agent-nuvira 3.1.3 → 3.3.0

package/dist/cli/chat.d.ts

@@ -147,6 +147,22 @@ export declare class ChatCommand extends BaseCommand {
      * failing into dead ends — the fire-and-forget keeps the first message fast.
      */
     private coldStartProbeFired;
+    /**
+     * G5 — TRACE FIDELITY. The Auto-router decision snapshot for the current
+     * turn, stashed at resolve time so the reasoning trace can record WHY a step
+     * used the provider/model it did. The audit of the calculator session found
+     * 98/99 steps stamped with the session default `gemini/gemini-3.1-flash-lite`
+     * and **0/99** steps carrying a routing snapshot — so the Trace tab's "model
+     * used" column was really "the session's default model".
+     */
+    private lastRouteSnapshot?;
+    /**
+     * G5 — the provider/model of the LAST generation attempt (post-failover,
+     * post-default-resolution). Read by the chat trace recorder so a step names
+     * the model that actually ran it, instead of whatever the session default
+     * happened to be.
+     */
+    private lastAttempt?;
     /**
      * P3 — programmatic single-turn answer for the dashboard chat console.
      *
@@ -209,6 +225,17 @@ export declare class ChatCommand extends BaseCommand {
          * project, cwd-aware).
          */
         projectContext?: string;
+        /**
+         * Session 3 — CHANNEL / FORMAT POLICY for the STABLE layer.
+         *
+         * The gateway used to prepend the "RESPONSE FORMAT (non-negotiable…)" block
+         * to every INBOUND USER TURN — i.e. the same policy was re-injected as
+         * volatile content on every message (~80% of the visible user turn on
+         * WhatsApp turns). Policy is identical every turn, so it belongs in the
+         * system prompt: it is then byte-stable (prompt-cacheable) and shows up in
+         * the layered trace's `systemDigest` instead of polluting the ask.
+         */
+        systemPolicy?: string;
         /**
          * P4 — the attached project's directory. When set, the turn ALSO recalls
          * that project's prior sessions + facts (`autoRecall`) and injects them
@@ -252,6 +279,12 @@ export declare class ChatCommand extends BaseCommand {
          * carried out. Surfaces must not present such a turn as "in progress".
          */
         unfulfilledPromise?: boolean;
+        /**
+         * G13b — the request asked for an authored deliverable to be produced and
+         * the turn wrote nothing to disk. The reply may be excellent prose; the
+         * artifact does not exist, so no surface may read it as finished work.
+         */
+        undeliveredArtifact?: boolean;
