@claudexor/orchestrator 3.0.3 → 3.1.0

package/dist/attemptTelemetry.d.ts

@@ -5,7 +5,9 @@
  * no regex over prose, and a tool error is "recovered" only when the SAME
  * tool later succeeds against the SAME target.
  */
-import type { AttemptTelemetryRecord, AuthSourceKind, ExternalContextPolicy, HarnessEvent, RequestRequirementResolution, TaskContract, ToolKind } from "@claudexor/schema";
+import type { AttemptTelemetryRecord, AuthSourceKind, ExternalContextPolicy, HarnessEvent, RequestRequirementResolution, TaskContract, ToolKind, WorkState } from "@claudexor/schema";
+import { type TransientFailureObservation } from "./transientClassify.js";
+export { type TransientFailureObservation, classifyAdapterThrow } from "./transientClassify.js";
 export interface ToolErrorRecord {
     tool: string;
     kind: ToolKind;
@@ -23,6 +25,9 @@ export interface AttemptOutcomeState {
     webRequiredUnsatisfied: boolean;
     toolWarningsCount: number;
     status: AttemptOutcomeStatus;
+    /** D-16 model-attested work outcome (from the finalizer); absent on routes
+     * with no work_report transport. */
+    workState?: WorkState;
 }
 export interface WebEvidenceState {
     required: boolean;
@@ -31,10 +36,44 @@ export interface WebEvidenceState {
     attempted: boolean;
     satisfied: boolean;
     failed: boolean;
+    /** QA-042: retrieval strength — "verified" once any web result carried a
+     * typed successful retrieval, else "dispatched" once web activity completed
+     * with no typed outcome (codex), else "none". Never downgrades verified. */
+    verification: "verified" | "dispatched" | "none";
     tool: string | null;
     target: string | null;
     errorSummary: string | null;
 }
+/**
+ * QA-040: runtime browser-MCP evidence for one attempt. `requested` is set when
+ * the engine armed the browser injection (a fixed `browser` server namespace).
+ * A browser tool call/result matched to that injected server flips attempted/
+ * satisfied/failed — so a successful browser navigation is recognized as
+ * trusted live-web activity even though adapters normalize browser calls as
+ * `kind:"mcp"`. Spoof-resistant: only the engine-injected server name matches.
+ */
+export interface BrowserEvidenceState {
+    requested: boolean;
+    serverName: string | null;
+    attempted: boolean;
+    satisfied: boolean;
+    failed: boolean;
+}
+/**
+ * Delegation-belt runtime readiness for one attempt (QA-024). `requested` is
+ * set at attempt creation when a belt MCP server was injected into the spec;
+ * `ready`/`failed` are filled from the harness's `started` event (its
+ * `mcp_servers[<belt>].status`); `toolEvidence` flips when any `mcp__<belt>__*`
+ * tool actually runs. A requested belt that reports `failed` with no tool
+ * evidence is the false-success trap the outcome axis must catch.
+ */
+export interface DelegationBeltState {
+    requested: boolean;
+    serverName: string | null;
+    ready: boolean;
+    failed: boolean;
+    toolEvidence: boolean;
+}
 export interface AttemptTelemetry {
     requestRequirements: RequestRequirementResolution[];
     toolErrors: ToolErrorRecord[];
@@ -58,17 +97,32 @@ export interface AttemptTelemetry {
     /** Model hint the engine SENT this attempt (requested side; observedModel is
      * the disclosed side of the model x route truth). */
     requestedModel: string | null;
-    /** Adapter-declared transient failures seen during this attempt. */
-    transientFailures: {
-        kind: NonNullable<HarnessEvent["transient"]>["kind"];
-        retryDelayMs: number | null;
-    }[];
+    /** Adapter-declared transient failures seen during this attempt, each
+     * classified into the GH #31 typed taxonomy the retry policy gates on. */
+    transientFailures: TransientFailureObservation[];
     /** TYPED vendor rate-limit signals seen during this attempt (W5.4): the
      * rotation predicate reads these, never prose or plain transients. */
     rateLimits: {
         retryDelayMs: number | null;
         resetsAt: string | null;
     }[];
+    /** Delegation-belt runtime readiness (QA-024); requested=false on non-delegate attempts. */
+    delegationBelt: DelegationBeltState;
+    /** Browser-MCP runtime evidence (QA-040); requested=false unless the browser
