feat(workbench): application explorer — shell, Application tree, route workspace, one invocation API, page cuts (#600 PR 1)

.changeset/wb600-application-explorer.md

@@ -0,0 +1,5 @@
+---
+'agent-bundle': minor
+---
+
+Redesign the Workbench as an application explorer (#600 PR 1). The dev server gains one route invocation API — `POST /api/routes/invocations` renders any compiled route (MCP tool, resource, prompt, CLI route, script, or event route with a canonical or Claude/Codex/Cursor payload) through the production runtime and returns the render-event stream, final Agent Document, structured result, request context, providers, timings, and the MCP/CLI/host projections; `GET /api/routes/invocations[/<id>]` lists and replays this session's invocations and every completion is published as a `route.invocation` project event (diagnostics `AB8231`, `AB8232`, `AB8236`–`AB8238`). The foreground server serves the Workbench shell for its deep-link paths (`/routes/**`, `/trace`, `/problems`, `/sessions`, `/advanced`). Breaking for `agent-bundle/test`: `inspectWorkbenchSurface()` now reports the Application tree (`application`, with `workbenchLeafPath(leaf)`) and the populated Advanced sections instead of `pages`; `WorkbenchPageName` and `workbenchPageLabel` are removed. (#629)

.github/workflows/nightly.yml

@@ -55,34 +55,6 @@ jobs:
           if-no-files-found: ignore
           retention-days: 7
 
-  # Evidence-capture pool (rstest.evidence.config.ts): documentation-artifact
-  # browser journeys whose behavioural contracts the integration pool already
-  # proves per PR (see nightlyEvidenceTestFiles in rstest.integration-tests.ts).
-  # The pool's capture writes its screenshots and evidence.json to a temporary
-  # directory it removes on exit, so a green run leaves no artifact; the
-  # tracked copies in docs/assets are regenerated by hand (see
-  # examples/rsc-agent-runtime/README.md). Budget: install + build + the
-  # example's cold rsbuild in globalSetup + a 480 s capture watchdog.
-  evidence:
-    name: Evidence capture (Node 22.19)
-    runs-on: ubuntu-latest
-    timeout-minutes: 25
-    steps:
-      - uses: actions/checkout@v7
-      - uses: ./.github/actions/setup-workspace
-        with:
-          node-version: 22.19.0
-          playwright-browser: chromium
-      - run: pnpm test:evidence
-      - if: failure()
-        name: Preserve Playwright traces of failed browser tests
-        uses: actions/upload-artifact@v7
-        with:
-          name: playwright-traces-evidence
-          path: .rstest/playwright-traces
-          if-no-files-found: ignore
-          retention-days: 7
-
   # Official MCP server conformance runner (rstest.mcp-conformance.config.ts)
   # against one generated fixture over loopback HTTP. The expected-failure
   # allowlist lives in packages/agent-bundle/tests/fixtures/

docs/diagnostics.md

@@ -44,6 +44,7 @@ even when no error diagnostic was reported.
 | `AB8215`–`AB8218` | Workbench read-only host discovery route (`/api/discovery`): `AB8215` invalid path, `AB8216` query string or non-`GET` method (400/405), `AB8217` report over the 16 MiB response limit (413), `AB8218` discovery not available (503). |
 | `AB8219`–`AB8223` | Workbench live MCP probe route (user-initiated, read-only initialize + tools/list): `AB8219` invalid path, `AB8220` invalid request/method, `AB8221` probe target not found, `AB8222` response over the 16 MiB budget, `AB8223` probe unavailable. |
 | `AB8233`–`AB8235` | Workbench browser-side strict decoders rejecting a dev-server response: `AB8233` lifecycle replay, `AB8234` host discovery, `AB8235` MCP probe report. |
+| `AB8231`–`AB8232`, `AB8236`–`AB8238` | Workbench route invocation service (`/api/routes/invocations`): `AB8231` unknown route or invocation id, `AB8232` no published build / invocation manifest unavailable, `AB8236` render child timed out or crashed, `AB8237` malformed invocation request, and `AB8238` unknown fixture id. (`AB8233`–`AB8235` were already assigned to browser-side decoders.) |
 | `AB8110`–`AB8113` | Workbench standalone MCP Inspector routes (`/api/inspector/status`, `/api/inspector/launch`): `AB8110` invalid path, `AB8111` invalid request shape or query, `AB8112` the Inspector could not be launched (spawn failure, exit before publishing a URL, or the 30 s startup budget elapsed; 502), `AB8113` routes not available (404 when the launcher is not composed, 503 after shutdown). |
 | `AB8120`–`AB8123` | Workbench route manifest (`/api/routes/manifest`): `AB8120` invalid path, `AB8121` not available (404/409/503), `AB8122` query string on the request, `AB8123` the browser client could not decode the response (see below). |
 | `AB8024`–`AB8025` | Live host MCP proxy: epoch drift behind a host connection and dev-server unavailability (see below). |

