feat: add agent-bundle compiler and developer workbench

.github/dependabot.yml

@@ -0,0 +1,10 @@
+version: 2
+updates:
+  - package-ecosystem: npm
+    directory: /
+    schedule:
+      interval: weekly
+  - package-ecosystem: github-actions
+    directory: /
+    schedule:
+      interval: weekly

.github/workflows/ci.yml

@@ -0,0 +1,77 @@
+name: CI
+
+on:
+  pull_request:
+  push:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  # Fast end-to-end confidence gate: builds, validates, and deterministically evals the
+  # checked-in micro fixture through the real CLI. Never skip-gated and needs no native host.
+  spot-check:
+    name: Micro-eval spot-check (Node 22.19)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: pnpm/setup@v1
+        with:
+          cache: true
+          install: false
+          runtime: node@22.19.0
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm examples:check
+      - run: pnpm test:spot-check
+
+  verify:
+    name: Verify (Node ${{ matrix.node-version }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        node-version: ['22.19.0', '24', '26']
+    steps:
+      - uses: actions/checkout@v7
+      - uses: pnpm/setup@v1
+        with:
+          cache: true
+          install: false
+          runtime: node@${{ matrix.node-version }}
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm exec playwright install --with-deps chrome
+      # Build first: checked-in suites and API tests import the package's built type
+      # declarations, so typecheck requires dist (same order as `pnpm check`).
+      - run: pnpm build
+      - name: Package lint (publint)
+        run: pnpm lint:package
+      - run: pnpm typecheck
+      - run: pnpm lint
+      - run: pnpm test
+
+  release-gates:
+    name: Release gates (Node 22.19)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: pnpm/setup@v1
+        with:
+          cache: true
+          install: false
+          runtime: node@22.19.0
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm exec playwright install --with-deps chrome
+      - run: pnpm check:release
+
+  dependency-review:
+    if: github.event_name == 'pull_request'
+    name: Dependency review
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: actions/dependency-review-action@v4

.github/workflows/native-host-smoke.yml

@@ -0,0 +1,53 @@
+name: Native host smoke
+
+on:
+  workflow_dispatch:
+    inputs:
+      host:
+        description: Signed-in host CLI to exercise
+        required: true
+        default: both
+        type: choice
+        options: [both, claude, codex]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: native-host-smoke-${{ github.workflow }}-${{ inputs.host }}
+  cancel-in-progress: true
+
+jobs:
+  native-host-smoke:
+    # This runs only on a trusted runner with existing CLI login state; no workflow secrets are passed.
+    runs-on: self-hosted
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - host: claude
+            source_tests: packages/agent-bundle/tests/native-claude-contract.test.ts packages/agent-bundle/tests/eval-claude-harness.test.ts
+            packed_command: pnpm test:packed:native:claude
+          - host: codex
+            source_tests: packages/agent-bundle/tests/native-codex-contract.test.ts packages/agent-bundle/tests/eval-codex-home.test.ts
+            packed_command: pnpm test:packed:native:codex
+    steps:
+      - uses: actions/checkout@v7
+        if: inputs.host == 'both' || inputs.host == matrix.host
+      - uses: pnpm/setup@v1
+        if: inputs.host == 'both' || inputs.host == matrix.host
+        with:
+          cache: true
+          install: false
+          runtime: node@22.19.0
+      - run: pnpm install --frozen-lockfile
+        if: inputs.host == 'both' || inputs.host == matrix.host
+      - name: ${{ matrix.host }} source signed-in smoke
+        if: inputs.host == 'both' || inputs.host == matrix.host
+        env:
+          AGENT_BUNDLE_NATIVE_CLAUDE_SMOKE: ${{ matrix.host == 'claude' && '1' || '' }}
+          AGENT_BUNDLE_NATIVE_CODEX_SMOKE: ${{ matrix.host == 'codex' && '1' || '' }}
+        run: pnpm test -- ${{ matrix.source_tests }}
+      - name: ${{ matrix.host }} installed-tarball native smoke
+        if: inputs.host == 'both' || inputs.host == matrix.host
+        run: ${{ matrix.packed_command }}

.github/workflows/package-preview.yml

@@ -0,0 +1,25 @@
+name: Package preview
+
+on:
+  pull_request:
+
+permissions: {}
+
+concurrency:
+  group: package-preview-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  publish:
+    name: Publish pkg.pr.new preview
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - uses: pnpm/setup@v1
+        with:
+          cache: true
+          install: false
+          runtime: node@22.19.0
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm build
+      - run: pnpm preview:publish

.superpowers/sdd/2026-08-14-agent-bundle-workbench-evals-implementation/task-9a-mcp-launch-config-report.md