         provider?: string;
         model?: string;
     }>;

package/dist/cli/chat.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chat.d.ts","sourceRoot":"","sources":["../../src/cli/chat.ts"],"names":[],"mappings":"AAIA,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAEpC,OAAO,EAAE,WAAW,EAAc,MAAM,eAAe,CAAC;AAgCxD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAC;AAoBtD;;;;;GAKG;AACH,MAAM,WAAW,YAAY;IAC3B,EAAE,CAAC,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAC/B,EAAE,CAAC,EAAE,OAAO,CAAC;IACb,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,OAAO,EAA+B,KAAK,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACrF,OAAO,EAGL,KAAK,kBAAkB,EACxB,MAAM,4BAA4B,CAAC;AA0JpC,sEAAsE;AACtE,MAAM,WAAW,wBAAwB;IACvC,oDAAoD;IACpD,QAAQ,EAAE,OAAO,CAAC;IAClB,0EAA0E;IAC1E,WAAW,EAAE,OAAO,CAAC;CACtB;AAED,oEAAoE;AACpE,MAAM,WAAW,yBAAyB;IACxC,GAAG,CAAC,EAAE,OAAO,CAAC;IACd,0EAA0E;IAC1E,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,wBAAgB,uBAAuB,CACrC,MAAM,EAAE,aAAa,EACrB,IAAI,CAAC,EAAE,yBAAyB,GAC/B,wBAAwB,CA4B1B;AAID;;;;;;;;GAQG;AACH,wBAAsB,gBAAgB,CACpC,IAAI,EAAE,MAAM,EACZ,aAAa,EAAE,GAAG,EAClB,OAAO,CAAC,EAAE;IAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,GAC9C,OAAO,CAAC,IAAI,CAAC,CAYf;AAED;;;;;;;;;GASG;AACH;;;;;GAKG;AACH,eAAO,MAAM,yBAAyB,uBAA0B,CAAC;AAEjE;;;;;;;;;;;;;GAaG;AACH,wBAAsB,0BAA0B,CAAC,CAAC,EAChD,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,MAAM,CAAC,EAAE,WAAW,EACpB,OAAO,CAAC,EAAE,CAAC,aAAa,EAAE,MAAM,EAAE,GAAG,EAAE,OAAO,KAAK,IAAI,GACtD,OAAO,CAAC,CAAC,CAAC,CAWZ;AAkCD,qBAAa,WAAY,SAAQ,WAAW;IAC1C,OAAO,CAAC,WAAW,CAAS;IAE5B;;;;;;;;;;;;;OAaG;IACH,OAAO,CAAC,sBAAsB,CAA6B;IAE3D;;;;;;;;;OASG;IACH,OAAO,CAAC,mBAAmB,CAA6B;IAMxD;;;;;;OAMG;IACH,OAAO,CAAC,+BAA+B,CAAqB;IAE5D;;;;OAIG;IACH,OAAO,CAAC,SAAS,CAAmE;IAEpF;;;;;OAKG;IACH,OAAO,CAAC,mBAAmB,CAAS;IAEpC;;;;;;;;;OASG;IACG,UAAU,CACd,OAAO,EAAE,MAAM,EACf,IAAI,GAAE;QACJ,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,GAAG,CAAC,EAAE,OAAO,CAAC;QAAM,OAAO,CAAC,EAAE,KAAK,CAAC;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAA;SAAE,CAAC,CAAC;QACzE,OAAO,CAAC,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;QACjC;;;;;WAKG;QACH,YAAY,CAAC,EAAE,OAAO,CAAC;QACvB,+DAA+D;QAC/D,UAAU,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,CAAC;QACpC;;;;WAIG;QACH,UAAU,CAAC,EAAE,CAAC,KAAK,EAAE,SAAS,GAAG,QAAQ,EAAE,IAAI,EAAE,YAAY,KAAK,IAAI,CAAC;QACvE;;;;WAIG,CAAI,YAAY,CAAC,EAAE,CAAC,QAAQ,EAAE,OAAO,wBAAwB,EAAE,YAAY,KAAK,IAAI,CAAC;QACxF;;;;WAIG;QACH,SAAS,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,sBAAsB,EAAE,cAAc,KAAK,IAAI,CAAC;QAC7E,4EAA4E;QAC5E,YAAY,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,wBAAwB,EAAE,iBAAiB,KAAK,IAAI,CAAC;QACrF;;;WAGG;QACH,SAAS,CAAC,EAAE,OAAO,wBAAwB,EAAE,aAAa,CAAC;QAC3D,iGAAiG;QACjG,OAAO,CAAC,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;QACjC;;;;;;WAMG;QACH,cAAc,CAAC,EAAE,MAAM,CAAC;QACxB;;;;;;;WAOG;QACH,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB;;;;;WAKG;QACH,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;QAClC;;;;;WAKG;QACH,MAAM,CAAC,EAAE,WAAW,CAAC;KACjB,GACL,OAAO,CAAC;QACT,OAAO,EAAE,MAAM,CAAC;QAChB,SAAS,EAAE,kBAAkB,EAAE,CAAC;QAChC,gBAAgB,CAAC,EAAE,OAAO,CAAC;QAC3B,yEAAyE;QACzE,SAAS,CAAC,EAAE,OAAO,CAAC;QACpB,0EAA0E;QAC1E,OAAO,CAAC,EAAE,OAAO,CAAC;QAClB,4EAA4E;QAC5E,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;QACrB;;;WAGG;QACH,qBAAqB,CAAC,EAAE,OAAO,CAAC;QAChC;;;WAGG;QACH,kBAAkB,CAAC,EAAE,OAAO,CAAC;QAC7B,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,KAAK,CAAC,EAAE,MAAM,CAAC;KAChB,CAAC;IAkHA,MAAM,IAAI,OAAO;YAgBH,OAAO;IAkXrB;;;;;;;;;;;;;OAaG;YACW,aAAa;IA6d3B;;;;;;OAMG;IACH;;;;;;;;;;;;OAYG;IACH,OAAO,CAAC,aAAa;IAUrB,OAAO,CAAC,kBAAkB;IAmX1B;;;;OAIG;YACW,eAAe;IAkC7B;;;;;OAKG;IACH,OAAO,CAAC,cAAc;IAUtB;;;;;;;;;;;;;;;OAeG;IACH;;;;;;;;;;;OAWG;IACH,OAAO,CAAC,yBAAyB;YAgBnB,eAAe;IAI7B;;;;OAIG;IACH;;;;;;;OAOG;YACW,gBAAgB;IAwP9B;;;;;;;;OAQG;IACH,OAAO,CAAC,kBAAkB;YAmFZ,aAAa;CAwG5B"}
+{"version":3,"file":"chat.d.ts","sourceRoot":"","sources":["../../src/cli/chat.ts"],"names":[],"mappings":"AAIA,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAEpC,OAAO,EAAE,WAAW,EAAc,MAAM,eAAe,CAAC;AAiCxD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAC;AAqBtD;;;;;GAKG;AACH,MAAM,WAAW,YAAY;IAC3B,EAAE,CAAC,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;IACb,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAC/B,EAAE,CAAC,EAAE,OAAO,CAAC;IACb,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,MAAM,CAAC;IACf,UAAU,CAAC,EAAE,MAAM,CAAC;CACrB;AAED,OAAO,EAA+B,KAAK,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACrF,OAAO,EAGL,KAAK,kBAAkB,EACxB,MAAM,4BAA4B,CAAC;AA0JpC,sEAAsE;AACtE,MAAM,WAAW,wBAAwB;IACvC,oDAAoD;IACpD,QAAQ,EAAE,OAAO,CAAC;IAClB,0EAA0E;IAC1E,WAAW,EAAE,OAAO,CAAC;CACtB;AAED,oEAAoE;AACpE,MAAM,WAAW,yBAAyB;IACxC,GAAG,CAAC,EAAE,OAAO,CAAC;IACd,0EAA0E;IAC1E,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,wBAAgB,uBAAuB,CACrC,MAAM,EAAE,aAAa,EACrB,IAAI,CAAC,EAAE,yBAAyB,GAC/B,wBAAwB,CA4B1B;AAID;;;;;;;;GAQG;AACH,wBAAsB,gBAAgB,CACpC,IAAI,EAAE,MAAM,EACZ,aAAa,EAAE,GAAG,EAClB,OAAO,CAAC,EAAE;IAAE,QAAQ,CAAC,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,GAC9C,OAAO,CAAC,IAAI,CAAC,CAYf;AAED;;;;;;;;;GASG;AACH;;;;;GAKG;AACH,eAAO,MAAM,yBAAyB,uBAA0B,CAAC;AAEjE;;;;;;;;;;;;;GAaG;AACH,wBAAsB,0BAA0B,CAAC,CAAC,EAChD,OAAO,EAAE,MAAM,OAAO,CAAC,CAAC,CAAC,EACzB,MAAM,CAAC,EAAE,WAAW,EACpB,OAAO,CAAC,EAAE,CAAC,aAAa,EAAE,MAAM,EAAE,GAAG,