examples/hooks-and-scripts/README.md

@@ -20,25 +20,25 @@ modes covered.
 
 ## Workbench walkthrough
 
-1. **Overview** is the Bundle dashboard. It relates the authored Hook to its
-   emitted artifact, exercise trace, and evaluation pages; its status is the
-   authoritative current-or-stale epoch state.
-2. **Hooks** defaults to the Claude `sessionStart` binding and its populated
-   inline canonical JSON, including `"source": "workbench"`. Run the
+1. The shell header reports the authoritative current-or-stale epoch state and
+   links build diagnostics to **Problems**.
+2. Under **Application → Events / Hooks**, select the Claude `sessionStart`
+   binding and its populated inline canonical JSON, including
+   `"source": "workbench"`. Run the
    simulation, then use **Replay saved simulation** to rerun exactly that
    epoch-bound input. The result directs the release session through both
    checks.
-3. **Playground** defaults to Script execution, the Claude target, and
-   `verify-release`. Run it and wait until the session is finalized: the
+3. Under **Application → Scripts**, select `verify-release` and the Claude
+   target, then run it. The
    emitted script reads the packaged `release/release-manifest.json` relative
    to its module and reports release 2.4.0 ready for packaging.
 4. Change the target to portable and select `detect-risk`. Its emitted script
    reads `release/risk-register.json`, reports high-severity `REL-204`, exits
    with code 2, and finalizes a durable blocking trace.
-5. **Logs** filters those producer records by producer, level, kind, or
-   context; open a record to inspect raw details. **Artifacts** is the emitted
-   file/provenance view, while **Comparisons** aligns outcomes only after two
-   recorded eval runs.
+5. **Advanced → Raw logs** filters those producer records by producer, level,
+   kind, or context; open a record to inspect raw details. **Advanced →
+   Artifact** is the emitted file/provenance view, while **Advanced → Evals →
+   Compare** aligns outcomes only after two recorded eval runs.
 
 ## Reversible diagnostic walkthrough
 

examples/host-test/README.md

@@ -123,17 +123,18 @@ the printed command, open the Agents pane, and use the same scenario prompt.
 
 ## Workbench walkthrough
 
-1. **Overview** lists the twenty event routes, both MCP servers, the skill, and
-   the routed CLI with their per-target capability judgments — `workspace/open`
+1. **Application** lists the twenty event routes, both MCP servers, the skill,
+   and the routed CLI with their per-target capability judgments — `workspace/open`
    is Cursor-only, `task/*` and `file/change` are Claude-only, and portable
    carries no hooks at all.
-2. **Hooks** simulates any family with canonical input; the route appends a
-   record to the log and returns an empty result (only `session/start` speaks
-   an `additional_context` line naming the log path).
-3. **Playground** runs `host-test dump` and the `dump` tool against the same
-   log, so a simulated hook is visible from the MCP surface immediately.
-4. **Lifecycles** replays checked-in native receipts and shows the request
-   context — and lineage — each replay mounted.
+2. Under **Application → Events / Hooks**, select any family and run it with
+   canonical input; the route appends a record to the log and returns an empty
+   result (only `session/start` speaks an `additional_context` line naming the
+   log path).
+3. Run the `host-test dump` CLI leaf or the MCP `dump` Tool leaf against the
+   same log, so a simulated hook is visible from the MCP surface immediately.
+4. Use the selected event route's **Replay** tab for checked-in native receipts;
+   each replay shows its request context and lineage.
 
