dsh-llm-verifier 0.6.0 → 0.6.2

package/lib/types/verifier-activity.d.ts

@@ -0,0 +1,189 @@
+/**
+ * In-memory view of the verifier's in-flight cycles, for the chat chip.
+ *
+ * WHY THIS EXISTS
+ * ---------------
+ * Every automatic stage runs judge calls while the chat shows nothing about it. Two of them are
+ * silent windows rather than "no output at all":
+ *
+ * - P06 buffers the main reply, so while a cycle runs there is no streaming text whatsoever —
+ *   indistinguishable from a hang;
+ * - the routed reviews (compare/select/track) and the final acceptance run inside
+ *   `agent/turn-stopping`, where the reply is already on screen but the turn stays open for as long
+ *   as the judges take (a bounded but real pause, and a passing acceptance says nothing at all).
+ *
+ * The harness renders the host's own inline indicator ("Compacting context…") from a *session event*
+ * pair, and that route is closed to an out-of-repo plugin: the persistence read path
+ * (`validateStoredEvents` in `@deepseek-ai/dsh-session-persistence`) refuses to interpret a log
+ * containing an event type outside the repository-generated `KNOWN_SESSION_EVENT_TYPES` unless the
+ * stored envelope carries `ignorable: true`, and `Session.append` cannot set that marker — only a
+ * seed/restore envelope can. Appending our own type would therefore make the topic unloadable.
+ *
+ * So the indicator is deliberately host-only state: the router (the one place that knows a cycle was
+ * granted) and the P06 selector publish what they are doing here, the plugin's existing RPC answers
+ * a read of it, and the client polls that read only while a turn runs. Nothing here is durable (a
+ * reload shows nothing, which is correct: the cycle it described is gone) and nothing here touches
+ * the model history, the session log, the statistics or any verdict.
+ *
+ * Pure bookkeeping: no I/O, no clock of its own (callers pass `now`), so it is directly testable.
+ */
+import type { Reservation, RouterCycleObserver } from './router.ts';
+/** Which part of the plugin owns a cycle. */
+export type ActivityStage = 'process' | 'route' | 'final';
+/** What a cycle is doing right now, as far as the UI can tell it apart. */
+export type ActivityPhase = 'generating' | 'comparing' | 'classifying' | 'reviewing' | 'accepting';
+/** The routed decision a `route` cycle resolved to, once classification promoted it. */
+export type ActivityDestination = 'compare' | 'select' | 'track';
+/** How a finished cycle resolved, from the user's point of view. */
+export type ActivityOutcome = 'replaced' | 'kept' | 'same' | 'failed' | 'accepted' | 'rejected';
+/** One cycle in flight. */
+export interface ActivityActive {
+    cycleId: string;
+    stage: ActivityStage;
+    phase: ActivityPhase;
+    /** Routing target of a `route` cycle; absent while it is still classifying and for the others. */
+    destination?: ActivityDestination;
+    /** Candidates in the P06 cycle, the original reply included (2..4). */
+    candidates?: number;
+    /** Model calls this cycle reserved, for the copy that quotes a size. */
+    expectedCalls?: number;
+    /** `provider/model` of P06's second generator, when `autoProcessAlternativeModel` is set. */
+    alternativeModel?: string;
+    startedAt: number;
+}
+/** One cycle that just finished; kept only long enough for the chip to explain the pause. */
+export interface ActivitySettled {
+    cycleId: string;
+    stage: ActivityStage;
+    outcome: ActivityOutcome;
+    candidates?: number;
+    /** Wall clock the cycle settled at. */
+    at: number;
+}
+/** What one session currently has to show. Both fields absent means "nothing to say". */
+export interface ActivityView {
+    active?: ActivityActive;
+    settled?: ActivitySettled;
+}
+/**
+ * How long a settled cycle keeps explaining itself.
+ *
+ * Long enough to be read after the reply lands (the chat scrolls as soon as the winner replays),
+ * short enough that a stale line never becomes part of the transcript.