@@ -0,0 +1,29 @@
+# Task 9A: MCP launch configuration report
+
+## Outcome
+
+- Added the Session-local `Launch configuration` display using only `McpBrowserSessionModel.config`.
+- `stdio` renders transport, command, ordered arguments, cwd, and lexical environment rows with explicit empty states.
+- `streamable-http` renders only its transport and sanitized URL.
+- The existing Inspector-config download and display now both derive from the same local `config` snapshot.
+- No server or model sanitizer changes were needed.
+
+## TDD evidence
+
+- RED: `npm test -- packages/workbench/tests/mcp-page.test.ts` exited 1 because the current page did not contain `Launch configuration`.
+- RED: `npm exec rstest -- --config rstest.config.ts run packages/workbench/tests/overview.e2e.test.ts --testNamePattern 'renders the safe launch configuration for one real artifact MCP session'` exited 1 at the missing `.mcp-page-launch-configuration` locator.
+- GREEN: the same dedicated Chrome command exited 0 (1 passed, 7 skipped). It opens a real artifact session, asserts safe values and secret absence, checks 390px overflow, and records page errors.
+
+## Verification
+
+- `npm test -- packages/workbench/tests/mcp-page.test.ts packages/workbench/tests/mcp-session-model.test.ts packages/workbench/tests/mcp-session-controller.test.ts` — 38 passed.
+- `npm run typecheck --workspace agent-bundle-workbench` — passed.
+- `npm exec rslint -- packages/workbench/src/mcp/mcp-page.tsx packages/workbench/tests/mcp-page.test.ts packages/workbench/tests/overview.e2e.test.ts` — 0 errors, 0 warnings.
+- `npm run build --workspace agent-bundle-workbench` — production Rsbuild passed.
+- `git diff --check` — passed.
+- Review regression: `git diff --unified=0 bc25a0c -- packages/workbench/tests/overview.e2e.test.ts` shows only the dedicated helper and test outside the broad lifecycle body.
+- TraceDecay file diagnostics — 0 diagnostics; unsafe-pattern and redundancy scans found no production findings.
+
+## Browser note
+
+The in-app Browser was unavailable, so the task-authorized Chrome fixture was used. Launch-configuration display, secret-absence, responsive-layout, and page-error coverage live solely in the dedicated Chrome test; the pre-existing broad lifecycle body was restored to its `bc25a0c` form. Its normal close-button click still has the independently reproduced baseline timeout, including when the Task 9A session subsection is temporarily removed.

.superpowers/sdd/2026-08-14-agent-bundle-workbench-evals-implementation/task-9b-mcp-session-timeout-report.md

@@ -0,0 +1,34 @@
+# Task 9B: MCP session timeout report
+
+## Delivered contract
+
+- `timeoutMs` is selected before session open, defaults to 5,000 ms, and is stored immutably on the session rather than in `McpSessionBinding`.
+- The MCP create route admits only a finite, positive own JSON number and returns the established `AB8016` invalid-shape diagnostic for invalid values or smuggled fields.
+- Initialization, catalog reads, prompt/resource/tool calls, and restart resolve omitted timeouts from the session value; internal callers can still supply an explicit request override.
+- Route, transport, controller, and page retain the value. The page exposes an accessible `Session timeout (ms)` control, validates locally, locks it once opening begins, and displays the exact active timeout.
+
+## TDD evidence
+
+- Backend RED: a create request with `timeoutMs` received HTTP 400; the opened session exposed no timeout.
+- Browser RED: transport POST body omitted `timeoutMs`, controller factory options omitted it, and page markup had no labeled control.
+- Chrome RED: dedicated interaction contract was written before the page control existed.
+- GREEN: all contracts pass with the implementation below.
+
+## Verification
+
+- `npm test -- packages/agent-bundle/tests/mcp-session-service.test.ts packages/agent-bundle/tests/mcp-session-routes.test.ts` — 34 passed.
+- `npm run typecheck --workspace agent-bundle-workbench` — passed.
+- Scoped `npx rslint` for the five MCP sources and focused tests — 0 errors, 0 warnings.
+- `npm test -- packages/workbench/tests/agent-bundle-remote-transport.test.ts packages/workbench/tests/mcp-session-controller.test.ts packages/workbench/tests/mcp-page.test.ts` — 41 passed.
+- `npx rstest --config rstest.config.ts packages/workbench/tests/mcp-session-timeout.e2e.test.ts --reporter verbose` — 1 Chrome test passed, including restart, invalid local input/no route request, no page errors, and 390 px overflow check.
+- `npm run build --workspace agent-bundle-workbench` and `git diff --check` — passed.
+- TraceDecay diagnostics reported 0 errors; simplify, redundancy, unsafe-pattern, and test-risk scans reported no findings in scope.
+
+## Commits
+
+1. `9f868fa feat(mcp): persist session request timeouts`
+2. `feat(workbench): control MCP session timeouts`
+
+## Scope guard
+
+No per-operation browser timeout, launch-config/download, trace-export, Inspector, Playground, eval, raw-log, or broad lifecycle E2E changes were made.

