Feature request: support the Oh My Pi (omp) agent harness (transcript adapter or documented custom watch schema)

[FernandeZ-hjm]

Feature request: support the Oh My Pi (omp) agent harness

TL;DR claude-mem's read path (MCP search tools) already works inside omp sessions, but the capture path doesn't: the worker never sees omp sessions (activeSessions: 0), because omp doesn't emit Claude Code hook events and has no built-in transcript watcher schema. Request: an omp transcript adapter, or a documented custom-schema format so the community can write one.

What is omp

Oh My Pi is a coding-agent harness (terminal UI, subagents, MCP xd:// tool devices) used with models such as kimi-code. Like Claude Code/Codex/Cursor, it persists full session transcripts as JSONL:

~/.omp/agent/sessions/<project-slug>/<timestamp>_<session-id>.jsonl
~/.omp/agent/sessions/<project-slug>/<timestamp>_<session-id>/<subagent>.jsonl

Example event types observed in a real transcript:

title, session, model_change, thinking_level_change,
custom_message, message, text, custom, thinking, task, path, title_change

Example lines:

{"type":"session","version":3,"id":"019f8762-...","timestamp":"2026-07-22T01:13:02.183Z","cwd":"/Users/x/AI 工作站","title":"Check claude-mem injection status"}
{"type":"message", ...}

Current behavior in omp

• The claude-mem MCP tools (mcp-search: search / timeline / get_observations / corpora) load fine in omp and can read existing memory.
• observation_add rejects with requires CLAUDE_MEM_RUNTIME=server (omp runs worker mode).
• Worker /api/health reports activeSessions: 0 while an omp session is actively running — nothing is captured, nothing is injected.

Why this should be feasible

claude-mem already generalizes capture beyond Claude Code:

• built-in watchers for codex / cursor / gemini transcripts (sessions_codex_count, sessions_gemini_count in the DB)
• user-defined watches + schemas via ~/.claude-mem/transcript-watch.json (resolveSchema reads config.schemas[name])

So the missing piece is either:

1. A built-in omp schema in the transcript watcher (preferred), or
2. Documentation of the custom schema format expected by transcript-watch.json — the watcher code is bundled/minified, so writing a correct third-party schema today requires reverse engineering. With a documented schema contract (event mapping: user prompt / assistant message / tool call / tool result / session start/end), the community could contribute adapters for omp and other harnesses.

Offer

Happy to provide full sample omp transcripts, test builds, or contribute the adapter if the schema contract is documented.