@ggui-ai/protocol 0.2.0-alpha.4 → 0.4.0-rc.0

package/dist/version.d.ts

@@ -6,6 +6,778 @@
  * schema change; the most recent change anchors {@link PROTOCOL_VERSION}.
  *
  * --------------------------------------------------------------------
+ * Render failure envelope — in-result isError tool errors (2026-07-23,
+ * BREAKING, pre-launch). Ruling B. A `ggui_render` that fails or is
+ * rejected is a `tools/call` RESULT carrying `isError: true` — never a
+ * thrown JSON-RPC error. Before this, failures returned a
+ * SUCCESS-shaped envelope: an accident of the 2026-05-14 codeReady
+ * byte-trim, not a design, which mounted error renders as blank divs
+ * and let render meters count failures as billable.
+ *
+ *   rb1. Failures stay SCHEMA-CONFORMANT. MCP SDK clients validate
+ *        `structuredContent` against `outputSchema` even when
+ *        `isError` is set, so the failure envelope must satisfy
+ *        `renderOutputSchema` — it is not a free-form error shape.
+ *        SPEC §7.1 / §7.9 Plane 3.
+ *   rb2. `renderOutputSchema` gains optional `error: {code, message}`,
+ *        present iff `isError`. `code` is the closed enum
+ *        `RenderErrorCode` = PRODUCTION_FAILED | VALIDATION_ERROR |
+ *        NO_PLATFORM_KEY | NO_CREDENTIALS; `message` is agent-facing
+ *        guidance, not a stack.
+ *   rb3. `resourceUri` becomes OPTIONAL — present iff the render is
+ *        mountable. A failure carries no resource, so consumers must
+ *        stop assuming presence. `updateOutputSchema` was synced for
+ *        the same stripped-field class.
+ *   rb4. The phantom JSON-RPC error code -32004 (PRODUCTION_FAILED) is
+ *        RETIRED from `MCP_ERROR_CODES`. It was never emitted; the
+ *        in-result contract above replaces it.
+ *   rb5. NO `_meta` on failures (nothing to project), and the error
+ *        session still COMMITS — the handshakeId is consumed, so
+ *        recovery is re-handshake, not retry.
+ *
+ * Conformance-kit verdict: minor-intent while `draft-`, so
+ * PROTOCOL_VERSION is unchanged and no migration doc is required (see
+ * docs/protocol/VERSION-POLICY.md). Recorded here because the wire
+ * shape changed even though the stamp did not.
+ *
+ * --------------------------------------------------------------------
+ * Consolidated cleanliness release (2026-06-12, BREAKING, pre-launch).
+ * Four strands from the 2026-06-12 cleanliness + orphan audits, every
+ * finding adversarially re-verified before action: the dv* entries
+ * purge dead wire vocabulary and SSoT-wire the MCP tool input schemas;
+ * the ug* and dg* entries clean the `@ggui-ai/ui-gen` + `@ggui-ai/design`
+ * published artifacts (dead generation island, prompt-builder
+ * consolidation, packaging — no wire change); the tw* entries
+ * single-source twin/mirror surfaces across the react / react-native /
+ * server packages (no wire change).
+ *
+ *   dv1. **MCP tool input schemas are SSoT-wired.** `schemas/mcp.ts`'s
+ *      "the server handler imports these" docstring is now TRUE: the
+ *      handlers in `@ggui-ai/mcp-server-handlers` (+ the hosted pod's
+ *      `ggui_discover` / `ggui_request_credential`) import the new
+ *      `*InputShape` raw shapes (consume / emit / get-session /
+ *      list-featured / search / render-blueprint / discover /
+ *      request-credential / render) instead of hand-authoring mirrors.
+ *      The schemas reconciled TOWARD the shipped handler behavior (the
+ *      live wire): `sessionId` gains `.min(1)`, consume `timeout` is
+ *      `.int()`, search `limit` is int 1..100, render-blueprint loses
+ *      the never-read `props` field, list-featured + discover are
+ *      EMPTY shapes (the filter vocabulary died in the pre-launch
+ *      scrub), and the aspirational `.strict()` posture drops to match
+ *      the `z.object(shape)` strip-unknown-keys behavior every live
+ *      handler actually enforces. `renderInputSchema` — the one drift
+ *      case decided the other way, toward the truthful contract —
+ *      gains the live `themeId` + `infra.model` fields the published
+ *      schema was missing (it would have REJECTED valid live wire
+ *      calls), and the handler adopts the schema's richer handshakeId
+ *      validation messages. Hand types `GguiConsumeInput` /
+ *      `GguiGetSessionInput` / `GguiSearchBlueprintsInput` /
+ *      `GguiRenderBlueprintInput` / `GguiRequestCredentialInput` are
+ *      now `z.infer` of those schemas ("kept in sync by convention"
+ *      dies); `ggui_handshake` / `ggui_update` keep their documented
+ *      deliberate handler-side divergences (loose-on-purpose draft
+ *      contract; optional sessionId for in-process dispatch).
+ *
+ *   dv2. **`EventType` collapses to its one real member
+ *      `'data:submit'`** (actions-vs-context: actions drive turns;
+ *      there is no third category). The 8 producer-less members
+ *      (`data:change`, `lifecycle:focus`/`blur`,
+ *      `interaction:click`/`hover`/`scroll`, `error:validation`/
+ *      `connection`) are deleted along with their dead client lanes:
+ *      the RN WebViewRenderer's legacy bridge emissions and both SDKs'
+ *      never-firing `GguiRenderProps.onInteraction` callback. Server
+ *      runtime behavior is unchanged — non-submit wire envelopes were
+ *      already ledger-only and never reached `ggui_consume`.
+ *
+ *   dv3. **`'url'` / `'close'` / `'internal:progress'` leave the WS
+ *      message union** — pre-flatten vocabulary that survived the
+ *      rc4/rc7 frame purges. `'url'` + `UrlPayload`: zero emitters
+ *      since the post-R5 `/r/` short-code drop. `'close'` +
+ *      `ClosePayload`: SPEC §12.2 already records the close frame
+ *      deleted in draft-2026-05-27 ("no terminal close ceremony");
+ *      the server's consumer-with-no-emitter socket-router arm goes
+ *      with it — clients just close the socket. `'internal:progress'`
+ *      + `InternalProgressPayload` + `ProgressStep`: zero producers,
+ *      zero handler arms; rc7's "ProgressStep survives" retention was
+ *      circular (its only consumer was InternalProgressPayload).
+ *
+ *   dv4. **Emitterless union members deleted**: `GenerationStrategy`
+ *      loses `'creative'`; `ScreenMechanic` loses `'drag'` / `'swipe'`
+ *      (no authored blueprint ever used them; the generator infers
+ *      gesture mechanics from the prompt).
+ *
+ *   dv5. **Client pre-ack error classification matches the real
+ *      wire**: the iframe-runtime's `classifyPreAckError` /
+ *      `fromAuthFailure` / `ProtocolError`'s `auth` arm drop the
+ *      never-emitted `AUTH_REJECTED` / `TOKEN_EXPIRED` codes (`git
+ *      log -S` proves no first-party server ever minted them) in
+ *      favor of the codes servers actually send pre-ack:
+ *      `BOOTSTRAP_EXPIRED` / `BOOTSTRAP_INVALID` /
+ *      `BOOTSTRAP_SESSION_MISMATCH` / `BOOTSTRAP_APP_MISMATCH` /
+ *      `UNAUTHENTICATED` (+ the kept `SESSION_NOT_FOUND`). A real
+ *      ws-token expiry now classifies as `auth` instead of falling
+ *      into the generic `protocol` bucket. The conformance kit's
+ *      authored ProtocolError mirror follows. `AUTH_REJECTED` remains
+ *      ONLY on the §5.5.2 postMessage `BootstrapFailureReason` plane,
+ *      where SPEC documents it.