+     * injection was armed for this attempt. */
+    browser: BrowserEvidenceState;
+    /** D-16: a terminal `capacity_exhausted` context signal was observed this
+     * attempt (never a transient; consumed by the finalizer, not the retry loop). */
+    contextExhausted: boolean;
+    /** D-16d: the typed cause of the terminal capacity exhaustion (last-wins),
+     * or null when none observed. The continuation controller keys eligibility on
+     * `repeated_refill` (claude's rapid-refill breaker), never on `prompt_too_long`
+     * (an irreducible packet). */
+    contextExhaustedCause: NonNullable<HarnessEvent["context"]>["cause"] | null;
+    /** D-16c: the raw `{work_report}` payload a `side_tool` route surfaced on its
+     * final message (claude StructuredOutput tool), or null. The unwrap validates
+     * it while the markdown answer stays the deliverable. */
+    sideToolWorkReport: unknown;
     /** Contract/outcome truth for this attempt, produced by the orchestrator. */
     outcome: AttemptOutcomeState | null;
     /** Token usage summed across this attempt's usage events (money stays in the
@@ -89,7 +143,30 @@ export interface AttemptTelemetry {
         unknownUsd: number;
     };
 }
-export declare function createAttemptTelemetry(policy: ExternalContextPolicy, webRequired: boolean, effectiveMode?: ExternalContextPolicy, requestRequirements?: RequestRequirementResolution[], requestedModel?: string | null): AttemptTelemetry;
+export declare function createAttemptTelemetry(policy: ExternalContextPolicy, webRequired: boolean, effectiveMode?: ExternalContextPolicy, requestRequirements?: RequestRequirementResolution[], requestedModel?: string | null, 
+/** The delegation-belt MCP server name injected into THIS attempt's spec, or
+ * null when no belt was injected (QA-024). Non-null marks the belt requested. */
+beltServerName?: string | null, 
+/** The browser MCP server name armed for THIS attempt (fixed `browser`
+ * namespace), or null when the browser was not injected (QA-040). Non-null
+ * marks the browser requested/armed. */
+browserServerName?: string | null): AttemptTelemetry;
+/**
+ * The delegation belt was requested (--delegate injected it) but never became
+ * operational: the harness reported the server `failed` and no belt tool ever
+ * ran (QA-024). This is the false-success trap — the harness may have answered
+ * from its own native subagent with no Claudexor sub-run provenance. A belt
+ * that was ready-but-unused is NOT unavailable (docs leave the spawn decision to
+ * the harness); only a startup failure counts.
+ */
+export declare function delegationBeltUnavailable(t: AttemptTelemetry): boolean;
+/**
+ * QA-040: the browser was armed but never called while generic web evidence
+ * satisfied the run — a legitimately policy-satisfied run (e.g. web_search
+ * answered) that nonetheless did not exercise the requested browser. Disclosure
+ * only: it must NOT fail the run.
+ */
+export declare function browserUnused(t: AttemptTelemetry): boolean;
 /**
  * Observe a normalized harness event into the attempt telemetry. Governance is
  * fully typed: only the `tool` ToolRef on tool_call/tool_result/file_change
@@ -104,6 +181,10 @@ export declare function setAttemptOutcome(t: AttemptTelemetry, opts: {
     gatesPassed: boolean | null;
     harnessErrored: boolean;
     webRequiredUnsatisfied: boolean;
+    /** D-16 work_state from the unified finalizer (INV-116): a needs_input/
+     * incomplete veto rides HERE without flipping `status` — the lifecycle
+     * stays succeeded-class; applyability and the CLI exit read the axis. */
+    workState?: WorkState;
 }): void;
 /** Bounded telemetry summary for events/artifacts (full detail lives in telemetry.yaml). */
 export declare function telemetrySummary(t: AttemptTelemetry): Record<string, unknown>;