EAAE,OAAO,KAAK,IAAI,GACtD,OAAO,CAAC,CAAC,CAAC,CAWZ;AAkCD,qBAAa,WAAY,SAAQ,WAAW;IAC1C,OAAO,CAAC,WAAW,CAAS;IAE5B;;;;;;;;;;;;;OAaG;IACH,OAAO,CAAC,sBAAsB,CAA6B;IAE3D;;;;;;;;;OASG;IACH,OAAO,CAAC,mBAAmB,CAA6B;IAMxD;;;;;;OAMG;IACH,OAAO,CAAC,+BAA+B,CAAqB;IAE5D;;;;OAIG;IACH,OAAO,CAAC,SAAS,CAAmE;IAEpF;;;;;OAKG;IACH,OAAO,CAAC,mBAAmB,CAAS;IAEpC;;;;;;;OAOG;IACH,OAAO,CAAC,iBAAiB,CAAC,CAAgE;IAE1F;;;;;OAKG;IACH,OAAO,CAAC,WAAW,CAAC,CAAuC;IAE3D;;;;;;;;;OASG;IACG,UAAU,CACd,OAAO,EAAE,MAAM,EACf,IAAI,GAAE;QACJ,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,GAAG,CAAC,EAAE,OAAO,CAAC;QAAM,OAAO,CAAC,EAAE,KAAK,CAAC;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,OAAO,EAAE,MAAM,CAAA;SAAE,CAAC,CAAC;QACzE,OAAO,CAAC,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;QACjC;;;;;WAKG;QACH,YAAY,CAAC,EAAE,OAAO,CAAC;QACvB,+DAA+D;QAC/D,UAAU,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,CAAC;QACpC;;;;WAIG;QACH,UAAU,CAAC,EAAE,CAAC,KAAK,EAAE,SAAS,GAAG,QAAQ,EAAE,IAAI,EAAE,YAAY,KAAK,IAAI,CAAC;QACvE;;;;WAIG,CAAI,YAAY,CAAC,EAAE,CAAC,QAAQ,EAAE,OAAO,wBAAwB,EAAE,YAAY,KAAK,IAAI,CAAC;QACxF;;;;WAIG;QACH,SAAS,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,sBAAsB,EAAE,cAAc,KAAK,IAAI,CAAC;QAC7E,4EAA4E;QAC5E,YAAY,CAAC,EAAE,CAAC,OAAO,EAAE,OAAO,wBAAwB,EAAE,iBAAiB,KAAK,IAAI,CAAC;QACrF;;;WAGG;QACH,SAAS,CAAC,EAAE,OAAO,wBAAwB,EAAE,aAAa,CAAC;QAC3D,iGAAiG;QACjG,OAAO,CAAC,EAAE,WAAW,CAAC,SAAS,CAAC,CAAC;QACjC;;;;;;WAMG;QACH,cAAc,CAAC,EAAE,MAAM,CAAC;QACxB;;;;;;;;;WASG;QACH,YAAY,CAAC,EAAE,MAAM,CAAC;QACtB;;;;;;;WAOG;QACH,WAAW,CAAC,EAAE,MAAM,CAAC;QACrB;;;;;WAKG;QACH,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;QAClC;;;;;WAKG;QACH,MAAM,CAAC,EAAE,WAAW,CAAC;KACjB,GACL,OAAO,CAAC;QACT,OAAO,EAAE,MAAM,CAAC;QAChB,SAAS,EAAE,kBAAkB,EAAE,CAAC;QAChC,gBAAgB,CAAC,EAAE,OAAO,CAAC;QAC3B,yEAAyE;QACzE,SAAS,CAAC,EAAE,OAAO,CAAC;QACpB,0EAA0E;QAC1E,OAAO,CAAC,EAAE,OAAO,CAAC;QAClB,4EAA4E;QAC5E,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;QACrB;;;WAGG;QACH,qBAAqB,CAAC,EAAE,OAAO,CAAC;QAChC;;;WAGG;QACH,kBAAkB,CAAC,EAAE,OAAO,CAAC;QAC7B;;;;WAIG;QACH,mBAAmB,CAAC,EAAE,OAAO,CAAC;QAC9B,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,KAAK,CAAC,EAAE,MAAM,CAAC;KAChB,CAAC;IAmHA,MAAM,IAAI,OAAO;YAgBH,OAAO;IAkXrB;;;;;;;;;;;;;OAaG;YACW,aAAa;IA0lB3B;;;;;;OAMG;IACH;;;;;;;;;;;;OAYG;IACH,OAAO,CAAC,aAAa;IAUrB,OAAO,CAAC,kBAAkB;IAsX1B;;;;OAIG;YACW,eAAe;IAkC7B;;;;;OAKG;IACH,OAAO,CAAC,cAAc;IAUtB;;;;;;;;;;;;;;;OAeG;IACH;;;;;;;;;;;OAWG;IACH,OAAO,CAAC,yBAAyB;YAgBnB,eAAe;IAI7B;;;;OAIG;IACH;;;;;;;OAOG;YACW,gBAAgB;IA4Q9B;;;;;;;;OAQG;IACH,OAAO,CAAC,kBAAkB;YAmFZ,aAAa;CAwG5B"}

package/dist/cli/chat.js

@@ -23,6 +23,7 @@ import { getAutoRouter, isAutoModel, isAutoProvider, governanceVerdict } from '.