+ *
+ *   dv6. **Dead published exports pruned** (zero consumers + zero
+ *      docs — published-API ladder rung 3): the Streamable-Invoke
+ *      narrowing aliases (`TextDelta`, `InputJsonDelta`,
+ *      `ContentBlockDeltaPayload`, `MessageStartEvent`,
+ *      `ContentBlockStartEvent`, `ContentBlockDeltaEvent`,
+ *      `ContentBlockStopEvent`, `MessageDeltaEvent`,
+ *      `MessageStopEvent`, `PingEvent`, `ErrorEvent`, `InvokeRequest`
+ *      — the zod schemas in `schemas/invoke.ts` stay); the
+ *      model-catalog helpers (`PROVIDER_INFO`, `isValidModelId`,
+ *      `getModelsForProvider`, `getModelsForTier`, `selectModelByTier`
+ *      — `MODEL_REGISTRY` / `ModelId` / `DEFAULT_MODEL` stay);
+ *      `DESIGN_SYSTEM_VERSION` + the StyleFingerprint cluster
+ *      (`StyleFingerprintInput`, `computeStyleFingerprint`);
+ *      `DeclareToolCatalogInput` (the schema + Output type stay);
+ *      inference utilities `GadgetNames` / `InferContextNames` /
+ *      `InferContextValue` / `PropsOf`; and the featured-blueprints
+ *      wire-liars `GguiListFeaturedBlueprintsInput` / `...Output` +
+ *      `GguiDiscoverInput` / `GguiDiscoverOutput` (the live handlers'
+ *      shapes are the wire; the dead handler `category` filter branch
+ *      parsed from an empty schema dies too). Sibling-package prunes
+ *      in the same slice: `@ggui-ai/react` loses the explicit
+ *      back-compat alias `extractBootstrapMeta` (+ the unreachable
+ *      `SelfRepairBoundary` default export); `@ggui-ai/mcp-server`
+ *      loses dead auth fragments (`formatClearUserSessionCookieHeader`
+ *      with its false /logout docstrings, `OAuthAuthResult`);
+ *      `@ggui-ai/iframe-runtime` drops the dead `./runtime.js`
+ *      exports-map subpath.
+ *
+ *   dv7. **protocol src adopts its own `isRecord`** at ~20 guarded
+ *      `as Record<string, unknown>` cast sites (no wire change; one
+ *      deliberate tightening: `isGguiSubmitActionInput` now rejects
+ *      ARRAY payloads for unknown extension kinds — payload-object
+ *      presence is the documented invariant).
+ *
+ *   dv8. **The Component Bridge is now inbound-only** (SPEC §8
+ *      rewritten). The legacy user-data half — postMessage type
+ *      `'ggui-user-data'` (`BRIDGE_EVENTS.USER_DATA_POST`),
+ *      CustomEvent `'ggui:user-data'` (`BRIDGE_EVENTS.USER_DATA`),
+ *      lifecycle types `'ggui-resize'` / `'ggui-render-success'` /
+ *      `'ggui-render-error'` (`RESIZE` / `RENDER_SUCCESS` /
+ *      `RENDER_ERROR`), and the `SRCDOC_USER_DATA_BRIDGE` capture
+ *      snippet — is deleted end-to-end: zero injectors, zero
+ *      dispatchers, zero receivers existed (the consuming
+ *      IframeComponentRenderer died in the component-renderer
+ *      retirement; both SDKs' `GguiRender` listeners could never
+ *      fire). Outbound user interaction has ONE path: component
+ *      action hooks (`useAction`) → `data:submit` `ActionEnvelope`
+ *      over the live channel → consume buffer (§5, §7.3, §8.4).
+ *      `BRIDGE_EVENTS.AGENT_DATA_POST` / `AGENT_DATA` +
+ *      `SRCDOC_AGENT_DATA_BRIDGE` (the live agent→component half)
+ *      are unchanged.
+ *
+ *   dv9. **`ProgressUpdate` deleted** — the producer-internal
+ *      generation-progress type had exactly one consumer chain: the
+ *      never-implemented `UiGenerator.stream?()` /
+ *      `UiGenerateEvent` seam in `@ggui-ai/mcp-server-core`, which
+ *      dies in the same slice (the one consumer-position handler
+ *      documented that it deliberately ignores it; provisional
+ *      preview on `_ggui:preview` is the shipped
+ *      progress-visibility channel). Client-visible progress
+ *      vocabulary is unchanged.
+ *
+ *   dv10. **`'ggui:request-repair'` client event deleted** with its
+ *      whole dead seam: no listener has existed since the BaseShell
+ *      deletion (2026-04-23) — `GguiSessionRenderer`'s dispatch, the
+ *      never-supplied `onRequestRepair` option threading in
+ *      `@ggui-ai/react` (`ReactComponentRendererProps`) and
+ *      `@ggui-ai/iframe-runtime` (`RenderItemOptions` /
+ *      `ReactRootMountOptions`), and the false "render-item wires
+ *      this to the envelope it dispatches" docstring all go. The
+ *      `SelfRepairBoundary` tier (separate, honestly-`@experimental`
+ *      surface) is untouched.
+ *
+ *   dv11. **Inbound MCP-Apps hosting plumbing removed; the wire
+ *      variant stays.** The producer-less server island —
+ *      `CreateGguiServerOptions.connectors` (its forwarded value was
+ *      silently swallowed by a conditional spread into a handler
+ *      that never read it), the unmountable `/mcp-apps/resource` +
+ *      `/mcp-apps/tools-call` proxy module, `ConnectorRegistry` /
+ *      `InMemoryConnectorRegistry`, the never-set
+ *      `mcpAppsServerBaseUrl` renderer option, and the
+ *      hardcoded-false `isMcpAppsGguiSession` provisional-preview
+ *      gate input — is deleted (rebuild-from-git when a real
+ *      embedding customer materializes; the producing
+ *      `shortcuts.mcpApps` input died in the handshake-first
+ *      collapse). `McpAppsGguiSession` / `McpAppsSource` and every
+ *      union-forced client/handler guard branch REMAIN published
+ *      wire surface per SPEC §6 + the MCP-Apps composition strategy.
+ *
+ *   dv12. **`ggui.json#theme` (ThemeDocument) drops the optional
+ *      `canvas` group**, and `DtcgTheme` / `ParsedTheme` drop
+ *      `canvas` / `canvasConfig` + the `--ggui-canvas-*` CSS
+ *      variables (design-package + open file-format vocabulary, not
+ *      MCP wire). The consuming GenerativeCanvas shells left with
+ *      the guuey split; no renderer, prompt, or app shell read the
+ *      tokens since. Operator theme files carrying a `canvas` block
+ *      now fail strict parse — delete the block.