package/dist/attemptTelemetry.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"attemptTelemetry.d.ts","sourceRoot":"","sources":["../src/attemptTelemetry.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,OAAO,KAAK,EACV,sBAAsB,EACtB,cAAc,EACd,qBAAqB,EACrB,YAAY,EACZ,4BAA4B,EAC5B,YAAY,EACZ,QAAQ,EACT,MAAM,mBAAmB,CAAC;AAG3B,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,QAAQ,CAAC;IACf,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,qHAAqH;IACrH,SAAS,EAAE,OAAO,CAAC;CACpB;AAED,MAAM,MAAM,oBAAoB,GAAG,SAAS,GAAG,uBAAuB,GAAG,SAAS,GAAG,QAAQ,CAAC;AAE9F,MAAM,WAAW,mBAAmB;IAClC,kBAAkB,EAAE,OAAO,CAAC;IAC5B,WAAW,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,cAAc,EAAE,OAAO,CAAC;IACxB,sBAAsB,EAAE,OAAO,CAAC;IAChC,iBAAiB,EAAE,MAAM,CAAC;IAC1B,MAAM,EAAE,oBAAoB,CAAC;CAC9B;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,EAAE,OAAO,CAAC;IAClB,IAAI,EAAE,qBAAqB,CAAC;IAC5B,aAAa,EAAE,qBAAqB,CAAC;IACrC,SAAS,EAAE,OAAO,CAAC;IACnB,SAAS,EAAE,OAAO,CAAC;IACnB,MAAM,EAAE,OAAO,CAAC;IAChB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;CAC7B;AAED,MAAM,WAAW,gBAAgB;IAC/B,mBAAmB,EAAE,4BAA4B,EAAE,CAAC;IACpD,UAAU,EAAE,eAAe,EAAE,CAAC;IAC9B,+EAA+E;IAC/E,iBAAiB,EAAE,MAAM,CAAC;IAC1B,wEAAwE;IACxE,aAAa,EAAE,MAAM,CAAC;IACtB,GAAG,EAAE,gBAAgB,CAAC;IACtB,4EAA4E;IAC5E,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,8FAA8F;IAC9F,QAAQ,EAAE,eAAe,GAAG,SAAS,GAAG,IAAI,CAAC;IAC7C,+EAA+E;IAC/E,eAAe,EAAE,eAAe,GAAG,SAAS,GAAG,IAAI,CAAC;IACpD,gFAAgF;IAChF,UAAU,EAAE,cAAc,GAAG,IAAI,CAAC;IAClC;;kFAE8E;IAC9E,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB;yDACqD;IACrD,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,oEAAoE;IACpE,iBAAiB,EAAE;QACjB,IAAI,EAAE,WAAW,CAAC,YAAY,CAAC,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QACrD,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;KAC7B,EAAE,CAAC;IACJ;0EACsE;IACtE,UAAU,EAAE;QAAE,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE,CAAC;IACvE,6EAA6E;IAC7E,OAAO,EAAE,mBAAmB,GAAG,IAAI,CAAC;IACpC;;;;qFAIiF;IACjF,KAAK,EAAE;QACL,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;QAC3B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAC5B,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;KAClC,CAAC;IACF;6EACyE;IACzE,SAAS,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAE,CAAC;CAC1E;AAED,wBAAgB,sBAAsB,CACpC,MAAM,EAAE,qBAAqB,EAC7B,WAAW,EAAE,OAAO,EACpB,aAAa,GAAE,qBAA8B,EAC7C,mBAAmB,GAAE,4BAA4B,EAAO,EACxD,cAAc,GAAE,MAAM,GAAG,IAAW,GACnC,gBAAgB,CA6BlB;AAQD;;;;;GAKG;AACH,wBAAgB,uBAAuB,CAAC,CAAC,EAAE,gBAAgB,EAAE,EAAE,EAAE,YAAY,GAAG,IAAI,CAmJnF;AAID,wBAAgB,qBAAqB,CAAC,CAAC,EAAE,gBAAgB,GAAG,eAAe,EAAE,CAE5E;AAED,wBAAgB,YAAY,CAAC,CAAC,EAAE,gBAAgB,GAAG,eAAe,EAAE,CAQnE;AAED,wBAAgB,iBAAiB,CAC/B,CAAC,EAAE,gBAAgB,EACnB,IAAI,EAAE;IACJ,kBAAkB,EAAE,OAAO,CAAC;IAC5B,WAAW,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,cAAc,EAAE,OAAO,CAAC;IACxB,sBAAsB,EAAE,OAAO,CAAC;CACjC,GACA,IAAI,CAkBN;AAWD,4FAA4F;AAC5F,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,gBAAgB,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAuC7E;AAED,wBAAgB,sBAAsB,CACpC,SAAS,EAAE,MAAM,EACjB,SAAS,EAAE,MAAM,EACjB,CAAC,EAAE,gBAAgB,GAClB,sBAAsB,CAoDxB;AAED;;mEAEmE;AACnE,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,sBAAsB,EAAE,GAAG;IACzE,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;CACpC,CAcA;AAED,wBAAgB,uBAAuB,CACrC,OAAO,EAAE,sBAAsB,EAAE,EACjC,QAAQ,EAAE,YAAY,GACrB,sBAAsB,CAAC,KAAK,CAAC,CAoB/B;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAI3D"}
+{"version":3,"file":"attemptTelemetry.d.ts","sourceRoot":"","sources":["../src/attemptTelemetry.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,OAAO,KAAK,EACV,sBAAsB,EACtB,cAAc,EACd,qBAAqB,EACrB,YAAY,EACZ,4BAA4B,EAC5B,YAAY,EACZ,QAAQ,EACR,SAAS,EACV,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EACL,KAAK,2BAA2B,EAKjC,MAAM,wBAAwB,CAAC;AAEhC,OAAO,EAAE,KAAK,2BAA2B,EAAE,oBAAoB,EAAE,MAAM,wBAAwB,CAAC;AAEhG,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,EAAE,QAAQ,CAAC;IACf,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,qHAAqH;IACrH,SAAS,EAAE,OAAO,CAAC;CACpB;AAED,MAAM,MAAM,oBAAoB,GAAG,SAAS,GAAG,uBAAuB,GAAG,SAAS,GAAG,QAAQ,CAAC;AAE9F,MAAM,WAAW,mBAAmB;IAClC,kBAAkB,EAAE,OAAO,CAAC;IAC5B,WAAW,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,cAAc,EAAE,OAAO,CAAC;IACxB,sBAAsB,EAAE,OAAO,CAAC;IAChC,iBAAiB,EAAE,MAAM,CAAC;IAC1B,MAAM,EAAE,oBAAoB,CAAC;IAC7B;wCACoC;IACpC,SAAS,CAAC,EAAE,SAAS,CAAC;CACvB;AAED,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,EAAE,OAAO,CAAC;IAClB,IAAI,EAAE,qBAAqB,CAAC;IAC5B,aAAa,EAAE,qBAAqB,CAAC;IACrC,SAAS,EAAE,OAAO,CAAC;IACnB,SAAS,EAAE,OAAO,CAAC;IACnB,MAAM,EAAE,OAAO,CAAC;IAChB;;gFAE4E;IAC5E,YAAY,EAAE,UAAU,GAAG,YAAY,GAAG,MAAM,CAAC;IACjD,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;CAC7B;