.superpowers/sdd/2026-08-14-agent-bundle-workbench-evals-implementation/task-9c-integrated-inspector-report.md

@@ -0,0 +1,35 @@
+# Task 9C: Integrated Inspector report
+
+## Delivered contract
+
+- The Workbench rail and page union no longer expose Inspector as a peer page.
+- MCP owns an internal, accessible Playground / Inspector tab control. Both tab panels stay mounted and are hidden and inert when inactive, so switching does not recreate either presentation.
+- `McpPage` and `InspectorSessionAdapter` share the existing controller; the adapter receives the app-level subscription of that same controller model. No second transport, session, bootstrap, route client, or iframe was introduced.
+- `#mcp` is the sole MCP route. Direct `#mcp` and the rail MCP action select Playground, and internal tabs leave browser history unchanged. Unsupported `#inspector` remains in the URL, resolves to Overview, renders no Inspector tab, and opens no MCP session.
+- The responsive tab shell keeps the 390 px page width within the viewport.
+
+## TDD evidence
+
+- RED: the dedicated-Inspector Chrome fixture was first rewritten to require canonical `#mcp`, no Inspector rail link, a selected internal Inspector tab, idle/no POST, the shared real-session flow, and a separate unsupported-`#inspector` Overview fallback with no session. The pre-change dedicated page retained `#inspector` and had no internal Inspector tab.
+- GREEN: the same real Chrome fixture now verifies the unsupported hash stays `#inspector` while showing Overview without a session, then exercises the canonical `#mcp` flow with one POST, exact shared Inspector catalogs/protocol/logging, preserved Playground form state, reset propagation, no page errors, and 390 px overflow.
+
+## Verification
+
+- `npm exec rstest -- run packages/workbench/tests/inspector-shell.e2e.test.ts --reporter verbose` — 3 Chrome tests passed: unsupported-hash fallback, production `#mcp` session flow, and explicit development artifact.
+- `npm exec rstest -- run packages/workbench/tests/inspector-session-adapter.test.ts packages/workbench/tests/inspector-session-adapter-fixture.test.ts packages/workbench/tests/mcp-session-controller.test.ts packages/workbench/tests/mcp-page.test.ts --reporter verbose` — 47 tests passed.
+- `npm run typecheck --workspace agent-bundle-workbench` — passed.
+- `npx rslint packages/workbench/src/main.tsx packages/workbench/tests/inspector-shell.e2e.test.ts` — 0 errors, 0 warnings.
+- `npm run build --workspace agent-bundle-workbench` and `git diff --check` — passed.
+- TraceDecay package diagnostics — 0 errors, 0 warnings; simplify scan — no findings (the index noted only `styles.css` stale); unsafe-pattern scan — 0 matches; test-risk scan — no items.
+- Code-simplifier and deslop review — no behavior-preserving cleanup needed.
+
+## Scope guard
+
+No Inspector vendor/adapter source, export wiring, MCP launch/timeout/App logic, backend routes, or evaluator files changed.
+
+## Fix round 1: complete tab semantics
+
+- RED: the expanded real Chrome session contract failed because the selected Playground tab had no `tabindex="0"`; ArrowRight therefore did not move focus or select Inspector.
+- GREEN: the two internal tabs now use roving `tabIndex` and select/focus through ArrowLeft, ArrowRight, Home, and End without mutating browser history.
+- The existing Chrome fixture now also proves exactly one mounted panel of each kind, inactive `hidden` + `inert` state, selected-panel exposure, stable history through keyboard and pointer switches, retained Inspector Logging state after a Playground roundtrip, and no horizontal overflow for either presentation at 390 px.
+- Verification: Inspector Chrome fixture 3/3; focused Inspector adapter/controller/page suite 47/47; Workbench typecheck, scoped Rslint, production Rsbuild, and `git diff --check` all passed. Code-simplifier/deslop found no cleanup needed.