+ *
+ *   dv13. **`checkSecurity` + `SecurityCheckResult` deleted;
+ *      `FULLSTACK_IMPORT_PREFIXES` un-exported** (published npm
+ *      surface, no wire change — DEAD-PROTO-04 follow-up). The
+ *      convenience checker had zero consumers repo-wide: every live
+ *      validation path runs `validateComponentDetailed` in
+ *      `@ggui-ai/ui-gen`, which imports `DANGEROUS_PATTERNS` directly
+ *      and emits richer per-violation diagnostics. The prefix list
+ *      stays as the internal input of the still-published
+ *      `classifyUi`. `DANGEROUS_PATTERNS` / `DangerousPattern` /
+ *      `UiClass` / `classifyUi` are unchanged.
+ *
+ *   ug1. **`@ggui-ai/ui-gen` superseded coding-agent generation
+ *      island deleted** (no wire change — published npm surface +
+ *      dist bytes). The pre-harness virtual-git orchestration
+ *      (`coding-agent/{index,agent,planner,file-agent,boilerplate,
+ *      prompts,trace,context}.ts`, `runCodingAgent` / `runPlanner` /
+ *      `runFileAgent`), the legacy `apply_diff` unified-diff tool
+ *      case + `diff-processor.ts` (reachable only from the island),
+ *      the island-only `fullToolSchemas` / `initialToolSchemas` /
+ *      `executeToolBatch` registries, the text-only
+ *      `runEnforcedCodingLoop` / `buildFeedback` loop, and the
+ *      legacy prompt constants (`LEAN_PROMPT`, `CODER_PROMPT`,
+ *      `ENFORCED_CODER_PROMPT`, `SELF_CHECK_RULES`,
+ *      `buildEnforcedCoderPrompt`, `injectDesignSpec`,
+ *      `injectFeedback`, `CONTEXT_TOOL_NAMES`, `BUILD_TOOL_NAMES`)
+ *      are gone — all superseded by `harness/coding/run-coding-turn`
+ *      + `boilerplate/system-prompt`. The two stale tsup entries
+ *      that kept shipping unroutable `dist/coding-agent/*` bytes
+ *      after the rc13 subpath prune die with it. KEPT (live via
+ *      `harness/coding/*`): `workspace.ts`, `tools.ts`, `types.ts`
+ *      (pruned to its consumed members), `self-check.ts`,
+ *      `tag-balance.ts`.
+ *
+ *   ug2. **`@ggui-ai/ui-gen/harness/prompts` re-exports the
+ *      canonical prompt-context builders** from `contract-context.ts`
+ *      (the module the live `createUiGenerator` path uses) instead of
+ *      carrying a hand-maintained twin that had silently drifted (no
+ *      variance block, no stream `source.tool` annotation, hooks-only
+ *      gadget teaching). One module now owns the LLM-visible strings,
+ *      and the duplicate mutable `RenderingContext` in
+ *      `harness/result-types` collapses to a re-export of the
+ *      canonical readonly type. The LIVE generation path's emitted
+ *      prompt is byte-identical (snapshot-proven); the multi-sdk
+ *      benchmark's deep-import path now exercises production prompt
+ *      content instead of the drifted fork.
+ *
+ *   ug3. **`runTier0Checks` / `runTier0` (`@ggui-ai/ui-gen/check`)
+ *      drop the never-read `compiledCode` positional parameter.**
+ *      Compile failure has been signaled by `buildErrors` alone since
+ *      the S6 tier-0 migration; the documented-but-ignored parameter
+ *      was an unflagged contract gap on a published signature.
+ *
+ *   ug4. **`provider-adapter-mock` leaves the published artifact**
+ *      — relocated to `provider-adapter-mock.test-util.ts` (excluded
+ *      from the tarball), its stale tsup entry deleted. rc13 already
+ *      killed the `./provider-adapter-mock` subpath; the dist bytes
+ *      had been shipping with no exports route to reach them.
+ *
+ *   ug5. **`@ggui-ai/ui-gen` stops bundling provider SDKs**
+ *      (packaging only): `@google/adk`, `@openai/agents`,
+ *      `puppeteer-core` move devDependencies → optional
+ *      peerDependencies + tsup externals. They are lazy
+ *      `await import(...)`s in the Google/OpenAI adapters + visual
+ *      evaluator, but as devDeps tsup inlined their full transitive
+ *      trees into every entry chunk. Tarball: 84.7MB → 16.1MB
+ *      unpacked (17.7MB → 4.0MB packed). Consumers using those
+ *      adapters install the SDK themselves.
+ *
+ *   ug6. **`./advanced` subpath actually builds** (packaging only):
+ *      `src/advanced/index.ts` was never a tsup entry — the documented
+ *      `@ggui-ai/ui-gen/advanced` exports-map target only resolved
+ *      because a stale pre-tsup `dist/advanced/` orphan kept shipping,
+ *      and the clean-build hygiene in this release exposed the dangling
+ *      subpath. The entry is now registered (the DTS graph growth needs
+ *      an 8GB heap cap on the build script).
+ *
+ *   dg1. **`@ggui-ai/design` stops shipping Storybook stories**
+ *      (packaging only): `files[]` negations + `tsconfig.build.json`
+ *      excludes drop the 37 `src/*.stories.tsx` + their compiled dist
+ *      output (209 pack entries, incl. the orphaned
+ *      `dist/blueprints` / `dist/templates` story remnants of deleted
+ *      src dirs); pack list 891 → 601 files.
+ *
+ *   tw1. **`ai.ggui/render` slice deps single-sourced** (published
+ *      npm surface, no wire change). `@ggui-ai/mcp-server-handlers`
+ *      exports `RenderSliceMetaDeps` + `assembleRenderSliceBase`;
+ *      `GguiRenderHandlerDeps` and `GguiUpdateHandlerDeps` now
+ *      `extend` the shared interface instead of hand-mirroring
+ *      `mintWsToken` / `runtimeUrl` / `themeId` / `themeMode` /
+ *      `themeProvider`, and both handlers assemble the envelope base
+ *      (runtimeUrl fallback, `token`→`wsToken` remap, layered theme
+ *      resolution) through the one helper. DELETED with it: update's
+ *      vestigial `streamWebSocketLocalTools` dep (declared + plumbed
+ *      through `@ggui-ai/mcp-server`'s `UpdateDeps`, never read — the
+ *      post-update slice is a deliberate props-only subset that never
+ *      emitted the field; it remains mount-time bootstrap data owned
+ *      by `ggui_render`).
+ *
+ *   tw2. **`@ggui-ai/design` `ClickableProps.onClick` /
+ *      `PressableProps.onPress` widen to
+ *      `MouseEvent | KeyboardEvent`** (published type surface; runtime
+ *      unchanged). Enter/Space keyboard activation always delivered a
+ *      `KeyboardEvent` through a handler typed `MouseEvent` via a
+ *      double cast — the union makes the published type honest and
+ *      deletes both casts. Handlers annotating the event parameter as
+ *      bare `MouseEvent` must widen (or drop) the annotation.
+ *
+ *   tw3. **react ↔ react-native twin-parity gate extended** (test
+ *      infrastructure; both SDK suites). The websocket/provider
+ *      family (`WebSocketManager`, `EventBuffer`, `useWebSocket`,
+ *      `GguiProvider` + the two websocket test files) joins a new
+ *      `DOCUMENTED_DELTA_TWINS` structural gate (both copies exist,
+ *      RN file-top "Platform delta" header required, exported
+ *      surfaces equal modulo annotated platform-only exports), and
+ *      the reserved-validators A2UI adapter trio
+ *      (react / react-native / mcp-server) is pinned by a new
+ *      comment-stripped `CODE_IDENTICAL_MIRRORS` gate. One behavioral
+ *      convergence: the RN `useWebSocket.sendAction` now re-stamps
+ *      `schemaVersion` via `makeActionEnvelope` exactly like the web
+ *      hook (RN's docstring already claimed symmetry); RN gains the
+ *      formerly web-only `useWebSocket` test suite.
