fix(opencode): support OpenCode 2 plugin API (V1/V2 dual export)

[Scratchydisk]

Supersedes #1169 (closed to allow further local testing — same branch, force-pushed since to rebase onto current main, which is why GitHub wouldn't let me reopen it directly).

Summary

OpenCode 2's plugin loader ignores the V1 { id, server } shape this package's src/adapters/opencode/plugin.ts exports — per OpenCode's own migration guide (opencode.ai/v2/docs/migrate-v1): "V1 plugin implementations do not run in V2." It only recognizes a default export produced by Plugin.define({ id, setup }) from the new @opencode/plugin package.

Repro against the currently-published package (verified against a real opencode@2.0.3/2.0.4 install)

// opencode.jsonc
{ "plugins": ["context-mode"] }

level=WARN message="failed to load plugin" target=context-mode
cause="Cause([Fail(PluginModule.LoadError: Plugin must export a default
definition with an id and an effect or setup function. (cause:
SchemaError(Missing key at ["default"]["effect"] Missing key at
["default"]["setup"])))])"

ctx_* tools never register, no hooks fire, and opencode plugin list silently shows nothing for it — worse than a crash, since there's no indication anything is wrong until a user notices the tools are gone.

Fix

Adds a V2 setup() entrypoint alongside the existing V1 server(), following OpenCode's own documented "support V1 and V2 from one package" pattern (opencode.ai/v2/docs/build/plugins#support-v1): spread Plugin.define({ id, setup }) into the same default-export object that already carries server. OpenCode 1.x and KiloCode keep calling server() — completely unchanged — while OpenCode 2 reads id/setup() off the same object and ignores server().

The V2 setup() re-implements all five hooks using OpenCode 2's documented mapping (opencode.ai/v2/docs/build/plugins/migrate-v1):

V1 hook	V2 API
tool.execute.before	ctx.tool.hook("execute.before", ...)
tool.execute.after	ctx.tool.hook("execute.after", ...)
event (message.updated)	ctx.event.subscribe()
chat.message	ctx.session.hook("prompt", ...)
experimental.session.compacting	ctx.session.hook("compaction", ...)
experimental.chat.system.transform	ctx.session.hook("context", ...)
tool map	ctx.tool.transform(editor => editor.add(...))

The custom ctx_* tools now declare JSON Schema (required by V2) instead of a Zod shape, reusing the existing zod3ShapeToV4 conversion (added for #574) plus Zod v4's own toJSONSchema().

@opencode/plugin is imported dynamically (await import(...) inside a try/catch), not as a static top-level import, so a stale/partial install missing that dependency degrades to V1-only instead of breaking the whole module for existing OpenCode 1.x/KiloCode users.

Verification

This branch has now been tested two ways:

1. Sandboxed, against the real binary: patched a real opencode@2.0.3 package cache and confirmed the unmodified published package fails with the exact error above while this branch's build loads with no error.
2. Live production install: patched ~/.cache/opencode/.../node_modules/context-mode on a machine actually running this plugin day-to-day (real opencode.json with "plugins": ["context-mode", ...], real provider credentials), restarted the OpenCode background service (which picked it up along with an unrelated OpenCode 2.0.3→2.0.4 self-update), and confirmed:
   • Every plugin load since the restart is clean — zero "failed to load plugin" entries for context-mode in the service log (previously, every single load attempt failed with the schema error above).
   • Ran a live session that explicitly invoked the ctx_stats tool and got back a fully-formed real report pulled from the actual session database (event counts, per-tool capture stats, lifetime cost) — end-to-end proof that Plugin.define/setup() → ctx.tool.transform → tool execution → session DB all work correctly under a real OpenCode 2 session.

All 122 pre-existing opencode-plugin/adapters/opencode/parse-opencode-usage tests pass unchanged — zero behavior change to the V1/KiloCode path. Plus 2 new tests locking in the dual-export shape. npm run typecheck and npm run build both clean.

Known gap for maintainer review

The exact field names on OpenCode 2's ToolExecuteBefore/ToolExecuteCompleted/ToolExecuteFailed event types aren't spelled out in the public docs beyond short examples (event.tool, event.input, event.status, event.result, event.error). I inferred event.sessionID by analogy with every other hook interface OpenCode documents (SessionRequestHook, SessionRetryHook, PermissionEvaluation, etc. all expose it), and defensively no-op when it's absent. The routing-enforcement path (tool.execute.before/after → session event capture) hasn't yet been exercised against a live agent turn that actually triggers a blocked/modified tool call — worth a maintainer smoke test before release, though the ctx_* tool-execution path itself (registration, JSON Schema conversion, execute, result handling) is now confirmed working live.

Test plan

• npm run typecheck
• npm run build
• npx vitest run tests/opencode-plugin.test.ts tests/adapters/opencode.test.ts tests/session/parse-opencode-usage.test.ts (122 passed)
• Live OpenCode 2 session: plugin loads cleanly, ctx_stats tool call succeeds end-to-end against a real session DB
• Maintainer: exercise tool.execute.before/after (routing enforcement + session capture) against a live OpenCode 2 session with a real blocked/modified tool call