 import { estimateTokens } from '../learning/cost-tracker.js';
 import { getModelRegistry } from '../learning/model-registry.js';
 import { refreshModelRegistry } from '../inference/model-probe.js';
+import { startWarmupDaemon } from '../learning/model-warmup.js';
 import { recordRoutingDecision } from '../learning/routing-history.js';
 import { shouldConfirmFailover, promptFailoverChoice } from './failover-prompt.js';
 import { buildAutoResolveOptions } from '../learning/resolve-options.js';
@@ -35,7 +36,8 @@ import { resolveDispatch } from '../nlu/actions.js';
 import { hasCodingAction, resolveAskKind } from '../nlu/conversation-gate.js';
 import { runToolLoop, extractFallbackToolCalls } from '../tools/tool-loop.js';
 import { detectAnswerQualityFailure, answerQualityError, toUserFacingGenerationError, isToolCallingUnsupported, stripToolCallArtifacts, } from '../inference/tool-call-utils.js';
-import { beginTrace, endTrace, recordStep, buildTraceOutcome } from '../learning/reasoning-trace.js';
+import { beginTrace, endTrace, recordStep, recordTraceEvent, buildTraceOutcome } from '../learning/reasoning-trace.js';
+import { recordWorkingState, getWorkingState, formatWorkingState } from '../learning/working-state.js';
 import { getLoopExposureMode } from '../tools/toolsets.js';
 import { resolveModelHarnessProfile, shouldSkipNativeTools } from '../learning/model-harness.js';
 import { resolveAdapterDefault, hasCredentials } from '../learning/model-selection.js';
@@ -347,6 +349,22 @@ export class ChatCommand extends BaseCommand {
      * failing into dead ends — the fire-and-forget keeps the first message fast.
      */
     coldStartProbeFired = false;
+    /**
+     * G5 — TRACE FIDELITY. The Auto-router decision snapshot for the current
+     * turn, stashed at resolve time so the reasoning trace can record WHY a step
+     * used the provider/model it did. The audit of the calculator session found
+     * 98/99 steps stamped with the session default `gemini/gemini-3.1-flash-lite`
+     * and **0/99** steps carrying a routing snapshot — so the Trace tab's "model
+     * used" column was really "the session's default model".
+     */
+    lastRouteSnapshot;
+    /**
+     * G5 — the provider/model of the LAST generation attempt (post-failover,
+     * post-default-resolution). Read by the chat trace recorder so a step names
+     * the model that actually ran it, instead of whatever the session default
+     * happened to be.
+     */
+    lastAttempt;
     /**
      * P3 — programmatic single-turn answer for the dashboard chat console.
      *
@@ -429,7 +447,7 @@ export class ChatCommand extends BaseCommand {
         }
         const parsed = parseRequestSync(message);
         const dispatchDecision = resolvePipelineDispatch(parsed, { dev: opts.dev, text: message });
-        const answer = await this.runChatAnswer(message, opts.history ?? [], { type, provider, model }, { provider: mergedOpts.provider, model: mergedOpts.model, dev: mergedOpts.dev, cache: true }, true, { auto: autoMode }, parsed, { askUser: opts.askUser, onProgress: opts.onProgress, onToolCall: opts.onToolCall, onPlanChange: opts.onPlanChange, onGitDiff: opts.onGitDiff, onSkillDraft: opts.onSkillDraft, planStore: opts.planStore ?? this.planStore, gateway: opts.gateway, projectContext: opts.projectContext, recallContext: recallBlock, projectPath: opts.projectPath, onToken: opts.onToken, signal: opts.signal, continuation: opts.continuation });
+        const answer = await this.runChatAnswer(message, opts.history ?? [], { type, provider, model }, { provider: mergedOpts.provider, model: mergedOpts.model, dev: mergedOpts.dev, cache: true }, true, { auto: autoMode }, parsed, { askUser: opts.askUser, onProgress: opts.onProgress, onToolCall: opts.onToolCall, onPlanChange: opts.onPlanChange, onGitDiff: opts.onGitDiff, onSkillDraft: opts.onSkillDraft, planStore: opts.planStore ?? this.planStore, gateway: opts.gateway, projectContext: opts.projectContext, recallContext: recallBlock, projectPath: opts.projectPath, onToken: opts.onToken, signal: opts.signal, continuation: opts.continuation, systemPolicy: opts.systemPolicy });
         // No-model fallback: the tool loop could not generate a single response
         // AND the rules assessed a high-confidence pipeline intent — run the
         // pipeline directly (rules decide only when the model is unavailable; the
@@ -456,6 +474,7 @@ export class ChatCommand extends BaseCommand {
             toolCalls: answer.toolCalls,
             unverifiedActionClaim: answer.unverifiedActionClaim,
             unfulfilledPromise: answer.unfulfilledPromise,
+            undeliveredArtifact: answer.undeliveredArtifact,
             provider: type,
             model,
         };
@@ -884,8 +903,20 @@ export class ChatCommand extends BaseCommand {
                 ambientProjectContext = undefined;
             }
         }
+        // G4 — carry THIS project's working state (files changed, verification debt,
+        // user-reported regressions) into the turn, so the model does not re-derive
+        // what previous turns already established. This is the fix for the
+        // calculator session's core drift (it re-diagnosed the same root cause six
+        // times, then undid its own earlier fixes).