+ *
+ * --------------------------------------------------------------------
+ * Blueprint provenance (`BlueprintSource` union + `PortableBlueprint`
+ * v2) + retired-surface cleanup (2026-06-11, BREAKING, pre-launch).
+ * One release, four strands: the bp* entries introduce provenance on
+ * the wire; the rc* entries delete vocabulary that had no emitter;
+ * the ab* entries close the action→consume loop on the WS transport
+ * and collapse the duplicated client wire-config pipeline; the hc*
+ * entries clean the helper-package published surfaces (dead exports,
+ * packaging, conformance execution, component re-homing — no wire
+ * change) plus the one additive ops-tool schema change (hc4) and one
+ * published-symbol rename (hc5); the cc1 entry restores the SPEC §7.3
+ * `ggui_consume` timeout bound on the wire; the pa1 entry de-vendors
+ * the CLI publish-auth surface (no wire change).
+ *
+ *   bp1. **`BlueprintSource` discriminated union introduced**
+ *      (`kind: 'llm' | 'user' | 'curated'`), replacing the flat
+ *      provenance strings that previously lived on three separate
+ *      seams (catalog entries, screen blueprints, the runtime cache
+ *      registry). The `llm` arm REQUIRES `generator` + `model`; the
+ *      legacy `heuristic` arm is deleted (zero mint sites repo-wide).
+ *      `parseBlueprintSource` / `isBlueprintSource` validate at trust
+ *      boundaries — legacy flat strings are rejected, never coerced.
+ *
+ *   bp2. **`PortableBlueprint` v2**
+ *      (`PORTABLE_BLUEPRINT_SCHEMA_VERSION = 2`): `source` and
+ *      `generatorProtocolVersion` are REQUIRED (the latter's
+ *      optional-for-back-compat shim dies). `fromPortableBlueprint`
+ *      is the validating importer: v1 artifacts and era-mismatched
+ *      records are REJECTED — a rejected seed entry is just a
+ *      cold-gen; the fix is to re-export the pool, never to coerce.
+ *      `toolIdentityCatalogHash` stays optional (an offline pool
+ *      export has no runtime handshake catalog to hash).
+ *
+ *   bp3. **Provenance is stamped at every mint site** and persisted
+ *      on stored blueprint rows. Rows whose `source` does not parse
+ *      are dropped at the load boundary, never coerced — blueprints
+ *      are a cache (invalidation = regeneration, never data loss).
+ *
+ *   bp4. **MVB `Blueprint.generator` → `Blueprint.source`** — the
+ *      multi-variant blueprint record (and `blueprintSchema`) joins
+ *      the single provenance vocabulary; the bare engine slug dies.
+ *      `createdBy` is KEPT — a different axis (who initiated the
+ *      mint): an `ggui_ops_generate_blueprint` row is
+ *      `createdBy: 'operator'` AND `source.kind: 'llm'`. Ripples:
+ *      `BlueprintMeta.generator: string` (required) →
+ *      `BlueprintMeta.source?: BlueprintSource`, present iff
+ *      `origin === 'cache'` (the matched row's stored provenance);
+ *      absent on `agent`/`synth` — gen pending, nothing truthful to
+ *      report (the old field stamped a static deployment slug even
+ *      on cache hits, itself a fabrication). The
+ *      `HandshakeDecisionAdapter.generatorSlug` seam retires with
+ *      it. `ggui_ops_generate_blueprint` output `generator` →
+ *      `source` (llm arm, from the engine's own metadata stamp).
+ *      `ggui_ops_register_blueprint` input `generator` DELETED and
+ *      output `generator` → `source` (always `{kind: 'user'}`) —
+ *      the handler no longer fabricates a registry-default slug for
+ *      hand-authored bytes. `ggui_ops_list_blueprints` keeps the
+ *      `generator` filter, redefined to match `llm`-sourced rows
+ *      only. New Zod mirrors exported: `blueprintSourceSchema` +
+ *      per-arm `llm`/`user`/`curated` schemas.
+ *
+ *   rc1. **`SubscribePayload.appId` required → optional (widening).**
+ *      Absent appId no longer flows through raw: the server resolves
+ *      the caller's identity-default app before any store work —
+ *      token-bound appId, else the deployment's identity mapping
+ *      (`defaultAppIdFromIdentity`, same rule as the MCP route), else
+ *      the deployment default. A PRESENT appId must still match the
+ *      GguiSession's bound app or subscribe fails `APP_MISMATCH`.
+ *      Closes the absent-appId bifurcation (silent provisioning of a
+ *      GguiSession row with `appId: undefined` on in-memory stores).
+ *
+ *   rc2. **`CanvasLifecyclePayload` → `GguiLifecyclePayload`** (and
+ *      `validateCanvasLifecyclePayload` → `validateGguiLifecyclePayload`).
+ *      Internal type rename only — the `_ggui:lifecycle` reserved
+ *      channel and its event vocabulary are unchanged on the wire.
+ *      Canvas MODE is discarded (MCP Apps display modes are the
+ *      model); the lifecycle channel is generation-progress, not
+ *      canvas.
+ *
+ *   rc3. **`unknown_generator` leaves `GadgetGateErrorCode`.** The
+ *      throwing gate had no production call site (render's strict
+ *      override path carries no generator field); the live surface is
+ *      the forgiving `GENERATOR_UNKNOWN` handshake draft finding.
+ *      Render-time generator selection is deferred to v2.
+ *
+ *   rc4. **`'feedback'` leaves the WS message union** (both
+ *      `WebSocketMessageType` and the `WebSocketMessage` variant,
+ *      which carried a bare `JsonObject` payload). The frame was
+ *      vocabulary with no emitter in either direction: no client ever
+ *      constructed it, and the only server touch rejected it
+ *      `NOT_IMPLEMENTED`. The iframe-runtime's inbound no-op stub
+ *      handler (a shim for servers that never existed) is deleted
+ *      with it. Conformance-kit-invisible (the kit never covered the
+ *      frame). If love/dislike feedback ships later it re-enters with
+ *      a real producer + consumer and a typed payload.
+ *
+ *   rc5. **`SubscribePayload.bootstrap` dead duplicate deleted.** The
+ *      R4 rename (r4.1 below) moved the live-channel auth credential
+ *      to `wsToken`, but the retired slot survived on the published
+ *      wire type as a "generic transport-bootstrap" judgment call with
+ *      ZERO emitters or readers in either direction. The field, its
+ *      boundary-test pin, and every docstring still teaching it as
+ *      live are deleted; `wsToken` is the only subscribe credential.