+ */
+export declare const ACTIVITY_SETTLED_TTL_MS = 20000;
+/**
+ * Safety bound on an in-flight record.
+ *
+ * Every cycle is bounded by the configured phase timeout and settles on every path we know of, so
+ * this only ever fires for a cycle that leaked (a bug, or a host that never calls back). Without it
+ * a leaked record would pin the chip in "working…" forever, because the client keeps polling while
+ * it has something to render.
+ */
+export declare const ACTIVITY_ACTIVE_TTL_MS = 600000;
+/**
+ * Map one finished P06 cycle onto the outcome the chip distinguishes.
+ *
+ * `replayed` is the host-facing fact ("which stream did the host actually receive") and wins over
+ * the outcome label: a cycle recorded as `candidate-selected` whose winner was withheld before the
+ * first chunk is corrected to `replayed: 'original'` and must read as "kept", not as a replacement.
+ * @param outcome - the durable outcome string the cycle was recorded with.
+ * @param replayed - which stream the host actually received.
+ * @returns The chip-level outcome.
+ */
+export declare function classifyProcessOutcome(outcome: string, replayed: 'original' | 'candidate' | 'none'): ActivityOutcome;
+/**
+ * Map one granted reservation onto the chip's active record.
+ *
+ * `process` is deliberately NOT mapped: P06 publishes its own richer record (candidate count,
+ * generation vs comparison), and two writers for one cycle id would race. `plan_review` and
+ * `team_task` are not announced either — the plan gate already shows a pending tool row in the
+ * chat, so a chip line would only repeat it.
+ * @param reservation - the reservation the router just granted.
+ * @param startedAt - wall clock the cycle began at.
+ * @returns The active record, or undefined for a phase the chip stays silent about.
+ */
+export declare function reservationActivity(reservation: Pick<Reservation, 'id' | 'phase' | 'expectedCalls'>, startedAt: number): ActivityActive | undefined;
+/**
+ * Map a promotion onto the update the chip needs.
+ *
+ * A classification cycle that resolved to a decision stays the SAME reservation (that is what
+ * promotion is for), so the chip must be moved from "classifying" to "reviewing" instead of leaving
+ * a second cycle behind.
+ * @param reservation - the promoted reservation, already carrying its execution phase.
+ * @returns The patch, or undefined when the phase is not one the chip announces.
+ */
+export declare function reservationPromotion(reservation: Pick<Reservation, 'phase' | 'expectedCalls'>): Pick<ActivityActive, 'phase' | 'expectedCalls'> & {
+    destination?: ActivityDestination;
+} | undefined;
+/**
+ * Map a router settlement onto the chip's settled row.
+ *
+ * A routed verdict is steered into the chat by the caller right after the cycle settles, so a second
+ * transient line would only repeat it; the final gate is the opposite (a PASSING acceptance steers
+ * nothing), and that is exactly the case worth showing. `undefined` means "clear the active row and
+ * leave nothing behind".
+ * @param reservation - the reservation being settled.
+ * @param outcome - whether the router committed or failed it.
+ * @returns The stage and outcome to keep, or undefined to clear silently.
+ */
+export declare function reservationSettled(reservation: Pick<Reservation, 'id' | 'phase'>, outcome: 'committed' | 'failed'): {
+    stage: ActivityStage;
+    outcome: ActivityOutcome;
+} | undefined;
+/**
+ * Build the router's advisory observer over one activity table.
+ *
+ * The router supplies the reservation (it is the only component that knows a cycle was granted and
+ * when it settled); this module owns what the chip says about it.
+ * @param activities - the table the plugin's RPC reads.
+ * @param now - wall clock used for the records and their TTLs.
+ * @returns The observer to hand to `new AutoVerifierRouter(...)`.
+ */
+export declare function createActivityObserver(activities: VerifierActivities, now?: () => number): RouterCycleObserver;
+/**
+ * Per-session activity table.
+ *
+ * One active cycle per session, because the router owns one in-flight reservation per agent at a
+ * time; a settled cycle replaces the previous settled one (the chip is a status line, not a history
+ * — the statistics sidecar is the history).
