feat(#85): add OMP (Oh My Pi) as a 4th delivery format

[jsirish]

Summary

• Adds .omp-plugin/ as a 4th throughline delivery format, targeting OMP (Oh My Pi) - a TypeScript coding-agent harness with its own plugin/hook/skill/MCP system.
• Follows the Codex reuse pattern, not OpenCode's full port: OMP's hooks are in-process pi.on() event handlers, not subprocess hooks, so hooks/post/throughline.ts is a thin shim that translates OMP's own event shapes into the JSON the existing hooks/*.sh scripts already jq-parse, then shells out to them. No capture/redaction logic is duplicated in TypeScript.
• Skills are auto-discovered via .omp-plugin/skills, a symlink to the shared skills/ directory - same pattern as .omp-plugin/hooks -> ../hooks. Zero extra files, zero drift risk.
• Install: git clone + omp plugin link .omp-plugin (a live checkout, not a package install - git pull to update).

How this differs from the initial plan

The plan (session notes, not committed) assumed OMP's hook API surface based on docs alone. Installing the real package locally (npm install @oh-my-pi/pi-coding-agent) and reading its actual TypeScript source corrected several assumptions:

• Hooks discovered under hooks/pre|post/ are bound through OMP's Extensions runtime (ExtensionAPI), not the separate, unused-at-runtime HookAPI module the docs and initial research implied.
• pi.exec() has no stdin option, so the shim calls node:child_process directly.
• omp plugin link <path> (not omp plugin install <git-url>) is the mechanism that installs a subdirectory of a monorepo - this is what makes .omp-plugin/ viable at all without a standalone npm package.
• OMP's own tool schemas (bash/write/grep use command/path/pattern, confirmed against src/tools/*.ts) needed translating into Claude Code's tool-shape (Bash/Edit/Write/Grep with tool_input.*) for session-capture.sh to parse correctly.

What's verified vs. still open

Verified (against real OMP source + a bun-based integration test suite that fires synthetic events at the real shim and asserts on real hooks/*.sh output, not a mocked child_process):

• All 6 registered events fire the right hook script with the right JSON shape.
• session_start sends the onboarding markdown via pi.sendMessage.
• tool_result correctly captures bash/edit/write/grep/mcp__*, skips read/glob (matching Claude Code's own exclusion), and drops unrecognized non-MCP tools rather than guessing a shape.
• session_before_compact/session_compact stamp the compaction boundary and re-inject the buffer tail (OMP splits Claude Code's single re-fired SessionStart(source=compact) into two events).
• session_shutdown stamps the buffer as ended.
• All 186 existing shell hook tests still pass - no regression to the Claude Code/Codex path.

Still open, tracked in #85 (needs a real OMP session, not just static/simulated verification):

• Exact field names for web_search/task tool inputs (best-effort mapping in place; falls back to empty fields rather than crashing if wrong).
• session_shutdown reliability on a clean exit (/quit/EOF) vs. only SIGINT/SIGTERM.
• Confirming pi.sendMessage() at session_start actually reaches the model on turn 1, not just the TUI.

I don't have a configured OMP+model setup in this environment to complete that last-mile verification myself - #85 stays open for it.

Test plan

• bun test in .omp-plugin/ - 9/9 pass (integration tests against the real shim + real hook scripts)
• bun run typecheck in .omp-plugin/ - clean
• sh tests/run.sh - 186/186 pass, no regression
• local-ci.sh --strict - all checks PASS (shellcheck, manifest validation, plugin version agreement, hook tests, typography, opencode plugin, omp plugin)
• Live verification against a real OMP session (see "still open" above) - needs a maintainer with OMP + a configured model to complete

🤖 Generated with Claude Code

[jsirish]

Live-verified against a real OMP session (brew-installed `omp` 18.2.6, Jason's Dynamic LLM gateway model). All previously-open items from #85 now confirmed working except web_search/task field names (untested, low risk, degrades gracefully):

• `session_start` -> `sendMessage()` reaches the model on turn 1 (model referenced the real injected HANDOFF/buffer content in its first reply)
• `before_agent_start` prompt capture
• `tool_result` capture for bash/grep with correct field extraction
• `session_shutdown` fires on a clean `--print`-mode exit
• Skills auto-discovery (all 4 skills visible to the model)
• Git-worktree data-dir sharing resolves correctly to the main checkout

Details in #85. Ready to merge from my side.