@wrenai/genbi 0.0.4 → 0.0.6

package/README.md

@@ -25,10 +25,10 @@ the rest:
 
 ## Developer preview
 
-`@wrenai/genbi` is published at `0.0.4`. Nothing here is under a compatibility
+`@wrenai/genbi` is published at `0.0.5`. Nothing here is under a compatibility
 commitment yet — commands, environment variables and interfaces move without
-notice. The Codex runtime is supported in code but has not been exercised end
-to end; Claude is the one that is. See
+notice. Runtime support is limited to the exact versions and execution scopes in the
+packaged certification records; a supported protocol alone does not enable a runtime. See
 [MAINTAINING.md](./MAINTAINING.md#versioning) for what the version number does
 and does not promise.
 
@@ -136,6 +136,47 @@ against real projects and depend on how well the configured model behaves. If a
 test you didn't touch fails and mentions a real query, check whether it's one of
 these before going looking for your own bug.
 
+## Component execution contracts
+
+Composed execution uses the versioned Vercel bundle or native host plan. The host owns
+component selection, reachable children, semantic context, model tiers and tool grants.
+The default dashboard calls `answer_query` for each panel; only the answer component has
+query authority. Child calls receive independent inputs, products and evidence. Successful
+root results are persisted only after all component resources have closed.
+
+The shared host limits an execution to depth 8, 32 child-component invocations per root
+execution, 40 executed steps globally, 12 executed steps per child invocation and 120 seconds. These are host call, step and time limits;
+they do not promise a hard cap on internal model requests made by a vendor runtime.
+
+Result verification uses observed tool/child data. Locked additivity and drill guards
+require query provenance derived by the governed Wren process and matching prepared context.
+The conservative SQL subset includes canonical cube measures, dimensions and simple literal
+filters. Unsupported expressions receive no semantic proof, so guarded execution refuses
+rather than inferring validity from model prose. Filter and order dimensions count toward
+the drill limit. Guarded parents currently refuse child evidence without semantic provenance.
+
+Native composition requires a separately certified, server-provided runtime for the same
+approved account, with fresh per-step resources and revocation checks. Native host interfaces and passing deterministic
+probes do not activate this capability. Claude retains its scope entry; Codex Analysis selects
+one answer or dashboard entry before session creation and keeps that pin for the session.
+A compatible CLI/package tuple and runtime certification are required before enabling the
+native path; the pinned 0.13.0 CLI alone does not provide the new host contracts.
+
+Codex Native sessions use the official CLI's remote terminal interface. A credential-free
+renderer runs in the vendor sandbox and connects to one host-owned session over a private
+Unix socket. The browser terminal supports text input, resize, reconnect and cancellation;
+model, account, project scope and tools remain pinned by the server. General shell access,
+configuration changes, login changes and cross-session commands are unavailable. Ending a
+session closes its processes; restart creates a fresh session rather than resuming a provider
+thread. Remote terminal admission requires its own `remote-tui-v1` certification evidence in
+addition to the driver certification.
+
+`test/component-default-parity.test.ts` compares the real default profile across Vercel and
+both native producer formats using deterministic model/database responses. It exercises the
+host broker, context verifier, result normalizer and native admission, including repeated
+child calls, caller tool isolation, root persistence and correlated events. It does not
+perform account-backed runtime certification.
+
 ## Architecture
 
 ```
@@ -160,3 +201,24 @@ and the UI renders it.
 It is a plain SPA rather than an SSR framework on purpose: it only talks to the
 BFF over HTTP/SSE and builds to static assets, so a future desktop shell or PWA
 stays a thin increment.
+
+
+### Structured Ask query evidence
+
+The direct Codex runtime carries host-normalized query results to the answer independently
+of the model's final text. Tables preserve the executed rows and SQL. A model-written
+`verified` flag cannot promote a text answer. If any root call fails or its evidence cannot
+be represented, the answer falls back to unverified text. Multiple successful query roots
+retain their table/SQL pairs in call order.
+
+The query verification badge covers rows and SQL, not the model's explanation of business
+meaning. Time-range and metric interpretation appears separately. Missing tool-supplied
+lineage or filters is shown as unknown. Questions without a period use the available data
+range and disclose that assumption; relative periods distinguish the current calendar from
+the latest data date. Explanations should identify date bounds, date field, week start,
+timezone (or unknown), partial periods and metric meanings. These are model instructions,
+not proof of business semantics; ambiguous definitions still require clarification.
+
+Structured Ask allows up to five minutes for the outer conversation (including the bounded
+component call and explanation). Individual component/step deadlines remain unchanged. An
+explicit caller timeout is preserved, and outer timeout failures are reported distinctly.