+ *
+ *   rc6. **The `_ggui:contract-error` surface is deleted entirely.**
+ *      The channel was normative vocabulary with ZERO emitters
+ *      anywhere (SPEC carried an explicit "no first-party emitter"
+ *      gap since draft-2026-06-10); pre-launch, vocabulary nothing
+ *      produces is dead surface. Gone: `CONTRACT_ERROR_CHANNEL`, the
+ *      `ContractErrorPayload` type, the `ContractErrorCode` union,
+ *      `makeContractErrorPayload`, `validateContractErrorPayload` +
+ *      its `BUILTIN_RESERVED_VALIDATORS` entry, the iframe-runtime's
+ *      `fromContractErrorPayload` parser + the `ProtocolError`
+ *      `'contract'` variant, the conformance kit's authored
+ *      `ContractErrorCode` mirror, and SPEC §4.4's payload section +
+ *      §5.5.3's reserved-MAY dual-emission path. The reserved
+ *      `_ggui:` NAMESPACE policy survives unchanged (agents and
+ *      renderers MUST NOT author `_ggui:*` deliveries — SPEC §4.4 is
+ *      now that authority section); `_ggui:lifecycle` remains the
+ *      protocol-owned built-in-validated reserved channel. Codes the
+ *      union once named keep living on their REAL surfaces: WS error
+ *      frames (`SESSION_NOT_FOUND`, `AUTH_REJECTED`,
+ *      `CONTRACT_VIOLATION`), tool `structuredContent` rejections
+ *      (`INVALID_ACTION_KIND`, `PIPE_NOT_FOUND`, `CONTEXT_TOO_LARGE`),
+ *      and push-time schema-compat rejection (`SCHEMA_MISMATCH_ERROR`).
+ *      `sanitizeCausedBy` survives — its consumer is the
+ *      `channel_error` frame's `details` diagnostic, not this payload.
+ *
+ *   rc7. **`'stream'` / `'progress'` / `'agent-msg'` leave the WS
+ *      message union** (both `WebSocketMessageType` and the
+ *      `WebSocketMessage` variants). Frame census of every
+ *      first-party server send site: only `data` / `ack` / `render` /
+ *      `props_update` / `pong` / `error` / `drain_ack` / `channel_*`
+ *      are ever emitted — the three frames were pure vocabulary, and
+ *      client-visible generation progress already flows as
+ *      `{type:'data'}` envelopes on the reserved `_ggui:lifecycle`
+ *      channel (SPEC §4.5). Deleted with the members: the
+ *      `StreamPayload` / `ProgressPayload` / `AgentMsgPayload` /
+ *      `AgentMsgType` payload types (`ProgressStep` survived this
+ *      purge as `InternalProgressPayload`'s vocabulary — circular;
+ *      dv3 above deletes both), the
+ *      `BRIDGE_EVENTS.AGENT_STREAM` / `AGENT_MSG` / `AGENT_LOGS`
+ *      CustomEvent names + every dispatch site, the SDKs'
+ *      `GguiRenderProps.onStream` / `onProgress` dead callbacks (and
+ *      the client-fabricated synthetic matchType "progress" event),
+ *      and the RN SDK's listener-with-no-emitter hook
+ *      `useGenerationProgress`. SPEC §2.8 (NDA) + §8.2 now teach the
+ *      lifecycle channel as the one progress mechanism. The
+ *      conformance kit's stream-update matcher behavior is unchanged
+ *      (it only ever accepted `{type:'data'}`); its docs/tests now
+ *      pin the text-chunk frame as retired rather than "a different
+ *      wire type".
+ *
+ *   rc8. **Client-side data-binding vocabulary leaves the protocol**
+ *      (`types/data-bindings.ts`: `ClientToolName`, `ClientToolConfig`,
+ *      `DataBindings`, `TypedToolConfig` + the per-tool config types).
+ *      No wire envelope ever carried these; the only consumers were
+ *      the two SDKs' orphaned client-side tools subsystems
+ *      (registry / hooks / resolver / built-ins, the `/query` TanStack
+ *      integration, and the `/testing` mock-tools helpers), deleted in
+ *      the same slice (`@ggui-ai/react` + `@ggui-ai/react-native` lose
+ *      the `tools`, `/query`, and `/testing` surfaces and the dead
+ *      `queryClient` / `sessionId` provider plumbing). A future
+ *      native-controller story reintroduces a resolver WITH a real
+ *      producer, per the protocol-and-contract bar.
+ *
+ *   rc9. **MCP Apps thin shell emits the canonical
+ *      `MALFORMED_BOOTSTRAP` reason.** The `ui://ggui/render` shell's
+ *      `ggui:bootstrap-failed` postMessage carried the out-of-
+ *      vocabulary `BOOTSTRAP_MALFORMED` (transposed words) — outside
+ *      the SPEC §5.5.2 `BootstrapFailureReason` union, so spec-
+ *      conformant hosts switch-casing on `reason` misclassified the
+ *      known failure as unknown. The emit and its docstring now use
+ *      the canonical member. (The CSP script hash is derived from the
+ *      shell body at module load, so it tracks automatically.)
+ *
+ *   rc10. **`EVENT_NOT_ALLOWED` WS error code deleted** along with the
+ *      no-op `assertEventAllowed` stub and the never-thrown
+ *      `EventNotAllowedError` class in `@ggui-ai/mcp-server`. The
+ *      pre-Phase-B `subscription.events` allowlist lost its wire
+ *      shape in the session-stack collapse; the surviving ceremony
+ *      was a defined failure mode with no reachable emitter. Inbound
+ *      actions are gated by `assertActionContract` alone; docstrings
+ *      across protocol / handlers / SDKs now say so. Per-render event
+ *      policy, if wanted, re-enters as a designed slice with a real
+ *      wire shape.
+ *
+ *   rc11. **`BOOTSTRAP_EXPIRED` guidance message names the live
+ *      refresh tool.** The live channel's expired-credential error
+ *      message (and five server docstrings + the cloud e2e G14 spec)
+ *      still instructed callers to invoke the retired
+ *      `ggui_runtime_refresh_bootstrap` name (renamed in r4.1). All
+ *      now name `ggui_runtime_refresh_ws_token`; a grep-based
+ *      regression test in `@ggui-ai/mcp-server` locks the retired
+ *      name out of the package source.
+ *
+ *   rc12. **`UI_GEN_VERSION` deleted from `@ggui-ai/ui-gen`.** The
+ *      published constant claimed to be "kept in sync with
+ *      `package.json`" while shipping a stale value, and had zero
+ *      consumers repo-wide — a published-API lie with no reader. No
+ *      replacement; consumers that need the package version read it
+ *      from the package manifest.
+ *
+ *   rc13. **`@ggui-ai/ui-gen` export map pruned 55 → 37 subpaths.**
+ *      Fifteen never-imported subpath entries deleted
+ *      (`./evaluation/mcp-server`, `./evaluation/prompts`,
+ *      `./adapters/provider-router`, `./adapters/{claude,openai,
+ *      google}/raw`, `./coding-agent`, `./coding-agent/boilerplate`,
+ *      `./tools`, `./provider-adapter-mock`, `./compile`,
+ *      `./validation/ui-compiler`, `./design-system-docs`,
+ *      `./tools/get-wire`, `./tools/render-check`) plus the three
+ *      redundant `/index` duplicate aliases (`./harness/index`,
+ *      `./adapters/index`, `./evaluation/axis-checks/index` — the six
+ *      monorepo import sites retargeted to the bare forms). The
+ *      underlying modules stay (internal relative imports); only the
+ *      dead published surface dies. `./blueprint-hint`, `./strategy`,
+ *      `./user-request` were initially KEPT pending the
+ *      predefined-match pipeline existence decision; that decision
+ *      landed within this same release (wave-2 approved forks): the
+ *      predefined-match pipeline was deleted and all three subpaths
+ *      were removed with it.