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,oBAAoB;IACnC,SAAS,EAAE,OAAO,CAAC;IACnB,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,SAAS,EAAE,OAAO,CAAC;IACnB,SAAS,EAAE,OAAO,CAAC;IACnB,MAAM,EAAE,OAAO,CAAC;CACjB;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,mBAAmB;IAClC,SAAS,EAAE,OAAO,CAAC;IACnB,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,KAAK,EAAE,OAAO,CAAC;IACf,MAAM,EAAE,OAAO,CAAC;IAChB,YAAY,EAAE,OAAO,CAAC;CACvB;AAED,MAAM,WAAW,gBAAgB;IAC/B,mBAAmB,EAAE,4BAA4B,EAAE,CAAC;IACpD,UAAU,EAAE,eAAe,EAAE,CAAC;IAC9B,+EAA+E;IAC/E,iBAAiB,EAAE,MAAM,CAAC;IAC1B,wEAAwE;IACxE,aAAa,EAAE,MAAM,CAAC;IACtB,GAAG,EAAE,gBAAgB,CAAC;IACtB,4EAA4E;IAC5E,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,8FAA8F;IAC9F,QAAQ,EAAE,eAAe,GAAG,SAAS,GAAG,IAAI,CAAC;IAC7C,+EAA+E;IAC/E,eAAe,EAAE,eAAe,GAAG,SAAS,GAAG,IAAI,CAAC;IACpD,gFAAgF;IAChF,UAAU,EAAE,cAAc,GAAG,IAAI,CAAC;IAClC;;kFAE8E;IAC9E,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB;yDACqD;IACrD,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B;8EAC0E;IAC1E,iBAAiB,EAAE,2BAA2B,EAAE,CAAC;IACjD;0EACsE;IACtE,UAAU,EAAE;QAAE,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE,CAAC;IACvE,4FAA4F;IAC5F,cAAc,EAAE,mBAAmB,CAAC;IACpC;+CAC2C;IAC3C,OAAO,EAAE,oBAAoB,CAAC;IAC9B;qFACiF;IACjF,gBAAgB,EAAE,OAAO,CAAC;IAC1B;;;kCAG8B;IAC9B,qBAAqB,EAAE,WAAW,CAAC,YAAY,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,IAAI,CAAC;IAC5E;;6DAEyD;IACzD,kBAAkB,EAAE,OAAO,CAAC;IAC5B,6EAA6E;IAC7E,OAAO,EAAE,mBAAmB,GAAG,IAAI,CAAC;IACpC;;;;qFAIiF;IACjF,KAAK,EAAE;QACL,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;QAC3B,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAC5B,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAC;KAClC,CAAC;IACF;6EACyE;IACzE,SAAS,EAAE;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAE,CAAC;CAC1E;AAED,wBAAgB,sBAAsB,CACpC,MAAM,EAAE,qBAAqB,EAC7B,WAAW,EAAE,OAAO,EACpB,aAAa,GAAE,qBAA8B,EAC7C,mBAAmB,GAAE,4BAA4B,EAAO,EACxD,cAAc,GAAE,MAAM,GAAG,IAAW;AACpC;iFACiF;AACjF,cAAc,GAAE,MAAM,GAAG,IAAW;AACpC;;wCAEwC;AACxC,iBAAiB,GAAE,MAAM,GAAG,IAAW,GACtC,gBAAgB,CA+ClB;AAgCD;;;;;;;GAOG;AACH,wBAAgB,yBAAyB,CAAC,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAEtE;AAgBD;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAE1D;AAUD;;;;;GAKG;AACH,wBAAgB,uBAAuB,CAAC,CAAC,EAAE,gBAAgB,EAAE,EAAE,EAAE,YAAY,GAAG,IAAI,CA+OnF;AAID,wBAAgB,qBAAqB,CAAC,CAAC,EAAE,gBAAgB,GAAG,eAAe,EAAE,CAE5E;AAED,wBAAgB,YAAY,CAAC,CAAC,EAAE,gBAAgB,GAAG,eAAe,EAAE,CAQnE;AAED,wBAAgB,iBAAiB,CAC/B,CAAC,EAAE,gBAAgB,EACnB,IAAI,EAAE;IACJ,kBAAkB,EAAE,OAAO,CAAC;IAC5B,WAAW,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,cAAc,EAAE,OAAO,CAAC;IACxB,sBAAsB,EAAE,OAAO,CAAC;IAChC;;6EAEyE;IACzE,SAAS,CAAC,EAAE,SAAS,CAAC;CACvB,GACA,IAAI,CA4BN;AAWD,4FAA4F;AAC5F,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,gBAAgB,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAiE7E;AAED,wBAAgB,sBAAsB,CACpC,SAAS,EAAE,MAAM,EACjB,SAAS,EAAE,MAAM,EACjB,CAAC,EAAE,gBAAgB,GAClB,sBAAsB,CAsFxB;AAED;;mEAEmE;AACnE,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,sBAAsB,EAAE,GAAG;IACzE,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;IAC5B,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,mBAAmB,EAAE,MAAM,GAAG,IAAI,CAAC;CACpC,CAcA;AAED,wBAAgB,uBAAuB,CACrC,OAAO,EAAE,sBAAsB,EAAE,EACjC,QAAQ,EAAE,YAAY,GACrB,sBAAsB,CAAC,KAAK,CAAC,CAqB/B;AAED;;;;;GAKG;AACH,wBAAgB,cAAc,CAAC,CAAC,EAAE,gBAAgB,GAAG,OAAO,CAI3D"}

package/dist/attemptTelemetry.js

@@ -1,5 +1,14 @@
 import { redactSecrets } from "@claudexor/util";
-export function createAttemptTelemetry(policy, webRequired, effectiveMode = policy, requestRequirements = [], requestedModel = null) {
+import { classifyCompletedCrash, classifyRateLimit, classifyStatusError, classifyTransientSignal, } from "./transientClassify.js";
+export { classifyAdapterThrow } from "./transientClassify.js";
+export function createAttemptTelemetry(policy, webRequired, effectiveMode = policy, requestRequirements = [], requestedModel = null, 
+/** The delegation-belt MCP server name injected into THIS attempt's spec, or
+ * null when no belt was injected (QA-024). Non-null marks the belt requested. */
+beltServerName = null, 
+/** The browser MCP server name armed for THIS attempt (fixed `browser`
+ * namespace), or null when the browser was not injected (QA-040). Non-null