+        const workingStatePath = ctxOverrides?.projectPath || process.cwd();
+        const workingStateBlock = formatWorkingState(getWorkingState(workingStatePath));
+        // Session 3 — channel/format policy lives in the STABLE layer. It is
+        // identical on every message, so keeping it here makes the system prompt
+        // byte-stable (prompt-cacheable) AND removes it from the volatile user
+        // turn, where it used to occupy ~80% of the ask on WhatsApp turns.
+        const systemPolicyBlock = ctxOverrides?.systemPolicy ? `\n\n${ctxOverrides.systemPolicy}` : '';
         const thread = [
-            { role: 'system', content: systemText + skillHint },
+            { role: 'system', content: systemText + systemPolicyBlock + skillHint },
             // P3 — the attached project's bounded snapshot (path + file tree +
             // symbol map) rides in before the conversation, exactly like --file
             // context: the model knows what it is looking at without being told.
@@ -901,6 +932,8 @@ export class ChatCommand extends BaseCommand {
             ...(ctxOverrides?.recallContext
                 ? [{ role: 'user', content: ctxOverrides.recallContext }]
                 : []),
+            // G4 — the deterministic working-state ledger (never a summary).
+            ...(workingStateBlock ? [{ role: 'user', content: workingStateBlock }] : []),
             ...(fileContext
                 ? [{ role: 'user', content: `[File context]\n${fileContext}` }]
                 : []),
@@ -925,6 +958,16 @@ export class ChatCommand extends BaseCommand {
         // loader channel the tool writes and the loop reads. 'all' (default)
         // keeps the pre-tiering behavior byte-identical.
         const loadedExtraTools = new Set();
+        // G3 — the files this turn mutates, observed on the tool event stream so
+        // the ledger can remember them (the loop reports tool NAMES, not paths).
+        const touchedFiles = new Set();
+        /**
+         * G18 — the turn's trace id, for the autonomy-gate events that tools emit
+         * DURING the loop. Assigned a few lines below (the trace begins once the
+         * thread and tool surface exist); the emit can only fire from tool
+         * execution, which is strictly after that assignment.
+         */
+        let traceIdForEvents;
         const toolContext = {
             configManager: this.configManager,
             loadedExtraTools,
@@ -932,6 +975,16 @@ export class ChatCommand extends BaseCommand {
             // agent operates inside the project (not the dashboard server's cwd).
             cwd: ctxOverrides?.projectPath || process.cwd(),
             emit: (event, data, source) => {
+                // G3 — collect mutated file paths from `tool:started` (which carries
+                // the arguments) so the working-state ledger knows what changed.
+                if (event === 'tool:started') {
+                    const t = data;
+                    if (t?.tool === 'edit_file' || t?.tool === 'write_file') {
+                        const p = t.args?.path ?? t.args?.file_path ?? t.args?.file;
+                        if (typeof p === 'string' && p)
+                            touchedFiles.add(p);
+                    }
+                }
                 // P0.6 — forward tool-call lifecycle events to the GUI before they
                 // reach the bus (the bus drives hooks; the override drives the card
                 // stream). Other events keep flowing to the bus untouched.
@@ -951,6 +1004,21 @@ export class ChatCommand extends BaseCommand {
                 if (ctxOverrides?.onSkillDraft && event === 'skill:draft') {
                     ctxOverrides.onSkillDraft(data);
                 }
+                // G18 — an autonomy gate DECIDING to proceed is a fact about the turn
+                // ("this change was applied without asking, and here is why"), not just
+                // a bus notification: record it on the turn's trace so the decision is
+                // auditable after the fact.