 ## Noninteractive checks
 

examples/mcp-app/README.md

@@ -28,33 +28,32 @@ and Claude artifacts; the App resource remains portable.
 
 ## Workbench walkthrough
 
-1. **Overview** opens on the Bundle dashboard. Its Author, Build, Exercise,
-   and Evaluate stages connect the source capability to its emitted artifact,
-   runtime evidence, and eval result.
-2. **Skills** defaults to `service-readiness`; compare its authored status
-   policy and readiness-report resource with generated output and its explicit
-   eval coverage. **Hooks** defaults to a populated Claude `sessionStart`
-   canonical input; run the simulation to attach the readiness workflow.
-3. **Playground** defaults to Script execution, the Claude target, and
-   `check-service-fixture`. Run it and wait for the finalized session. The
+1. The shell header connects build status and diagnostics to the current
+   emitted artifact.
+2. Under **Application → Skills**, select `service-readiness`; compare its
+   authored status policy and readiness-report resource with generated output
+   and its explicit eval coverage. Under **Application → Events / Hooks**,
+   select `sessionStart`, use the populated Claude canonical input, and run the
+   simulation to attach the readiness workflow.
+3. Under **Application → Scripts**, select `check-service-fixture`, choose the
+   Claude fixture, and run it. The
    emitted checker resolves the packaged status fixture beside its emitted
    module, so it succeeds without depending on the shell working directory.
-4. **Logs** exposes the resulting producer records. In **Artifacts**, select
-   portable to inspect `mcp-apps/status.html`; Codex and Claude retain their
-   host artifacts but not this portable App resource.
-5. Before recording two eval runs, **Comparisons** deliberately displays:
+4. **Advanced → Raw logs** exposes the resulting producer records. In
+   **Advanced → Artifact**, select portable to inspect `mcp-apps/status.html`;
+   Codex and Claude retain their host artifacts but not this portable App resource.
+5. Before recording two eval runs, **Advanced → Evals → Compare** deliberately displays:
    `At least two recorded runs are needed before a comparison can be aligned.`
    That is the precise empty state, not an error.
-6. In **MCP playground**, the defaults are portable and the `status` server.
-   Open the session, list tools, select `show-status`, choose `payments-api`,
-   and invoke it. Invocation history shows the degraded summary and labelled
+6. Under **Application → MCP → status → Tools**, select `show-status`, choose
+   `payments-api`, and run it. Invocation history shows the degraded summary and labelled
    Availability and P95 latency checks (the latter fails). Open the App preview:
    the rendered panel also shows `payments-api`, a text-labelled amber
    `degraded` indicator, the same summary, and passing/failing checks through
    the MCP Apps bridge. Inspect the
-   protocol trace, use **Restart MCP session**, then close, reset, and reopen
-   it to exercise the lifecycle.
-7. **Evals** defaults to the deterministic `mcp-app-status` suite. Run
+   protocol trace in the route workspace; use **Advanced → Protocol** for
+   session restart, reset, and lifecycle inspection.
+7. **Advanced → Evals → Runs** defaults to the deterministic `mcp-app-status` suite. Run
    `status-is-healthy` and inspect its completed passing trial attributed to
    `service-readiness`; it reads only checked-in fixture data and needs no
    native login or API key.

examples/rsc-agent-runtime/README.md

@@ -338,7 +338,8 @@ surface is not treated as stable here: React `19.2.8`, `react-dom` `19.2.8`,
 Existing Agent Bundle skills, static MCPs, evaluations, and normal hooks neither require nor activate this runtime. Nothing under `packages/agent-bundle` imports the example or React/RSC runtime packages.
 
 `PlaygroundService` is the landed, provider-neutral durable whole-plugin
-authoring timeline foundation. Runtime Playground history is deliberately
+authoring timeline foundation. Runtime history now backs the selected
+Application route workspace rather than a separate destination. It remains
 provider-session-scoped and ephemeral in this example; wiring a provider
 adapter, authenticated API, timeline UI, durable Runtime export, or evaluation
 promotion onto that history is an explicit non-goal of this demo.

examples/skills-starter/README.md

@@ -31,17 +31,17 @@ required. Both eval suites are deterministic and read only checked-in fixtures.
 