+ * marks the browser requested/armed. */
+browserServerName = null) {
     return {
         requestRequirements,
         toolErrors: [],
@@ -12,6 +21,7 @@ export function createAttemptTelemetry(policy, webRequired, effectiveMode = poli
             attempted: false,
             satisfied: false,
             failed: false,
+            verification: "none",
             tool: null,
             target: null,
             errorSummary: null,
@@ -24,6 +34,23 @@ export function createAttemptTelemetry(policy, webRequired, effectiveMode = poli
         requestedModel,
         transientFailures: [],
         rateLimits: [],
+        delegationBelt: {
+            requested: beltServerName !== null,
+            serverName: beltServerName,
+            ready: false,
+            failed: false,
+            toolEvidence: false,
+        },
+        browser: {
+            requested: browserServerName !== null,
+            serverName: browserServerName,
+            attempted: false,
+            satisfied: false,
+            failed: false,
+        },
+        contextExhausted: false,
+        contextExhaustedCause: null,
+        sideToolWorkReport: null,
         outcome: null,
         usage: { inputTokens: null, outputTokens: null, cachedInputTokens: null },
         usageCost: { cashUsd: 0, valuationUsd: 0, unknownUsd: 0 },
@@ -34,6 +61,76 @@ export function createAttemptTelemetry(policy, webRequired, effectiveMode = poli
 function addToken(acc, value) {
     return value === undefined ? acc : (acc ?? 0) + value;
 }
+/**
+ * Read the injected belt server's status out of the harness `started` frame's
+ * `mcp_servers` list (QA-024). The shape is the vendor's — claude emits
+ * `{ name, status }` entries — so we defensively narrow each entry and match by
+ * the injected belt server name. `status:"failed"` (or "error") is the startup
+ * failure the outcome axis must not let terminalize a silent success.
+ */
+function observeBeltStartup(t, ev) {
+    const payload = ev.payload;
+    const servers = payload?.["mcp_servers"];
+    if (!Array.isArray(servers))
+        return;
+    for (const raw of servers) {
+        if (!raw || typeof raw !== "object")
+            continue;
+        const entry = raw;
+        if (entry.name !== t.delegationBelt.serverName)
+            continue;
+        const status = typeof entry.status === "string" ? entry.status.toLowerCase() : "";
+        if (status === "failed" || status === "error")
+            t.delegationBelt.failed = true;
+        else if (status === "connected" || status === "ready" || status === "ok") {
+            t.delegationBelt.ready = true;
+        }
+        return;
+    }
+}
+/**
+ * The delegation belt was requested (--delegate injected it) but never became
+ * operational: the harness reported the server `failed` and no belt tool ever
+ * ran (QA-024). This is the false-success trap — the harness may have answered
+ * from its own native subagent with no Claudexor sub-run provenance. A belt
+ * that was ready-but-unused is NOT unavailable (docs leave the spawn decision to
+ * the harness); only a startup failure counts.
+ */
+export function delegationBeltUnavailable(t) {
+    return t.delegationBelt.requested && t.delegationBelt.failed && !t.delegationBelt.toolEvidence;
+}
+/**
+ * QA-040: does this tool ref belong to the engine-armed browser MCP? Adapters
+ * normalize browser calls as `kind:"mcp"` (codex `browser:browser_navigate`,
+ * claude `mcp__browser__browser_navigate`), so the ToolKind cannot express
+ * "browser". Match on the ENGINE-INJECTED server namespace only — a user MCP
+ * server cannot spoof trusted browser evidence because the browser is matched
+ * solely when the engine armed it under its fixed injected name.
+ */
+function matchesBrowser(t, tool) {
+    const server = t.browser.serverName;
+    if (!t.browser.requested || !server)
+        return false;
+    return tool.name.startsWith(`mcp__${server}__`) || (tool.target ?? "").startsWith(`${server}:`);
+}
+/**
+ * QA-040: the browser was armed but never called while generic web evidence
+ * satisfied the run — a legitimately policy-satisfied run (e.g. web_search
+ * answered) that nonetheless did not exercise the requested browser. Disclosure
+ * only: it must NOT fail the run.