+ */
+export declare class VerifierActivities {
+    private readonly active;
+    private readonly settled;
+    /** Publish one cycle as in flight. */
+    begin(sessionId: string, activity: ActivityActive): void;
+    /**
+     * Move one in-flight cycle to a later phase.
+     *
+     * Guarded by cycle id: a cycle that already settled (or was replaced by a newer one) must not be
+     * moved back into "in flight" by a late async step.
+     * @returns whether the exact cycle was the active one.
+     */
+    update(sessionId: string, cycleId: string, patch: Partial<Pick<ActivityActive, 'phase' | 'destination' | 'candidates' | 'expectedCalls'>>): boolean;
+    /**
+     * Retire one cycle.
+     *
+     * A cycle nobody observed (the purchase itself failed, so `begin` never ran) leaves nothing
+     * behind: the chip must not claim a cycle that never started. A cycle that WAS observed stays
+     * correctable while its own settled record is the current one, which is what P06's late
+     * "the winner was not delivered" correction needs.
+     * @param sessionId - the session the cycle belongs to.
+     * @param cycleId - the cycle to retire.
+     * @param outcome - the settled row to keep, or undefined to clear without leaving one.
+     * @param now - wall clock used for the settled record.
+     * @returns whether the cycle was known and retired.
+     */
+    finish(sessionId: string, cycleId: string, outcome: ActivityOutcome | undefined, now: number): boolean;
+    /**
+     * What one session has to show right now, with expired records pruned in place.
+     * @param sessionId - the session to read.
+     * @param now - wall clock used for both TTLs.
+     */
+    read(sessionId: string, now: number): ActivityView;
+    /** Forget one session entirely (settings changed, task replaced, agent disposed). */
+    clear(sessionId: string): void;
+    /** Forget every session (settings change, shutdown). */
+    clearAll(): void;
+}
+//# sourceMappingURL=verifier-activity.d.ts.map

package/lib/types/verifier-activity.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"verifier-activity.d.ts","sourceRoot":"","sources":["../../src/verifier-activity.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AACH,OAAO,KAAK,EAAE,WAAW,EAAE,mBAAmB,EAAe,MAAM,aAAa,CAAA;AAEhF,6CAA6C;AAC7C,MAAM,MAAM,aAAa,GAAG,SAAS,GAAG,OAAO,GAAG,OAAO,CAAA;AAEzD,2EAA2E;AAC3E,MAAM,MAAM,aAAa,GAAG,YAAY,GAAG,WAAW,GAAG,aAAa,GAAG,WAAW,GAAG,WAAW,CAAA;AAElG,wFAAwF;AACxF,MAAM,MAAM,mBAAmB,GAAG,SAAS,GAAG,QAAQ,GAAG,OAAO,CAAA;AAEhE,oEAAoE;AACpE,MAAM,MAAM,eAAe,GAAG,UAAU,GAAG,MAAM,GAAG,MAAM,GAAG,QAAQ,GAAG,UAAU,GAAG,UAAU,CAAA;AAE/F,2BAA2B;AAC3B,MAAM,WAAW,cAAc;IAC7B,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,aAAa,CAAA;IACpB,KAAK,EAAE,aAAa,CAAA;IACpB,kGAAkG;IAClG,WAAW,CAAC,EAAE,mBAAmB,CAAA;IACjC,uEAAuE;IACvE,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,wEAAwE;IACxE,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,6FAA6F;IAC7F,