 ## Workbench walkthrough
 
-1. **Overview** opens on the Bundle dashboard. It summarizes the three Skills,
-   generated targets, build health, and the next useful actions.
-2. **Skills** lists `dependency-upgrade`, `incident-triage`, and
-   `release-review`. Browse their linked checklists and report templates. Switch
-   between Source and Generated to see whether a target copied or adapted the
-   authored document. Every Skill shows its deterministic outcome-eval coverage;
+1. The shell header summarizes build health and current diagnostics.
+2. Under **Application → Skills**, select `dependency-upgrade`,
+   `incident-triage`, or `release-review`. Browse its linked checklists and
+   report templates. Use the inspector to compare Source and Generated output
+   and see whether a target copied or adapted the authored document. Every
+   Skill shows its deterministic outcome-eval coverage;
    it is labeled indirect because the deterministic harness cannot observe host
    Skill activation.
-3. **Artifacts** defaults to the Claude target. Change the target to compare
+3. **Advanced → Artifact** defaults to the Claude target. Change the target to compare
    the portable, Codex, and Claude output trees and their provenance.
-4. **Evals** defaults to the `release-readiness` suite. Run its deterministic
+4. **Advanced → Evals → Runs** defaults to the `release-readiness` suite. Run its deterministic
    `release-artifact-is-ready` case and inspect the passing trial. It consumes
    only the checked-in evidence fixture, so no model login or API key is needed.
 5. To practice repair, make a reversible policy edit, press **Rebuild**, and

package.json

@@ -16,7 +16,6 @@
     "test:projection": "rstest --config rstest.projection.config.ts",
     "test:integration": "pnpm build && pnpm test:integration:run",
     "test:integration:run": "AGENT_BUNDLE_WORKBENCH_PREBUILT=1 AGENT_BUNDLE_PACKAGE_PREBUILT=1 rstest --config rstest.integration.config.ts",
-    "test:evidence": "pnpm build && AGENT_BUNDLE_WORKBENCH_PREBUILT=1 AGENT_BUNDLE_PACKAGE_PREBUILT=1 rstest --config rstest.evidence.config.ts",
     "test:mcp-conformance": "pnpm build && AGENT_BUNDLE_MCP_CONFORMANCE=1 rstest --config rstest.mcp-conformance.config.ts",
     "test:native-host": "pnpm build && AGENT_BUNDLE_NATIVE_HOST_CONTRACTS=1 rstest --config rstest.native-host.config.ts",
     "test:watch": "rstest --config rstest.config.ts --watch",

packages/agent-bundle/rslib.config.ts

@@ -175,6 +175,7 @@ export default defineConfig({
       // route module may import at run time without pulling the compiler
       // into its generated bundle.
       routes: './src/routes/public.ts',
+      'route-invocation-child': './src/dev/routes/route-invocation-child.ts',
       rstest: './src/rstest/index.ts',
       'terminal-capability': './src/terminal-capability.ts',
       test: './src/test/index.ts',

packages/agent-bundle/src/contracts/application.ts

@@ -0,0 +1,23 @@
+export {
+  applicationLeafForRouteId,
+  applicationLeaves,
+  applicationTreeForManifest,
+  filterApplicationTree,
+  findApplicationLeaf,
+  firstApplicationLeaf,
+} from '../dev/routes/application-tree.ts';
+export type {
+  ApplicationGroup,
+  ApplicationGroupKind,
+  ApplicationLeaf,
+  ApplicationLeafExecution,
+  ApplicationServerGroup,
+  ApplicationSubgroup,
+  ApplicationTree,
+  ApplicationTreeInspectionHook,
+  ApplicationTreeInspectionMcpServer,
+  ApplicationTreeInspectionScript,
+  ApplicationTreeManifestSources,
+  ApplicationTreeSkill,
+  ApplicationTreeState,
+} from '../dev/routes/application-tree.ts';