+                if (event === 'autonomy:write-applied' && traceIdForEvents) {
+                    const d = data;
+                    recordTraceEvent(traceIdForEvents, {
+                        kind: 'gate',
+                        gate: 'autonomy',
+                        ...(d?.tool ? { tool: d.tool } : {}),
+                        summary: d?.reason
+                            ? `proceeded without asking — ${d.reason}`
+                            : 'proceeded without asking — the request itself was the authorization',
+                    });
+                }
                 getEventBus().emit(event, data, source);
             },
             // P3 — the dashboard chat console injects a NON-TTY ask_user renderer
@@ -991,6 +1059,8 @@ export class ChatCommand extends BaseCommand {
             provider: session.type,
             model: session.model,
         });
+        // G18 — the tool-context emit (declared above) now has somewhere to write.
+        traceIdForEvents = chatTraceId;
         const seenStepDigests = new Set();
         const digestPrompt = (p) => {
             try {
@@ -1013,20 +1083,28 @@ export class ChatCommand extends BaseCommand {
                 const promptPreview = lastUserIdx !== -1
                     ? `${prompt.slice(0, 80)}…\n\n${prompt.slice(lastUserIdx)}`.slice(0, 500)
                     : prompt.slice(0, 300);
+                // G5 — the REAL per-call model/provider (post-failover) and the
+                // Auto-router snapshot, plus estimateTokens (the old chars/4 estimate
+                // misreported usage). Falls back to the session default only when no
+                // attempt was recorded (e.g. a cached/offline step).
                 recordStep(chatTraceId, {
                     agentType: 'chat',
                     description: message.slice(0, 120),
-                    provider: session.type,
-                    model: session.model ?? 'unknown',
+                    // Session 3 — the FULL prompt feeds the layered digests and the
+                    // one-time stable-layer capture (see `promptFull` in recordStep).
+                    promptFull: prompt,
+                    provider: this.lastAttempt?.provider || session.type,
+                    model: this.lastAttempt?.model || session.model || 'unknown',
                     promptDigest: digest,
                     promptPreview,
                     responsePreview: output.slice(0, 1000),
                     responseLength: output.length,
-                    inputTokens: Math.ceil(prompt.length / 4),
-                    outputTokens: Math.ceil(output.length / 4),
+                    inputTokens: estimateTokens(prompt),
+                    outputTokens: estimateTokens(output),
                     latencyMs,
                     success: ok,
                     error,
+                    ...(mode.auto && this.lastRouteSnapshot ? { routing: this.lastRouteSnapshot } : {}),
                 });
             }
             catch {
@@ -1069,6 +1147,10 @@ export class ChatCommand extends BaseCommand {
                 maxParallelReads: harness.maxParallelReads,
                 onToken: ctxOverrides?.onToken,
                 signal: ctxOverrides?.signal,
+                // G18 — the same sink the execute loop uses: tool calls, gate decisions
+                // and refusals land on the turn's trace, so the chat surface can answer
+                // "what did it actually run, and what did it decline?" from evidence.
+                onTraceEvent: (event) => recordTraceEvent(chatTraceId, event),
                 deps: {
                     callModel: callModelWithTrace,
                     executeTool: async (name, args, ctx) => {
@@ -1129,7 +1211,55 @@ export class ChatCommand extends BaseCommand {
             tools: result.successfulToolCalls ?? result.toolCalls,
             unverifiedActionClaim: result.unverifiedActionClaim,
             unfulfilledPromise: result.unfulfilledPromise,
+            unverifiedEdit: result.unverifiedEdit,
+            unverifiedEditClaim: result.unverifiedEditClaim,
+            undeliveredArtifact: result.undeliveredArtifact,
         }));
+        // G3 — record what this turn actually did so the NEXT turn starts from it
+        // (files changed, whether anything verified the work, and whether the user
+        // reported a regression). Best-effort: the ledger must never break a turn.
+        try {
+            const activity = result.successfulToolCalls ?? [];
+            const verified = activity.some((t) => t === 'run_terminal' || t === 'test' || t === 'browser' || t === 'run_cli');
+            recordWorkingState(workingStatePath, {
+                filesTouched: [...touchedFiles],
+                toolsUsed: activity,
+                verified,
+                unverifiedEdit: result.unverifiedEdit === true,
+                userMessage: message,
+            });
+        }
+        catch {
+            // Best-effort.