+ *
+ *   ab1. **WS `data:submit` action ingress now feeds the consume
+ *      pipe.** The live channel's `handleInboundAction` dual-writes
+ *      every accepted `data:submit` envelope: the retained
+ *      `user.submitted` ledger event (ack `payload.sequence` source —
+ *      unchanged) AND a canonical consume entry
+ *      (`{intent, actionData, uiContext, actionId, firedAt}`, the
+ *      SAME shape `ggui_runtime_submit_action`'s dispatch branch
+ *      appends) on the sessionId-keyed pending-events pipe — so
+ *      `ggui_consume` drains WS-originated gestures identically to
+ *      tools/call-relayed ones. `uiContext` is `{}` on this transport
+ *      (WS clients mirror no contextSpec snapshot); `actionId` is a
+ *      server-minted 8-hex correlation id; `firedAt` is server clock.
+ *      New optional `GguiSessionChannelOptions.pendingEventConsumer`
+ *      seam (`@ggui-ai/mcp-server`); `createGguiServer` threads its
+ *      shared pipe instance whenever it composed the default handler
+ *      set. Pipe-append failure degrades to ledger-only with a
+ *      `render_channel_consume_append_failed` warn — ack semantics
+ *      unchanged. This makes the long-documented SDK promise ("the
+ *      agent receives actions via `ggui_consume`") TRUE on the WS
+ *      transport; the `tool` hint is clarified as LEDGER-only
+ *      vocabulary (operator surfaces) — consume entries carry no tool
+ *      slot on either transport (SPEC §nextStep routing + walkthrough
+ *      updated to the canonical drained shape).
+ *
+ *   ab2. **Client wire-config pipeline unified in `@ggui-ai/wire`.**
+ *      New exports `buildWireConfig` (the ONE envelope-build →
+ *      validate → emit dispatch pipeline + bus-backed subscribe) and
+ *      `StreamBus` (+ `RESERVED_CHANNEL_REPLAY_MAX`), hoisted from
+ *      `@ggui-ai/iframe-runtime`. `buildRootWireConfig` becomes an
+ *      iframe-side adapter injecting its CSP-precompiled validator,
+ *      ProtocolError dual-emission, and the tools/call-vs-WS
+ *      transport seam; `@ggui-ai/react`'s `<GguiRender>` builds its
+ *      `WireConfig` through the same shared pipeline (transport = WS
+ *      `action` frame, violations → `onError`, `clientSeq` shared
+ *      with `api.action`). `<GguiRender>` inbound `data` frames now
+ *      land on a per-render `StreamBus` with the bounded reserved-
+ *      channel replay ring — `_ggui:preview` frames arriving between
+ *      the subscribe ack and the provisional renderer's mount replay
+ *      to the late subscriber instead of dropping (behavior now
+ *      identical across both first-party renderers, per MCP Apps
+ *      Compliance). `useChannelStream` subscribes via the new
+ *      internal `StreamBusContext` instead of `window` CustomEvents;
+ *      the `BRIDGE_EVENTS.AGENT_DATA` window dispatch survives as a
+ *      one-way host-shell broadcast. React-Native twins unchanged —
+ *      porting the shared builder + bus to RN is the pending RN
+ *      WireProvider slice.
+ *
+ *   hc1. **`@ggui-ai/mcp-server-handlers` dead public surface deleted
+ *      (no wire change).** `UpdateUnsupportedError` (error class with
+ *      zero throw sites; its `update_unsupported` code appeared in no
+ *      spec or transport mapping); the empty `./contract-tests`
+ *      subpath (an `export {}` module promising a
+ *      `runRenderHandlerContract` battery that never landed) + the
+ *      vestigial `peerDependenciesMeta.vitest`; the retired
+ *      intent-keyed cache tail (`generationCacheKey`,
+ *      `DEFAULT_CACHE_SIMILARITY_THRESHOLD`,
+ *      `GenerationCacheDeps.similarityThreshold` — zero readers). The
+ *      live-LLM dev probe no longer compiles into the published dist
+ *      (tsconfig.build exclude).
+ *
+ *   hc2. **`@ggui-ai/mcp-server-core` contract-tests become executed
+ *      truth.** `gguiSessionStoreContract` +
+ *      `runGguiSessionStoreConformance` now RUN against both shipped
+ *      stores (in-memory + sqlite, temp-file db) from the stores' own
+ *      test files; `vitest` is declared as an optional peer dependency
+ *      (the `./contract-tests` subpath imports it explicitly — the
+ *      prior Jest-globals claim was false). Sqlite row reads narrow
+ *      through validating guards instead of `as Row` casts (operator-
+ *      mutable file = trust boundary); phantom impl names
+ *      (PostgresGguiSessionStore, S3VectorsStore, FileVectorStore)
+ *      leave the seam docstrings; ws-token prose names the real
+ *      `wsToken` slice field and the G14 multi-use-within-TTL posture.
+ *
+ *   hc3. **`startSandboxProxyServer` re-homed to
+ *      `@ggui-ai/agent-server`** (from `@ggui-ai/dev-stack`; module +
+ *      test suite moved verbatim, exported from the agent-server
+ *      barrel). agent-server drops its dev-stack dependency, so a
+ *      published agent install no longer eagerly pulls the dev-stack
+ *      toolchain (esbuild, chokidar). `@ggui-ai/dev-stack` also drops
+ *      its zero-consumer discovery re-export block — the canonical
+ *      home stays `@ggui-ai/project-config/node`. No compat
+ *      re-exports (pre-launch).
+ *
+ *   hc4. **`ggui_ops_create_app` output schema gains optional
+ *      `connectUrl`** (per-app MCP connect URL; emitted only by
+ *      deployments exposing per-app ingress — the default OSS
+ *      composition has no ingress knowledge and never sets it). The
+ *      cloud pod's variant converges onto the full OSS output shape
+ *      (`{appId, displayName, systemPrompt?, createdAt, updatedAt,
+ *      connectUrl?}`), and the pod tool-registry parity gate now pins
+ *      OUTPUT-schema identity (alongside input) for every shared wire
+ *      name, so host output drift is structurally caught. Additive —
+ *      existing callers unaffected.
+ *
+ *   hc5. **`@ggui-ai/negotiator` `validateContractStructure` →
+ *      `validateContractRedundancy`** (TS-symbol rename, no wire
+ *      change). The name collided with `@ggui-ai/protocol`'s normative
+ *      structural validator of the same name while carrying different
+ *      semantics (advisory warn-only redundant-action heuristic vs
+ *      `ContractViolation` list). Protocol remains the sole owner of
+ *      the `validateContractStructure` name; the negotiator export now
+ *      says what it detects. No compat alias (pre-launch).