+ */
+export function browserUnused(t) {
+    return t.browser.requested && !t.browser.attempted && t.web.satisfied;
+}
+/** QA-042: raise the web retrieval strength, never downgrading a proven
+ * `verified`. A typed success promotes to verified; any other satisfied web
+ * activity is dispatch-strength (completed, content not typed-proven). */
+function bumpWebVerification(t, retrieval) {
+    if (retrieval === "verified")
+        t.web.verification = "verified";
+    else if (t.web.verification !== "verified")
+        t.web.verification = "dispatched";
+}
 /**
  * Observe a normalized harness event into the attempt telemetry. Governance is
  * fully typed: only the `tool` ToolRef on tool_call/tool_result/file_change
@@ -41,6 +138,21 @@ function addToken(acc, value) {
  * matching or tool-name heuristics.
  */
 export function observeAttemptTelemetry(t, ev) {
+    // Delegation belt readiness (QA-024): the harness's `started` frame lists its
+    // MCP servers and each one's status. When the engine injected a belt, read
+    // THAT server's status as first-class readiness truth — never prose. A
+    // `failed` belt with no later tool evidence is the false-success trap.
+    if (ev.type === "started" && t.delegationBelt.requested) {
+        observeBeltStartup(t, ev);
+    }
+    // Belt tool evidence: any `mcp__<belt>__*` tool call/result proves the belt
+    // was actually reachable and used (a real Claudexor sub-run path), which
+    // distinguishes a used belt from one the harness silently substituted.
+    if (t.delegationBelt.requested && t.delegationBelt.serverName && ev.tool?.name) {
+        if (ev.tool.name.startsWith(`mcp__${t.delegationBelt.serverName}`)) {
+            t.delegationBelt.toolEvidence = true;
+        }
+    }
     // Route evidence: remember the model identity the stream itself disclosed.
     if (ev.observed_model && !t.observedModel)
         t.observedModel = ev.observed_model;
@@ -83,18 +195,38 @@ export function observeAttemptTelemetry(t, ev) {
     // produced the deliverable.
     if (ev.credential_profile_id)
         t.profileId = ev.credential_profile_id;
-    if (ev.transient) {
-        t.transientFailures.push({
-            kind: ev.transient.kind,
-            retryDelayMs: ev.transient.retry_delay_ms ?? null,
-        });
-    }
+    // #31: classify every disclosed transient into the typed taxonomy (see
+    // transientClassify.ts). An adapter `transient` and a `rate_limit` are
+    // retryable failures; the vendor's typed `status.error_category` surfaces only
+    // the deterministic FAILURE classes (auth/capability/config) so required-
+    // actions attach the right remediation. Rate limits ALSO stay in rateLimits
+    // for the W5.4 rotation predicate.
+    if (ev.transient)
+        t.transientFailures.push(classifyTransientSignal(ev.transient));
     if (ev.rate_limit) {
         t.rateLimits.push({
             retryDelayMs: ev.rate_limit.retry_delay_ms ?? null,
             resetsAt: ev.rate_limit.resets_at ?? null,
         });
+        t.transientFailures.push(classifyRateLimit(ev.rate_limit.retry_delay_ms ?? null));
     }
+    if (ev.status?.error_category) {
+        const obs = classifyStatusError(ev.status.error_category, ev.status.retry_delay_ms ?? null);
+        if (obs)
+            t.transientFailures.push(obs);
+    }
+    // D-16 context signal: a terminal capacity_exhausted marks the attempt for the
+    // finalizer's interrupted/context_capacity_exhausted mapping. Context signals
+    // NEVER enter the transient-retry loop (they are not transient failures).
+    if (ev.context?.kind === "capacity_exhausted") {
+        t.contextExhausted = true;
+        t.contextExhaustedCause = ev.context.cause ?? null;
+    }
+    // D-16c side_tool: capture the raw work_report the StructuredOutput tool
+    // surfaced on the final message (last-wins). The unwrap validates it.
+    const sideTool = ev.payload?.["work_report_side_tool"];
+    if (sideTool !== undefined)
+        t.sideToolWorkReport = sideTool;
     // Token usage: SUM across the attempt's usage events (single-event adapters
     // sum == last-wins; codex per-turn needs the sum). Money is the ledger's job.
     if (ev.type === "usage" && ev.usage) {
@@ -107,6 +239,11 @@ export function observeAttemptTelemetry(t, ev) {
             Number(ev.payload?.["dropped_unrecognized_events"] ?? 0);
         if (Number.isFinite(dropped) && dropped > 0)
             t.droppedEvents += dropped;
+        // #31 process crash: the run loop discloses a non-aborted signal kill, a
+        // non-zero exit, or a spawn failure as TYPED payload fields (never prose).
+        const crash = classifyCompletedCrash(ev.payload);
+        if (crash)
+            t.transientFailures.push(crash);
         return;
     }
     const tool = ev.tool;
@@ -118,6 +255,15 @@ export function observeAttemptTelemetry(t, ev) {
             t.web.tool = tool.name;
             t.web.target = tool.target ?? t.web.target;
         }
+        // QA-040: an armed browser MCP call is trusted live-web ACTIVITY even though
+        // its kind is "mcp" — mark the browser attempted and count it as web
+        // attempted (a successful result below satisfies the generic web gate).
