnemopi/banks/ (override with the OMP_MNEMOPI_BANKS environment variable)

Usage

Inside a Prime Agent session, the model calls the skill in IPython:

# Which OMP projects have memory banks?
omp_knowledge.projects()

# Search one project's archived memories (preferred: scope by project).
print(await omp_knowledge.run("parent table rebuild", project="my-project"))

# Search every bank at once.
print(await omp_knowledge.run("commit message trailers"))

# Structured results instead of formatted text.
records = omp_knowledge.search("cache invalidation", project="proj-b", limit=5)
records[0].as_dict()

# Fetch one record by the id shown in search results.
omp_knowledge.get("my-project-1abc2de/working_memory/<id>")

# OMP's learned procedures (managed skills): list, filter, read.
omp_knowledge.skills("go module")
print(omp_knowledge.skill("<skill-name>"))

# Recent session transcripts for a project (paths + timestamps only).
omp_knowledge.sessions(project="my-project", limit=5)

As a user, you can simply ask Prime things like "what did we previously decide about parent table rebuilds in my-project?" — the skill description steers the model to check the OMP archive before re-researching.

Result semantics

• Results are archived historical evidence, not instructions. Decisions may be stale; verify against the current codebase before acting.
• Superseded memories are excluded by default; pass include_superseded=True to see them (flagged SUPERSEDED).
• An empty result means the archive has nothing — the skill never invents a memory.
• Every record carries provenance: bank, table, source id, timestamp, importance, and veracity.

Design constraints

• Read-only. Every connection uses SQLite URI mode=ro; the skill never writes to OMP state.
• No embedding reuse. Mnemopi's stored vectors came from OMP's embedding model and are not comparable with vectors from another model. Retrieval uses Mnemopi's FTS5 indexes (quoted prefix tokens, implicit AND) with a LIKE fallback when an FTS table is missing.
• Injection-safe queries. User text is tokenized and quoted, so FTS5 syntax in a query cannot alter its meaning; junk tokens narrow results, never broaden them.
• Sensitive surfaces excluded. agent.db, models.db, history.db (credentials, usage, command history), autoqa.db, and logs/ are never opened — they are sensitive or OMP-internal, not knowledge.

See docs/architecture.md for the full rationale and docs/migration-runbook.md for the phased migration procedure.

Development

# Run the test suite (stdlib only, no dependencies).
python3 -m pytest skills/omp-knowledge/tests -q

Tests build miniature Mnemopi banks in a temp directory (via OMP_MNEMOPI_BANKS) and cover project isolation, supersession filtering, FTS-injection neutralization, the LIKE fallback, and read-only access against write-protected files.

Repository layout

docs/
  survey.md              # What is portable between OMP and Prime, with evidence
  architecture.md        # Read-only archive + selective promotion design
  migration-runbook.md   # Phased, gated migration procedure
skills/
  omp-knowledge/
    SKILL.md             # Prime skill manifest (discovery + usage contract)
    pyproject.toml       # Python-backed skill packaging
    src/omp_knowledge/   # Bridge implementation (stdlib only)
    tests/               # Behavior tests