[omartelo]

lich's half of oh-my-pi support. The plugin's half is omartelo/lich-plugin#18 — this needs a release carrying omp/lich.js before an install can succeed, since the install fetches that path at the latest tag.

What

The plugin installs (internal/agentplugin/omp.go). omp has no plugin CLI, so lich writes the released extension into omp's own extensions/ directory — the file being there is the install, and the marker line above it is the only record of the version, since omp keeps no plugin state to ask. Settings › lich plugin lists it beside the others; no UI change was needed, the row comes from Status().

The MCP server is registered. omp takes no --mcp-config flag, and the only place it reads a server from is an mcpServers document beside the extensions, so the install merges a lich entry into mcp.json. Every other server and every key lich knows nothing about survive the round trip. A document lich cannot parse aborts the install before anything is written, rather than being replaced — overwriting one lich failed to understand would delete servers it cannot see.

omp sessions resume. omp -r <id> reopens a conversation, and omp files one JSONL per session — so ResumeAvailable has something to prove one still exists, which is exactly what opencode and Crush lack (both keep conversations in SQLite). Cards that ran omp now get the same "continue where it left off" prompt Claude Code and Codex cards get.

The contract carries omp. docs/hooks/ gains its column in every mapping table and its case in the session-start fixtures.

On the "blocker" the plugin side reported

The plugin PR reports that provider: "omp" would be rejected with a 400 until omp is registered. It would not: the endpoint validates against providers.Known (internal/terminal/transport.go), and omp has been in providers.Registry since it was added. What was missing was only a fixture case naming it — docs/hooks/fixtures/session-start.jsonl is a sample of cases, not the provider list (claude has no case of its own either). The case added here is now exercised by the real parser in TestHookFixturesMatchParsers, which is the proof.

Measured, not guessed

Everything provider-specific was read off the binary on this machine (omp v17.3.0):

• -r, --resume=<value> — resume by id.
• omp config path → ~/.omp/agent; with OMP_PROFILE=work → ~/.omp/profiles/work/agent; with PI_CODING_AGENT_DIR=/tmp/x → /tmp/x; with both, the profile wins. That precedence is pinned by a test, because getting it backwards means the install lands where omp is not reading and every restored card silently starts fresh.

Known gaps, both the harness's own

• No waiting. omp declares tool_approval_requested but no real run was observed emitting it, and a report wired to an event name that never fires is one that silently never arrives. So a session waiting on permission shows a spinner rather than a bell — Settings says so (OMP_APPROVAL_HINT), following the precedent of the Crush hint, because a missing bell reads as a broken install.
• No idle, for opencode's reason: the extension runs in-process, so what would report the CLI leaving dies with it.

Test plan

• gofmt -l . clean, go vet ./... clean
• go test ./internal/... green (1344), -race on agentplugin and terminal
• Cross-compile loop for linux/darwin/windows: go build + go vet clean on each
• biome check, tsc --noEmit, vitest run (875) and vite build green
• New coverage: install writes the extension with its marker and registers the server; a pre-existing mcp.json keeps its own servers and its unknown keys; an unparseable one is refused with nothing written; a 404 release leaves neither file behind; ompAgentDir precedence; resumeArgs and ResumeAvailable for omp, including that omp never matches a Codex rollout
• Not exercised against a real install — that needs the plugin release above. The chain is covered to the bytes on disk, which is where lich's half ends.

[omartelo]

Validated end to end against the real release (lich-plugin v0.10.0), on Linux, in an isolated omp agent directory (PI_CODING_AGENT_DIR) — the user's own omp config was never touched.

1. The install, run through this branch's own code path (Service.Install("omp"), real network, real release):

• extensions/lich.js written, 7232 bytes, first line // lich-plugin v0.10.0 — installed by lich, replace with the file from omartelo/lich-plugin
• installedVersion("omp") reads back 0.10.0
• a pre-existing mcp.json carrying {"mcpServers":{"notes":…},"other":1} came out with lich merged in and both notes and other intact:

{
  "mcpServers": {
    "lich": { "args": ["mcp"], "command": "…/lich" },
    "notes": { "command": "notes-server" }
  },
  "other": 1
}

2. omp loads what lich wrote, and reports. Running omp against a stub listener on the coordinates lich injects — no --hook flag, purely the autodiscovery of extensions/:

/session-start?token=…  {"session_id":"probe-session","provider_session_id":"019ffbbe-3f1b-7000-90dd-0184132e5d80","provider":"omp"}
/hook?token=…           {"session_id":"probe-session","state":"busy"}
/hook?token=…           {"session_id":"probe-session","state":"done"}

The turn itself died on a deliberately invalid API key, and the reports still arrived — start and end of turn do not depend on the model call.

3. Resume, against the transcript that run actually wrote. ResumeAvailable("omp", <that id>) → true, a made-up id → false, and the glob resolved the real file (sessions/-tmp/2026-08-13T15-29-37-051Z_<id>.jsonl). Then the invocation itself: omp -r <id> re-entered the same conversation — its session-start reported the identical provider_session_id rather than a new one, which is the whole contract resumeArgs rides on.

One thing worth knowing for the manual check: omp installs into ~/.cache/.bun/bin, which is on PATH only in a login shell. lich resolves its environment through $SHELL -lic env at boot (ResolveShellEnv), so Available is true in the app — but a bare non-interactive shell will not find omp, which is what makes the Settings row say "CLI not installed" if you go looking from one.

Probes were disposable and are gone; nothing in the diff changed since the checks above.