+        if (matchesBrowser(t, tool)) {
+            t.browser.attempted = true;
+            t.web.attempted = true;
+            t.web.tool = t.web.tool ?? tool.name;
+            t.web.target = tool.target ?? t.web.target;
+        }
         return;
     }
     if (ev.type !== "tool_result")
@@ -133,6 +279,10 @@ export function observeAttemptTelemetry(t, ev) {
             t.web.tool = tool.name;
             t.web.target = tool.target ?? t.web.target;
         }
+        if (matchesBrowser(t, tool)) {
+            t.browser.attempted = true;
+            t.web.attempted = true;
+        }
         return;
     }
     if (tool.status === "error") {
@@ -151,6 +301,19 @@ export function observeAttemptTelemetry(t, ev) {
             t.web.target = tool.target ?? t.web.target;
             t.web.errorSummary = redactSecrets(tool.error_summary ?? "web tool result marked error").slice(0, 1000);
         }
+        // QA-040: a failed armed-browser call is browser activity that did not
+        // satisfy. It contributes web `attempted` (a call was made) and, since the
+        // browser is a live-egress channel, a web failure — but a later successful
+        // web/browser call still recovers the generic gate (the ok branch below).
+        if (matchesBrowser(t, tool)) {
+            t.browser.attempted = true;
+            t.browser.failed = true;
+            t.web.attempted = true;
+            t.web.failed = true;
+            t.web.errorSummary =
+                t.web.errorSummary ??
+                    redactSecrets(tool.error_summary ?? "browser tool result marked error").slice(0, 1000);
+        }
         return;
     }
     // status === "ok": a later success of the SAME tool against the SAME target
@@ -191,6 +354,34 @@ export function observeAttemptTelemetry(t, ev) {
             : null;
         t.web.tool = tool.name;
         t.web.target = tool.target ?? t.web.target;
+        // QA-042: the retrieval STRENGTH. A typed `verified` retrieval (claude
+        // WebFetch content) proves content; codex `web_search`/`open_page` stamp
+        // `dispatched` (completed, no typed fetch outcome — a hidden 502 is
+        // indistinguishable), so the gate is satisfied at DISPATCH strength only.
+        // Absent stamp is treated as dispatch (never claim verified without proof).
+        bumpWebVerification(t, tool.web_retrieval);
+    }
+    // QA-040: a successful armed-browser call is trusted live-web evidence. It
+    // satisfies the generic web gate AND records browser runtime evidence — a
+    // real navigation/screenshot is a typed success, so it is `verified` strength
+    // (unlike a dispatch-only codex web_search). A user MCP server cannot reach
+    // here: matchesBrowser only accepts the engine-injected server namespace.
+    if (matchesBrowser(t, tool)) {
+        t.browser.attempted = true;
+        t.browser.satisfied = true;
+        t.web.attempted = true;
+        t.web.satisfied = true;
+        t.web.verification = "verified";
+        // The recovery loop above already marked a matching earlier browser error
+        // recovered; recompute the disclosed failure rollup from what remains.
+        t.browser.failed = t.toolErrors.some((e) => e.kind === "mcp" &&
+            matchesBrowser(t, { name: e.tool, target: e.target ?? undefined }) &&
+            !e.recovered);
+        t.web.failed = t.toolErrors.some((e) => e.kind === "web" && !e.recovered) || t.browser.failed;
+        if (!t.web.failed)
+            t.web.errorSummary = null;
+        t.web.tool = t.web.tool ?? tool.name;
+        t.web.target = tool.target ?? t.web.target;
     }
 }
 const TELEMETRY_TOOL_ERRORS_MAX = 20;
@@ -209,9 +400,18 @@ export function toolWarnings(t) {
 export function setAttemptOutcome(t, opts) {
     const warnings = toolWarnings(t).length;
     const contractFailed = !opts.deliverablePresent || opts.gatesPassed === false;
+    // QA-024: a requested belt that failed to start with no tool evidence is an
+    // explicitly-requested capability that never became operational — treated
+    // like an unsatisfied hard requirement (never a silent clean success). It
+    // rides the same axis order as web: it can only ELEVATE severity, never mask
+    // a harder failure. NOTE (D-16 seam): this producer maps belt-unavailable to
+    // `failed`; a future finalizer that prefers a softer disclosure would flip
+    // this to `success_with_warnings` — the typed telemetry fact
+    // (delegation_belt.*) is what a consumer reads either way.