packages/agent-bundle/src/contracts/invocations.ts

@@ -0,0 +1,26 @@
+/**
+ * Browser-consumable contract surface for dev-server route invocations — the
+ * one execution path behind the Workbench route workspace. Type-only: routes
+ * render on the server through the production runtime.
+ */
+export type {
+  RouteInvocationCliProjection,
+  RouteInvocationEvent,
+  RouteInvocationEventHost,
+  RouteInvocationEventOptions,
+  RouteInvocationEventPayload,
+  RouteInvocationHostProjection,
+  RouteInvocationKind,
+  RouteInvocationListResponse,
+  RouteInvocationProjection,
+  RouteInvocationProvider,
+  RouteInvocationProviderStatus,
+  RouteInvocationRequest,
+  RouteInvocationStatus,
+  RouteInvocationSummary,
+  RouteInvocationTiming,
+} from '../dev/routes/route-invocation.ts';
+export type {
+  RouteInvocation,
+  RouteInvocationResponse,
+} from '../dev/routes/route-invocation-result.ts';

packages/agent-bundle/src/contracts/workbench-shell.ts

@@ -0,0 +1,19 @@
+export {
+  applicationNodeKey,
+  applicationNodePath,
+  applicationNodeRefForPathSegments,
+  applicationNodeRefForRouteId,
+  isApplicationMcpNodeKind,
+  routeIdForApplicationNodeRef,
+  sameApplicationNodeRef,
+} from '../dev/routes/application-node.ts';
+export type {
+  ApplicationMcpNodeKind,
+  ApplicationNodeKind,
+  ApplicationNodeRef,
+} from '../dev/routes/application-node.ts';
+export {
+  isWorkbenchShellPath,
+  workbenchShellAreas,
+} from '../dev/workbench-shell-paths.ts';
+export type { WorkbenchShellArea } from '../dev/workbench-shell-paths.ts';

packages/agent-bundle/src/dev/coordinator.ts

@@ -86,6 +86,7 @@ export interface DevCoordinatorOptions {
   readonly initialPreparedProject?: PreparedProject;
   readonly ignoredPaths?: readonly string[];
   readonly onPreparedProject?: (prepared: PreparedProject) => Promise<void>;
+  readonly onPublishedProject?: (prepared: PreparedProject, epoch: ArtifactEpoch) => void;
   readonly outputPaths?: readonly string[];
   /** Rebuilds the framework-owned package build (bin/lib) after successful artifact rebuilds. */
   readonly packageBuildService?: DevPackageBuilder;
@@ -209,6 +210,7 @@ export class DevCoordinator {
   readonly #ignoredPaths: readonly string[];
   readonly #outputPaths: readonly string[];
   readonly #onPreparedProject: ((prepared: PreparedProject) => Promise<void>) | undefined;
+  readonly #onPublishedProject: ((prepared: PreparedProject, epoch: ArtifactEpoch) => void) | undefined;
   readonly #packageBuildService: DevPackageBuilder;
   readonly #prepareCommand: 'build' | 'dev';
   readonly #projectService: ProjectPreparer;
@@ -248,6 +250,7 @@ export class DevCoordinator {
     this.#now = options.now ?? (() => new Date());
     this.#nextPreparedProject = options.initialPreparedProject;
     this.#onPreparedProject = options.onPreparedProject;
+    this.#onPublishedProject = options.onPublishedProject;
     this.#packageBuildService = options.packageBuildService ?? new DevPackageBuildService();
     this.#outputPaths = Object.freeze([...new Set([
       ...(options.outputPaths ?? ['dist']),
@@ -563,6 +566,7 @@ export class DevCoordinator {
         startedAt: running.startedAt,
       });
       this.#activeEpoch = result.epoch;
+      this.#onPublishedProject?.(prepared, result.epoch);
       const artifact = artifactStatusFor(this.#activeEpoch, source.revision);
       this.#status = freezeProjectStatus({
         artifact,

packages/agent-bundle/src/dev/events.ts

@@ -82,6 +82,7 @@ const eventTypes = new Set<ProjectEventType>([
   'artifact.status',
   'dev.contract.status',
   'dev.host.sync',
+  'route.invocation',
   'runtime.event',
 ]);