.superpowers/sdd/2026-08-14-agent-bundle-workbench-evals-implementation/task-9d-current-protocol-trace-export-report.md

@@ -0,0 +1,53 @@
+# Task 9D report: current MCP protocol trace export
+
+## Scope guard
+
+Implemented only the browser-side current MCP protocol trace export. The export is deliberately distinct from Task 10's durable Playground session export: it has no persistence, append/finalize route, transport call, session POST, backend, hook, artifact, App, timeout, vendor, eval, or launch-config/environment/session-capability surface.
+
+The shared Task 9C controller/session remains the sole source of truth. The MCP page uses the existing generic Blob browser-download sink, and Inspector Protocol and Logging use their existing `onExportTrace` callback with the full underlying timeline rather than a filtered or cleared presentation list.
+
+Concurrent unrelated worktree edits (including Hooks, closure, config, and other MCP test changes) were preserved and are excluded from this commit.
+
+## RED
+
+Before production code, ran:
+
+```text
+npx rstest run packages/workbench/tests/mcp-page.test.ts packages/workbench/tests/inspector-session-adapter.test.ts
+```
+
+Expected RED: `packages/workbench/tests/mcp-page.test.ts` failed because `../src/mcp/mcp-protocol-trace.ts` did not exist. The new contracts covered the missing canonical builder/page trace action; the real Inspector Chrome flow was then extended to require the absent browser download wiring.
+
+## GREEN
+
+- Added `mcp-protocol-trace.ts`, a pure canonical, versionless builder with `kind: "agent-bundle.mcp-protocol-trace"`, explicit null session facts, full cursor/timeline/history, immediate detached JSON Blob serialization, trailing newline, JSON MIME type, and deterministic opaque-session/idle filename.
+- Added the labeled MCP Trace action and explanatory copy: it exports the current browser MCP trace, not a durable Playground session export.
+- Reused one generic browser Blob/download sink for Inspector config, MCP trace, and Inspector Protocol/Logging trace downloads.
+- Added unit coverage for exact export shape, raw ordered entries including replay gaps/invocations, cursor/history preservation, nulls, MIME/newline/filename, detached bytes after caller mutation, sensitive launch/config/session-capability omission, page sink handoff, and complete Inspector Protocol/Logging timeline handoff.
+- Extended the existing artifact-backed Inspector Chrome fixture to parse downloads from MCP Page, Inspector Protocol, and Inspector Logging; it proves matching canonical traces, the real `tools/call` payload, one session POST, no page errors, and 390 px no-overflow behavior.
+- Applied explicit `undefined` `useRef` initializers in the owned entry file to satisfy the current Workbench TypeScript version without changing runtime behavior.
+
+## Verification
+
+All post-change commands passed:
+
+```text
+npx rstest run packages/workbench/tests/mcp-page.test.ts packages/workbench/tests/inspector-session-adapter.test.ts packages/workbench/tests/mcp-session-controller.test.ts packages/workbench/tests/mcp-session-model.test.ts
+# 62 passed
+
+npx rstest run packages/workbench/tests/inspector-shell.e2e.test.ts
+# 3 passed: unsupported-hash, shared real-session, and development artifact coverage
+
+npm run typecheck --workspace agent-bundle-workbench
+npx rslint packages/workbench/src/mcp/mcp-protocol-trace.ts packages/workbench/src/mcp/mcp-page.tsx packages/workbench/src/main.tsx packages/workbench/tests/mcp-page.test.ts packages/workbench/tests/inspector-session-adapter.test.ts packages/workbench/tests/inspector-shell.e2e.test.ts
+npm run build --workspace agent-bundle-workbench
+git diff --check
+```
+
+TraceDecay MCP was unavailable. One required CLI fallback attempt failed exactly with:
+
+```text
+Error: config error: TraceDecay daemon socket '/home/zack/.tracedecay/daemon.sock' is not available. Run `tracedecay daemon install-service` and ensure the service is running.
+```
+
+Used local scoped source/test evidence after that single failed attempt. The requested code-simplifier and deslop reviews found no behavior-preserving cleanup to apply in the Task 9D hunks.

AGENTS.md

@@ -0,0 +1,16 @@
+# Repository guidance
+
+## Workbench platform scope
+
+- The developer Workbench is a desktop-only application.
+- Design, implementation, and browser acceptance should target desktop viewports and desktop interaction patterns.
+- Do not add mobile-specific layouts, responsive behavior, or mobile acceptance requirements unless the user explicitly requests them.
+- A mobile-only layout defect is not a release blocker for this repository.
+
+## Public examples
+
+- Treat `examples/*` as user-facing products, not test fixtures.
+- Use only public `agent-bundle` package exports and `workspace:*` dependencies.
+- Validate examples at a 1440×900 desktop viewport; mobile support is not required.
+- Never accept or capture a Workbench route while its loading state is still visible.
+- Browser acceptance must cover populated state plus the documented stale-diagnostic and repair flow.