fix(drive): repair compaction and open-picker probes for current V2

.changeset/await-plugin-activation.md

@@ -0,0 +1,5 @@
+---
+"opencode-drive": patch
+---
+
+Bump the pinned `@opencode-ai/client` to `0.0.0-dev-18911` so the generated SDK exposes `plugin.awaitActivation`, which scripts need before reading agents or models from a cold location on current V2 servers.

packages/drive/package.json

@@ -61,7 +61,7 @@
     "@effect/platform-node": "4.0.0-rc.112",
     "@effect/platform-node-shared": "4.0.0-rc.112",
     "@napi-rs/canvas": "1.0.2",
-    "@opencode-ai/client": "0.0.0-dev-18862",
+    "@opencode-ai/client": "0.0.0-dev-18911",
     "@opentui/core": "0.4.5",
     "@types/bun": "1.3.13",
     "@typescript/native-preview": "7.0.0-dev.20251207.1",

packages/drive/test/manual/tui-regressions/README.md

@@ -306,6 +306,13 @@ and terminal frame in `state-machine-failure.json`.
 | `cancelled` | Cancellation removes the known pending item while setup is blocked. The client still proposes a fresh compaction ID and completes it after healing; cancellation itself does not permanently reserve the old ID. |
 | `rollback`  | Killing connections before admission removes the unacknowledged row and displays the transport error. After healing, retry compaction and submit a recovery prompt successfully.                                 |
 
+The simulated compaction reply is a structured summary with the `##` section
+headings that OpenCode requires since PR #46751; a plain sentence now settles as
+`compaction.failed` ("did not match the required template") instead of
+`completed`. The fake model routes compaction requests on the stable prompt
+frame ("Summarize only the history shown"), not on template wording, which has
+already changed once since.
+
 The simulated model paces a busy-step checkpoint before holding the response
 open; this isolates admission behavior from stream-publication behavior. The
 bounded-flush tests separately verify complete single bursts without a later
@@ -364,7 +371,10 @@ Drive writes current V2 `cli.json` under its isolated `OPENCODE_CONFIG_DIR`.
 The probe expresses `session.new_location: "inherit"` and a `Ctrl-N` binding
 through `tuiConfig`, making the movement check an explicit fixture contract
 without a manual config-file workaround or installed-client changes. Fixture
-sessions also select the simulation model and build agent explicitly.
+sessions also select the simulation model and build agent explicitly. Since
+OpenCode PR #46639 plugins load after config, so a cold location answers
+`agent.list` and `model.default` with empty data until activation; the fixture
+calls `plugin.awaitActivation` for both project locations before reading them.
 
 Successful runs write `open-picker-report.json` and print `passed: true`.
 Failures retain the case, viewport, marker seed, checkpoint trace, server
@@ -413,6 +423,19 @@ bun run --cwd packages/drive drive start --name tui-multi-tool-interleavings \
   --dev "$OPENCODE_DEV"
 ```
 
+This probe currently fails on V2 at or after OpenCode PR #46724
+([`anomalyco/opencode#37650`](https://github.com/anomalyco/opencode/issues/37650)) with
+`InvalidRequestError: Expected JSON value at ["data"][2]["metadata"]["hidden"]`
+from `GET /api/session/:id/permission`. That is a server defect, not probe
+drift: the glob tool copies every optional input into its permission
+`metadata` (`path`, `limit`, and now `hidden`), so an omitted input leaves an
+`undefined` value that the response encoder for `Record(String, Unknown)`
+rejects, and the endpoint answers 400 while that permission is pending. Any
+model call that omits `hidden` (or `path`, or `limit`) trips it. The probe
+deliberately keeps its realistic `{ pattern, path, limit }` input rather than
+adding `hidden: false` to dodge the bug; expect it to pass again once the
+server drops or defaults undefined permission metadata.
+
 ## Plugin registry
 
 `plugin-registry.ts` checks that plugin registrations reach both the frontend and the model without a server restart. A discovered local plugin registers a command and a tool; the probe asserts the command through `command.list`, the tool through the captured provider request body, and the slash menu through the TUI. It then writes a second plugin into `.opencode/plugins/` at runtime, asserts both appear, removes it, and asserts it disappears. Activation is observed through `plugin.list`.

packages/drive/test/manual/tui-regressions/compact-admission.ts

@@ -7,6 +7,14 @@ import { saveFailure } from "./state-machine.js"
 // Run each scenario independently so a failed checkpoint cannot mask later cases.
 const scenario = process.env.OPENCODE_DRIVE_COMPACT_CASE ?? "ordered"
 const cols = Number(process.env.OPENCODE_DRIVE_COLS ?? 100)
+const compactionSummary = [
+  "## Objective",
+  "- Inspect the admission fixture README.",
+  "",
+  "## Work State",
+  "### Completed",
+  "- The fixture README was inspected.",
+].join("\n")
 
 export default defineScript({
   network: true,
@@ -55,8 +63,10 @@ export default defineScript({
         yield* llm.serve((request) => {
           const body = JSON.stringify(request.body)
           if (body.includes("title generator")) return Stream.make(Llm.text("Admission fixture"))
-          if (body.includes("Create a new anchored summary") || body.includes("Update the anchored summary below"))
-            return Stream.make(Llm.text("The fixture README was inspected."))
+          // Compaction (OpenCode PR #46751) asks for a structured summary and
+          // rejects a reply without one of its `##` section headings as
+          // `compaction.failed`. Match the prompt frame, not the template text.
+          if (body.includes("Summarize only the history shown")) return Stream.make(Llm.text(compactionSummary))
           // History includes previous markers. Only the newest prompt selects a reply.
           const marker = ["CA_WARM", "CA_HOLD", "CA_NEXT"].reduce((latest, candidate) =>
             body.lastIndexOf(candidate) > body.lastIndexOf(latest) ? candidate : latest,

packages/drive/test/manual/tui-regressions/open-picker.ts

@@ -59,10 +59,14 @@ export default defineScript({
         const location = yield* opencode.location.get({ location: { directory: `${artifacts}/files` } })
         const archive = yield* opencode.location.get({ location: { directory: `${artifacts}/files/archive` } })
         assert.notEqual(archive.project.id, location.project.id, "archive must be a distinct Git project")
+        // Plugins load after config (OpenCode PR #46639), so a cold location
+        // answers agent and model reads with empty data until activation.
+        yield* opencode.plugin.awaitActivation({ location })
+        yield* opencode.plugin.awaitActivation({ location: archive })
         const model = (yield* opencode.model.default({ location })).data
         const agent = (yield* opencode.agent.list({ location })).data.find((agent) => agent.id === "build")
-        assert(model)
-        assert(agent)
+        assert(model, "no default model after plugin activation")
+        assert(agent, "no build agent after plugin activation")
         // Created before frontend launch: these are neither open tabs nor event-
         // hydrated sessions. Distinct titles fit the 44-column picker as well.
         const older = yield* opencode.session