+ *
+ *   cc1. **`ggui_consume.timeout` bound restored to SPEC §7.3
+ *      `[0, 25]`.** The shipped handler had drifted to a schema max of
+ *      600s with a silent server-side truncation to a 120s default cap
+ *      (and a `maxTimeoutSeconds` deployment override). All three die:
+ *      the input schema now enforces integer `[0, 25]`, out-of-range
+ *      values reject `INVALID_PARAMS` (-32602) at the tool-argument
+ *      validation layer instead of being silently truncated, and the
+ *      override knob is deleted (zero consumers; the bound is
+ *      normative). SPEC §7.3's output section is simultaneously
+ *      rewritten to the REAL drained-row shape — `ConsumeEventEntry`
+ *      `{type:'action', sessionId, intent, actionData, uiContext,
+ *      actionId, firedAt}` (+ the optional `client.hostContext` echo)
+ *      — replacing the stale `{events: ActionEnvelope[]}` claim; the
+ *      dedup guidance moves from `(sessionId, clientSeq)` (an ingress
+ *      key) to `(sessionId, actionId)`. Agents that long-polled with
+ *      `timeout > 25` MUST re-call `ggui_consume` in a loop instead.
+ *
+ *   pa1. **CLI publish-auth de-vendored (no wire change).** The
+ *      `ggui {gadget,blueprint} publish` / `ggui keys register`
+ *      default auth path no longer talks to an identity provider from
+ *      the CLI: the direct-IdP `USER_PASSWORD_AUTH` + refresh flows,
+ *      the `@aws-sdk/client-cognito-identity-provider` dependency (the
+ *      only AWS SDK in any published `@ggui-ai/*` package), the
+ *      `GGUI_REGISTRY_COGNITO_POOL_ID` /
+ *      `GGUI_REGISTRY_COGNITO_APP_CLIENT_ID` env surface, and the
+ *      per-registry `~/.ggui/auth/<host>/token.json` cache are all
+ *      deleted. The default credential is the stored `ggui login`
+ *      session (`~/.ggui/auth.json`), refreshed via
+ *      `POST /v1/auth/refresh`; `--auth=bearer` is unchanged. Auth
+ *      errors are rewritten vendor-neutrally, and an HTTP 401 from
+ *      the registry surfaces as `auth_failed` with a credential-
+ *      kind-specific diagnosis. (The hosted-registry server side
+ *      landed 2026-06-11: the write routes' gateway authorizer
+ *      validates `cli_at_*` session bearers, so the default auth
+ *      path works end-to-end against `registry.ggui.ai`.)
+ *
+ * --------------------------------------------------------------------
+ * Synchronous wired-action dispatch retired — agent-routed consume
+ * loop is the single action model (2026-06-10, BREAKING, pre-launch):
+ *
+ *   wa1. **`StreamChannelEntry.tool` (refresh hint) deleted.** The
+ *      per-channel "refresh this channel from tool X after an action"
+ *      declaration is gone from the type and from
+ *      `streamChannelEntrySchema`. The entry schema is `.strict()`,
+ *      so a contract carrying a stray `tool` key now fails
+ *      `CTR_SHAPE_UNRECOGNIZED_KEYS` at lint time. The channel data
+ *      feed `StreamChannelEntry.source` (`{tool, args?}`) is a
+ *      DIFFERENT axis and survives unchanged.
+ *      `ResolvedStreamChannel` drops the matching passthrough.
+ *
+ *   wa2. **`ContractErrorCode` narrows.** `TOOL_NOT_FOUND`,
+ *      `TOOL_THREW`, and `TOOL_TIMEOUT` leave the named union — their
+ *      only emitters were the synchronous dispatch routers. The union
+ *      stays extensibly-closed (`(string & {})` branch);
+ *      `SCHEMA_VIOLATION` survives as the generic
+ *      value-violates-contract-declared-schema code.
+ *
+ *   wa3. **`ContractErrorPayload` narrows to
+ *      `{toolName, error, timestamp, schemaVersion?}`.**
+ *      `actionName` and `sourceAction` (the
+ *      `'wired-action' | 'refresh-stream'` dispatch provenance) are
+ *      deleted along with their `validateContractErrorPayload` arms
+ *      and the `makeContractErrorPayload` builder inputs. The
+ *      `_ggui:contract-error` reserved channel, the payload validator,
+ *      and `BUILTIN_RESERVED_VALIDATORS` all remain defined vocabulary.
+ *
+ *   wa4. **`actionNextSteps` deleted from the `ai.ggui/render`
+ *      slice.** The per-action nextStep map (SPEC name
+ *      `wiredActionTools`) existed to gate the client-side direct
+ *      `tools/call` arm, which is retired; the agent hint now derives
+ *      server-side from the render's `actionSpec[name].nextStep` when
+ *      the consume event is built. `RefreshInput` /
+ *      `EMPTY_REFRESH_INPUT` (router-only refresh-tool input types)
+ *      are deleted with the dispatch path.
+ *
+ * --------------------------------------------------------------------
+ * `AgentToolEntry` restructured — `{serverInfo?, toolInfo, usage, example}`
+ * (2026-06-02, BREAKING, pre-launch):
+ *
+ *   z1. **`AgentToolEntry` fields regrouped.** The flat per-tool entry
+ *      (`{description?, usage?, inputSchema?, outputSchema?, example?}`)
+ *      is restructured to separate MCP-native descriptor from ggui
+ *      authoring metadata:
+ *        - NEW `serverInfo?: {name, version?}` — owning MCP server
+ *          identity from the `initialize` handshake. OPTIONAL: populated
+ *          by the agent-server catalog-builder / a host that exposes it;
+ *          absent for hand-authored contracts. `(server, toolName)` is
+ *          the canonical cross-framework identity; `serverInfo.version`
+ *          is metadata, NOT identity.
+ *        - NEW `toolInfo: {inputSchema, description?, outputSchema?}` —
+ *          the MCP `tools/list` descriptor echoed verbatim (minus
+ *          `name`, which is the catalog key). `inputSchema` is now
+ *          REQUIRED (every MCP tool has one); the former top-level
+ *          `description` / `inputSchema` / `outputSchema` MOVED here.
+ *        - `usage` + `example` stay on the entry (the ggui authoring
+ *          layer, not MCP).
+ *      Entries remain keyed by bare `toolName` in
+ *      `agentCapabilities.tools` (unchanged). The canonical contract
+ *      hash strips ONLY `serverInfo.version` (metadata, not identity);
+ *      `serverInfo.name` IS identity — it disambiguates two servers that
+ *      expose a byte-identical bare tool name (e.g. Todoist vs Google-Tasks
+ *      `todo_add`) — and remains in the hash, alongside
+ *      `toolInfo.inputSchema` / `toolInfo.outputSchema`. (Refined
+ *      2026-06-03: name promoted to identity; `serverInfo.version`
+ *      relaxed to optional.)
+ *      Pre-launch no-backcompat: every contract author / catalog-builder
+ *      that emitted the flat shape MUST migrate to the nested
+ *      `toolInfo`. `agentToolEntrySchema` is `.strict()` on both the
+ *      outer entry and the inner `toolInfo` / `serverInfo` / `example`
+ *      objects.
+ *
+ * --------------------------------------------------------------------
  * R4 — auth-credential `bootstrap` renamed to `wsToken`; aggregate
  * polling endpoint deleted; console session-meta route renamed
  * (BREAKING, pre-launch):
@@ -16,7 +788,7 @@
  *      on the wire was the credential, so the term lost its meaning.