+    const beltUnavailable = delegationBeltUnavailable(t);
     const status = opts.webRequiredUnsatisfied
         ? "blocked"
-        : opts.harnessErrored || contractFailed
+        : opts.harnessErrored || contractFailed || beltUnavailable
             ? "failed"
             : warnings > 0
                 ? "success_with_warnings"
@@ -223,6 +423,7 @@ export function setAttemptOutcome(t, opts) {
         webRequiredUnsatisfied: opts.webRequiredUnsatisfied,
         toolWarningsCount: warnings,
         status,
+        ...(opts.workState ? { workState: opts.workState } : {}),
     };
 }
 function webStatus(t) {
@@ -246,10 +447,22 @@ export function telemetrySummary(t) {
             attempted: t.web.attempted,
             satisfied: t.web.satisfied,
             status: webStatus(t),
+            verification: t.web.verification,
             tool: t.web.tool,
             target: t.web.target,
             error_summary: t.web.errorSummary,
         },
+        ...(t.browser.requested
+            ? {
+                browser_evidence: {
+                    server_name: t.browser.serverName,
+                    attempted: t.browser.attempted,
+                    satisfied: t.browser.satisfied,
+                    failed: t.browser.failed,
+                    unused: browserUnused(t),
+                },
+            }
+            : {}),
         tool_errors_total: t.toolErrors.length,
         unrecovered_tool_errors: unrecovered.length,
         tool_errors: unrecovered
@@ -264,11 +477,25 @@ export function telemetrySummary(t) {
             }
             : {}),
         ...(t.outcome ? { outcome: t.outcome } : {}),
+        ...(t.delegationBelt.requested
+            ? {
+                delegation_belt: {
+                    server_name: t.delegationBelt.serverName,
+                    ready: t.delegationBelt.ready,
+                    failed: t.delegationBelt.failed,
+                    tool_evidence: t.delegationBelt.toolEvidence,
+                    unavailable: delegationBeltUnavailable(t),
+                },
+            }
+            : {}),
         ...(t.transientFailures.length > 0
             ? {
-                transient_failures: t.transientFailures
-                    .slice(-5)
-                    .map((e) => ({ kind: e.kind, retry_delay_ms: e.retryDelayMs })),
+                transient_failures: t.transientFailures.slice(-5).map((e) => ({
+                    kind: e.kind,
+                    category: e.category,
+                    retryable: e.retryable,
+                    retry_delay_ms: e.retryDelayMs,
+                })),
             }
             : {}),
         ...(t.droppedEvents > 0 ? { dropped_events: t.droppedEvents } : {}),
@@ -294,6 +521,7 @@ export function attemptTelemetryRecord(attemptId, harnessId, t) {
             attempted: t.web.attempted,
             satisfied: t.web.satisfied,
             status: webStatus(t),
+            verification: t.web.verification,
             tool: t.web.tool,
             target: t.web.target,
             error_summary: t.web.errorSummary,
@@ -310,17 +538,50 @@ export function attemptTelemetryRecord(attemptId, harnessId, t) {
         unrecovered_tool_errors: unrecoveredToolErrors(t).length,
         statusless_tool_results: t.statuslessResults,
         dropped_events: t.droppedEvents,
-        transient_failures: t.transientFailures
-            .slice(-TELEMETRY_TOOL_ERRORS_MAX)
-            .map((e) => ({ kind: e.kind, retry_delay_ms: e.retryDelayMs })),
+        transient_failures: t.transientFailures.slice(-TELEMETRY_TOOL_ERRORS_MAX).map((e) => ({
+            kind: e.kind,
+            category: e.category,
+            retryable: e.retryable,
+            retry_delay_ms: e.retryDelayMs,
+            http_status: e.httpStatus,
+            signal: e.signal,
+            adapter_code: e.adapterCode,
+        })),
         outcome: {
             deliverable_present: t.outcome?.deliverablePresent ?? false,
             gates_passed: t.outcome?.gatesPassed ?? null,
             harness_errored: t.outcome?.harnessErrored ?? false,
             web_required_unsatisfied: t.outcome?.webRequiredUnsatisfied ?? false,
+            delegation_belt_unavailable: delegationBeltUnavailable(t),
             tool_warnings_count: t.outcome?.toolWarningsCount ?? warnings.length,
             status: t.outcome?.status ?? (warnings.length > 0 ? "success_with_warnings" : "success"),
+            ...(t.outcome?.workState ? { work_state: t.outcome.workState } : {}),
         },
+        // Only present when a belt was actually injected into this attempt (QA-024).
+        ...(t.delegationBelt.requested
+            ? {
+                delegation_belt: {
+                    requested: t.delegationBelt.requested,
+                    server_name: t.delegationBelt.serverName,
+                    ready: t.delegationBelt.ready,
+                    failed: t.delegationBelt.failed,
+                    tool_evidence: t.delegationBelt.toolEvidence,
+                },
+            }
+            : {}),
+        // Only present when the browser was armed for this attempt (QA-040).
+        ...(t.browser.requested
+            ? {
+                browser: {
+                    requested: t.browser.requested,
+                    server_name: t.browser.serverName,
+                    attempted: t.browser.attempted,
+                    satisfied: t.browser.satisfied,
+                    failed: t.browser.failed,
+                    unused: browserUnused(t),
+                },
+            }
+            : {}),
         usage: {
             input_tokens: t.usage.inputTokens,
             output_tokens: t.usage.outputTokens,
@@ -360,6 +621,7 @@ export function aggregateRunWebEvidence(records, contract) {
         attempted: false,
         satisfied: false,
         status: contract.external_context.web_required ? "unverified" : "none",
+        verification: "none",
         tool: null,
         target: null,
         error_summary: null,