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,6FAA6F;AAC7F,MAAM,WAAW,eAAe;IAC9B,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,aAAa,CAAA;IACpB,OAAO,EAAE,eAAe,CAAA;IACxB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,uCAAuC;IACvC,EAAE,EAAE,MAAM,CAAA;CACX;AAED,yFAAyF;AACzF,MAAM,WAAW,YAAY;IAC3B,MAAM,CAAC,EAAE,cAAc,CAAA;IACvB,OAAO,CAAC,EAAE,eAAe,CAAA;CAC1B;AAED;;;;;GAKG;AACH,eAAO,MAAM,uBAAuB,QAAS,CAAA;AAE7C;;;;;;;GAOG;AACH,eAAO,MAAM,sBAAsB,SAAU,CAAA;AAE7C;;;;;;;;;GASG;AACH,wBAAgB,sBAAsB,CACpC,OAAO,EAAE,MAAM,EACf,QAAQ,EAAE,UAAU,GAAG,WAAW,GAAG,MAAM,GAC1C,eAAe,CAMjB;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,mBAAmB,CACjC,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,IAAI,GAAG,OAAO,GAAG,eAAe,CAAC,EAChE,SAAS,EAAE,MAAM,GAChB,cAAc,GAAG,SAAS,CAa5B;AAED;;;;;;;;GAQG;AACH,wBAAgB,oBAAoB,CAClC,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,OAAO,GAAG,eAAe,CAAC,GACxD,IAAI,CAAC,cAAc,EAAE,OAAO,GAAG,eAAe,CAAC,GAAG;IAAE,WAAW,CAAC,EAAE,mBAAmB,CAAA;CAAE,GAAG,SAAS,CASrG;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,kBAAkB,CAChC,WAAW,EAAE,IAAI,CAAC,WAAW,EAAE,IAAI,GAAG,OAAO,CAAC,EAC9C,OAAO,EAAE,WAAW,GAAG,QAAQ,GAC9B;IAAE,KAAK,EAAE,aAAa,CAAC;IAAC,OAAO,EAAE,eAAe,CAAA;CAAE,GAAG,SAAS,CAIhE;AAED;;;;;;;;GAQG;AACH,wBAAgB,sBAAsB,CACpC,UAAU,EAAE,kBAAkB,EAC9B,GAAG,GAAE,MAAM,MAAiB,GAC3B,mBAAmB,CAgBrB;AAED;;;;;;GAMG;AACH,qBAAa,kBAAkB;IAC7B,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAoC;IAC3D,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAqC;IAE7D,sCAAsC;IACtC,KAAK,CAAC,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,cAAc,GAAG,IAAI;IAIxD;;;;;;OAMG;IACH,MAAM,CACJ,SAAS,EAAE,MAAM,EACjB,OAAO,EAAE,MAAM,EACf,KAAK,EAAE,OAAO,CAAC,IAAI,CAAC,cAAc,EAAE,OAAO,GAAG,aAAa,GAAG,YAAY,GAAG,eAAe,CAAC,CAAC,GAC7F,OAAO;IAOV;;;;;;;;;;;;OAYG;IACH,MAAM,CACJ,SAAS,EAAE,MAAM,EACjB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,eAAe,GAAG,SAAS,EACpC,GAAG,EAAE,MAAM,GACV,OAAO;IAwBV;;;;OAIG;IACH,IAAI,CAAC,SAAS,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,GAAG,YAAY;IAalD,qFAAqF;IACrF,KAAK,CAAC,SAAS,EAAE,MAAM,GAAG,IAAI;IAK9B,wDAAwD;IACxD,QAAQ,IAAI,IAAI;CAIjB"}

package/lib/types/verifier-activity.js

@@ -0,0 +1,223 @@
+/**
+ * How long a settled cycle keeps explaining itself.
+ *
+ * Long enough to be read after the reply lands (the chat scrolls as soon as the winner replays),
+ * short enough that a stale line never becomes part of the transcript.
+ */
+export const ACTIVITY_SETTLED_TTL_MS = 20_000;
+/**
+ * Safety bound on an in-flight record.
+ *
+ * Every cycle is bounded by the configured phase timeout and settles on every path we know of, so
+ * this only ever fires for a cycle that leaked (a bug, or a host that never calls back). Without it
+ * a leaked record would pin the chip in "working…" forever, because the client keeps polling while
+ * it has something to render.
+ */
+export const ACTIVITY_ACTIVE_TTL_MS = 600_000;
+/**
+ * Map one finished P06 cycle onto the outcome the chip distinguishes.