+        }
+        // G1 + G2 — surface the unverified-edit warning ON THE CONSOLE too (the
+        // trace badge alone is invisible to a CLI/gateway user). Printed, never
+        // appended to `content`, so the delivered answer, the gateway bubble and
+        // the answer cache all stay clean.
+        try {
+            if (result.unverifiedEditClaim) {
+                logger.warn('   ⚠️  This reply asserts a code change, but NOTHING verified it (no test / typecheck / build / browser run). Treat the change as UNVERIFIED.');
+            }
+            else if (result.unverifiedEdit) {
+                logger.warn('   ⚠️  Files were changed this turn but no verification ran — the change is unverified.');
+            }
+        }
+        catch {
+            // Best-effort — a warning must never break the turn.
+        }
+        // G13b — the DELIVERABLE warning, on the console for the same reason: the
+        // trace badge is invisible to a CLI user, and this failure is the one that
+        // looks most like success. The reply reads as a finished 12-page story while
+        // no file exists, so the reader is told plainly that the deliverable is
+        // missing rather than left to discover it when the path is not there.
+        try {
+            if (result.undeliveredArtifact) {
+                logger.warn('   ⚠️  This request asked for a written deliverable, but NO file was written this turn — the text above is the answer, not the artifact.');
+            }
+        }
+        catch {
+            // Best-effort.
+        }
         // Finalize the turn (cache + memory + registry telemetry).
         // E3c: a generationFailed turn is NOT cached/persisted — the caller may
         // fall back to the rule decision, and the failure text must not pollute
@@ -1175,6 +1305,7 @@ export class ChatCommand extends BaseCommand {
             toolCalls: result.toolCalls,
             unverifiedActionClaim: result.unverifiedActionClaim,
             unfulfilledPromise: result.unfulfilledPromise,
+            undeliveredArtifact: result.undeliveredArtifact,
         };
     }
     /**
@@ -1287,6 +1418,9 @@ export class ChatCommand extends BaseCommand {
                 // telemetry name the model that failed, instead of "unknown".
                 if (effectiveModel)
                     session.model = effectiveModel;
+                // G5 — and name the attempt itself, so the trace recorder reports the
+                // real provider/model (post-failover) rather than the session default.
+                this.lastAttempt = { provider: typ, model: effectiveModel };
                 const nativeKey = `${typ}|${effectiveModel ?? ''}`;
                 // Answer-quality resilience: a CONFUSED reply — the model talking about
                 // the tool contract (e.g. apologizing that "the provided example call
@@ -1738,6 +1872,15 @@ export class ChatCommand extends BaseCommand {
             circuitBreakerStatus,
             ...(dispatch.taskIntentHint ? { taskIntentHint: dispatch.taskIntentHint } : {}),
         }, this.configManager));
+        // G5 — record the routing snapshot now (the winner + why), so the reasoning
+        // trace for this turn can show the decision instead of a blank column.
+        this.lastRouteSnapshot = {
+            provider: decision.provider,
+            model: decision.model,
+            score: decision.score,
+            complexity: String(decision.complexity),
+            explanation: decision.explanation,
+        };
         // Walk the ranked candidates (winner first) and return the first available
         // provider — never a provider that lacks a key or endpoint. Providers that
         // already failed this message (excludeProviders) OR earlier in this session
@@ -1750,6 +1893,17 @@ export class ChatCommand extends BaseCommand {
             return expiresAt !== undefined && expiresAt > exclusionTime;
         };
         // ── Cold-start probe (suggestion 3) ─────────────────────────────────────
+        // Start the background warmup/exploration daemon on EVERY chat session, not
+        // only a cold one (Models-page audit). It is idempotent (already-running is
+        // a no-op) and unref'd, so it cannot hold the process open — and it is the
+        // only thing that verifies models the router cannot see yet. Left cold-only,
+        // the verified pool could only ever shrink as staleness retired models.
+        try {
+            startWarmupDaemon(this.configManager);
+        }
+        catch {
+            // Best-effort — warmup must never break chat.
+        }
         // A fresh registry has zero verified models → routing would fall back to
         // credential-based defaults and possibly fail into dead ends. Fire ONE
         // background probe+spot-check so the registry learns from real API data.