  *      Renamed everywhere internally to `wsToken` (sharper, symmetric
  *      with `wsUrl`):
- *        - {@link SubscribePayload.bootstrap} → `SubscribePayload.wsToken`
+ *        - `SubscribePayload.bootstrap` → `SubscribePayload.wsToken`
  *        - `?bootstrap=<token>` WS upgrade query → `?wsToken=<token>`
  *        - `McpAppAiGguiSessionMeta.token` slice field → `wsToken`
  *        - `ChannelClientBootstrap.token` → `wsToken`
@@ -392,8 +1164,9 @@
  *      `assertContractSchemasValid(contract)` walks the six inner
  *      JSON Schema fields (`propsSpec.properties[*].schema`,
  *      `actionSpec[*].schema`, `streamSpec[*].schema`,
- *      `contextSpec[*].schema`, `agentCapabilities.tools[*].inputSchema`,
- *      `agentCapabilities.tools[*].outputSchema`) and runs
+ *      `contextSpec[*].schema`,
+ *      `agentCapabilities.tools[*].toolInfo.inputSchema`,
+ *      `agentCapabilities.tools[*].toolInfo.outputSchema`) and runs
  *      `compileForValidation()` on each. Ajv strict mode throws at
  *      compile-time on malformed schemas (unknown keywords, properties
  *      values that aren't schemas, array schemas with non-schema
@@ -434,6 +1207,8 @@
  *        - `handshakeOutputSchema`: dropped `reason`, `target`,
  *          `alternatives`, `contractHash`, `serverCapabilities`.
  *          `serverCapabilities` flows via `_meta.ggui.bootstrap` instead.
+ *          (`reason` was later reinstated as an optional ≤280-char wire
+ *          field — see the o-block below.)
  *        - `pushOutputSchema`: dropped `sessionId`, `shortCode`,
  *          `codeReady`, `handshakeId`, `decision`, `contract`,
  *          `interaction`, `contractHash`, `cache.*`, `codeUrl`,
@@ -1019,9 +1794,10 @@
  *   00000000. Protocol-level schema-compat invariant in
  *      `@ggui-ai/protocol/validation/schema-compat-invariants`:
  *      `CTR_SCHEMA_INCOMPAT` validates `actionSpec[*].schema` is a
- *      subset of the referenced `agentCapabilities.tools[*].inputSchema`,
+ *      subset of the referenced
+ *      `agentCapabilities.tools[*].toolInfo.inputSchema`,
  *      and `streamSpec[*].schema` is a superset of the referenced
- *      `agentCapabilities.tools[*].outputSchema`. Compares against the
+ *      `agentCapabilities.tools[*].toolInfo.outputSchema`. Compares against the
  *      contract's OWN agentCapabilities catalog (no runtime tool registry,
  *      no zod conversion) — author-visible bug class. Distinct from
  *      and complementary to the server-level F4 check
@@ -1546,7 +2322,13 @@
  *      `GGUI_SESSION_UI_META` → `GGUI_RENDER_UI_META`. Error codes
  *      renamed too: `SESSION_NOT_FOUND` → `RENDER_NOT_FOUND`;
  *      `STACK_ITEM_NOT_FOUND` → `RENDER_NOT_FOUND`;
- *      `CONCURRENT_SESSION_LIMIT` → `CONCURRENT_RENDER_LIMIT`.
+ *      `CONCURRENT_SESSION_LIMIT` → `CONCURRENT_RENDER_LIMIT`. The
+ *      live-channel wire-error codes were renamed in the same arc:
+ *      `SESSION_MISMATCH` → `RENDER_MISMATCH`;
+ *      `BOOTSTRAP_SESSION_MISMATCH` → `BOOTSTRAP_RENDER_MISMATCH`;
+ *      `DEVTOOL_COOKIE_SESSION_MISMATCH` →
+ *      `DEVTOOL_COOKIE_RENDER_MISMATCH`;
+ *      `SESSION_CREATE_FAILED` → `RENDER_CREATE_FAILED`.
  *
  *   b6. **Retired wire-side payloads deleted (no shims).** `PushPayload`,
  *      `PopPayload`, `GetStackPayload`, `SessionPayload`,
@@ -1592,12 +2374,39 @@
  *      strings on write; legacy numeric rows are coerced on read.
  *      Cross-deployment uniformity: the same field type ships from
  *      polling HTTP endpoint, WS replay frame, and store-side `observe()`.
+ *
+ *   b9. **GguiSession reintroduced as the render-OBJECT noun
+ *      (2026-06-05).** The flatten arc named the rendered object
+ *      `Render`; this restores `GguiSession` as that noun while `render`
+ *      stays the VERB. Source-breaking TS-symbol renames for
+ *      `@ggui-ai/*` importers: `Render`→`GguiSession` (+ `ComponentRender`,
+ *      `SystemRender`, `McpAppsRender`, `RenderBase`, `RenderStatus`),
+ *      `RenderEvent`→`GguiSessionEvent`, `RenderStore`→`GguiSessionStore`
+ *      (+ every impl), `RenderStreamBuffer`, `RenderNotFoundError`,
+ *      `RenderChannelServer`, `RenderApi`/`RenderInfo`/`RenderRef`,
+ *      `RenderRenderer`→`GguiSessionRenderer`. WIRE-breaking renames: the
+ *      canonical id `renderId`→`sessionId` (every payload/param/envelope/
+ *      URL/_meta field); tools `ggui_list_renders`→`ggui_list_sessions`
+ *      and `ggui_get_render`→`ggui_get_session`; error codes
+ *      `RENDER_NOT_FOUND`/`RENDER_MISMATCH`/`RENDER_CREATE_FAILED`/
+ *      `CONCURRENT_RENDER_LIMIT`→`SESSION_*` (numeric values unchanged);
+ *      REST routes `/api/renders`, `/admin/renders`, `/console/renders`
+ *      → `/…/sessions`; DOM `data-ggui-render-*`→`data-ggui-session-*`;
+ *      ack `renderToken`→`sessionToken`. The VERB `render` is KEPT
+ *      verbatim: the `ggui_render` (+ `ggui_render_blueprint`) tool,
+ *      `renderInputSchema`/`renderOutputSchema` + `GguiRenderInput`/
+ *      `GguiRenderOutput`, the `*Renderer` actor components, the
+ *      `ai.ggui/render` _meta slice key, the `ui://ggui/render` resource
+ *      URI (only its `{sessionId}` path segment follows the id rename),
+ *      and the WS frame discriminators `'render'` / `'render_event'`.
+ *      OSS-first; the cloud pod follows at its next image build. See
+ *      `docs/protocol/migrations/2026-06-05-gguisession-reintroduction.md`.
  */
-export declare const PROTOCOL_VERSION = "draft-2026-05-28";
+export declare const PROTOCOL_VERSION = "draft-2026-06-12";
 /**
  * Schema version stamped onto wire envelopes that opt into the
  * `schemaVersion` forward-compat field (see {@link ActionEnvelope},
- * {@link StreamEnvelope}, {@link ContractErrorPayload}).
+ * {@link StreamEnvelope}).
  *
  * Pre-launch semantics (current): producers SHOULD stamp; consumers
  * SHOULD NOT reject on mismatch — the field is advisory and lets old
@@ -1614,7 +2423,7 @@ export declare const PROTOCOL_VERSION = "draft-2026-05-28";
  * consumers can reference schema-versioning specifically without
  * coupling to the broader cache-invalidation constant.
  */
-export declare const PROTOCOL_SCHEMA_VERSION = "draft-2026-05-28";
+export declare const PROTOCOL_SCHEMA_VERSION = "draft-2026-06-12";
 /**
  * Canonical live-channel error code emitted when a peer's declared
  * {@link SubscribePayload.supportedVersions} /