+ *
+ * `replayed` is the host-facing fact ("which stream did the host actually receive") and wins over
+ * the outcome label: a cycle recorded as `candidate-selected` whose winner was withheld before the
+ * first chunk is corrected to `replayed: 'original'` and must read as "kept", not as a replacement.
+ * @param outcome - the durable outcome string the cycle was recorded with.
+ * @param replayed - which stream the host actually received.
+ * @returns The chip-level outcome.
+ */
+export function classifyProcessOutcome(outcome, replayed) {
+    if (replayed === 'candidate')
+        return 'replaced';
+    if (outcome === 'identical-candidate')
+        return 'same';
+    if (outcome === 'tie' || outcome === 'original-selected')
+        return 'kept';
+    if (outcome.startsWith('candidate-not-delivered'))
+        return 'kept';
+    return 'failed';
+}
+/**
+ * Map one granted reservation onto the chip's active record.
+ *
+ * `process` is deliberately NOT mapped: P06 publishes its own richer record (candidate count,
+ * generation vs comparison), and two writers for one cycle id would race. `plan_review` and
+ * `team_task` are not announced either — the plan gate already shows a pending tool row in the
+ * chat, so a chip line would only repeat it.
+ * @param reservation - the reservation the router just granted.
+ * @param startedAt - wall clock the cycle began at.
+ * @returns The active record, or undefined for a phase the chip stays silent about.
+ */
+export function reservationActivity(reservation, startedAt) {
+    switch (reservation.phase) {
+        case 'semantic':
+            return { cycleId: reservation.id, stage: 'route', phase: 'classifying', expectedCalls: reservation.expectedCalls, startedAt };
+        case 'compare':
+        case 'select':
+        case 'track':
+            return { cycleId: reservation.id, stage: 'route', phase: 'reviewing', destination: reservation.phase, expectedCalls: reservation.expectedCalls, startedAt };
+        case 'final':
+            return { cycleId: reservation.id, stage: 'final', phase: 'accepting', expectedCalls: reservation.expectedCalls, startedAt };
+        default:
+            return undefined;
+    }
+}
+/**
+ * Map a promotion onto the update the chip needs.
+ *
+ * A classification cycle that resolved to a decision stays the SAME reservation (that is what
+ * promotion is for), so the chip must be moved from "classifying" to "reviewing" instead of leaving
+ * a second cycle behind.
+ * @param reservation - the promoted reservation, already carrying its execution phase.
+ * @returns The patch, or undefined when the phase is not one the chip announces.
+ */
+export function reservationPromotion(reservation) {
+    switch (reservation.phase) {
+        case 'compare':
+        case 'select':
+        case 'track':
+            return { phase: 'reviewing', destination: reservation.phase, expectedCalls: reservation.expectedCalls };
+        default:
+            return undefined;
+    }
+}
+/**
+ * Map a router settlement onto the chip's settled row.
+ *
+ * A routed verdict is steered into the chat by the caller right after the cycle settles, so a second
+ * transient line would only repeat it; the final gate is the opposite (a PASSING acceptance steers
+ * nothing), and that is exactly the case worth showing. `undefined` means "clear the active row and
+ * leave nothing behind".
+ * @param reservation - the reservation being settled.
+ * @param outcome - whether the router committed or failed it.
+ * @returns The stage and outcome to keep, or undefined to clear silently.
+ */
+export function reservationSettled(reservation, outcome) {
+    return reservation.phase === 'final'
+        ? { stage: 'final', outcome: outcome === 'committed' ? 'accepted' : 'rejected' }
+        : undefined;
+}
+/**
+ * Build the router's advisory observer over one activity table.
+ *
+ * The router supplies the reservation (it is the only component that knows a cycle was granted and
+ * when it settled); this module owns what the chip says about it.
+ * @param activities - the table the plugin's RPC reads.
+ * @param now - wall clock used for the records and their TTLs.
+ * @returns The observer to hand to `new AutoVerifierRouter(...)`.
+ */
+export function createActivityObserver(activities, now = Date.now) {
+    const key = (agent) => String(agent.id);
+    return {
+        begin: (agent, reservation) => {
+            const activity = reservationActivity(reservation, now());
+            if (activity !== undefined)
+                activities.begin(key(agent), activity);
+        },
+        promoted: (agent, reservation) => {
+            const patch = reservationPromotion(reservation);
+            if (patch !== undefined)
+                activities.update(key(agent), reservation.id, patch);
+        },
+        settled: (agent, reservation, outcome) => {
+            const settled = reservationSettled(reservation, outcome);
+            activities.finish(key(agent), reservation.id, settled?.outcome, now());
+        },
+    };
+}
+/**
+ * Per-session activity table.
+ *
+ * One active cycle per session, because the router owns one in-flight reservation per agent at a
+ * time; a settled cycle replaces the previous settled one (the chip is a status line, not a history
+ * — the statistics sidecar is the history).
+ */
+export class VerifierActivities {
+    active = new Map();
+    settled = new Map();
+    /** Publish one cycle as in flight. */
+    begin(sessionId, activity) {
+        this.active.set(sessionId, activity);
+    }
+    /**
+     * Move one in-flight cycle to a later phase.
+     *
+     * Guarded by cycle id: a cycle that already settled (or was replaced by a newer one) must not be
+     * moved back into "in flight" by a late async step.
+     * @returns whether the exact cycle was the active one.
+     */
+    update(sessionId, cycleId, patch) {
+        const current = this.active.get(sessionId);
+        if (current === undefined || current.cycleId !== cycleId)
+            return false;
+        this.active.set(sessionId, { ...current, ...patch });
+        return true;
+    }
+    /**
+     * Retire one cycle.
+     *
+     * A cycle nobody observed (the purchase itself failed, so `begin` never ran) leaves nothing
+     * behind: the chip must not claim a cycle that never started. A cycle that WAS observed stays
+     * correctable while its own settled record is the current one, which is what P06's late
+     * "the winner was not delivered" correction needs.
+     * @param sessionId - the session the cycle belongs to.
+     * @param cycleId - the cycle to retire.
+     * @param outcome - the settled row to keep, or undefined to clear without leaving one.
+     * @param now - wall clock used for the settled record.
+     * @returns whether the cycle was known and retired.
+     */
+    finish(sessionId, cycleId, outcome, now) {
+        const active = this.active.get(sessionId);
+        const settled = this.settled.get(sessionId);
+        const known = active?.cycleId === cycleId
+            ? active
+            : settled?.cycleId === cycleId
+                ? settled
+                : undefined;
+        if (known === undefined)
+            return false;
+        if (active?.cycleId === cycleId)
+            this.active.delete(sessionId);
+        if (outcome === undefined) {
+            if (settled?.cycleId === cycleId)
+                this.settled.delete(sessionId);
+            return true;
+        }
+        this.settled.set(sessionId, {
+            cycleId,
+            stage: known.stage,
+            outcome,
+            ...(known.candidates === undefined ? {} : { candidates: known.candidates }),
+            at: now,
+        });
+        return true;
+    }
+    /**
+     * What one session has to show right now, with expired records pruned in place.
+     * @param sessionId - the session to read.
+     * @param now - wall clock used for both TTLs.
+     */
+    read(sessionId, now) {
+        const active = this.active.get(sessionId);
+        if (active !== undefined && now - active.startedAt >= ACTIVITY_ACTIVE_TTL_MS)
+            this.active.delete(sessionId);
+        const settled = this.settled.get(sessionId);
+        if (settled !== undefined && now - settled.at >= ACTIVITY_SETTLED_TTL_MS)
+            this.settled.delete(sessionId);
+        const currentActive = this.active.get(sessionId);
+        const currentSettled = this.settled.get(sessionId);
+        return {
+            ...(currentActive === undefined ? {} : { active: currentActive }),
+            ...(currentSettled === undefined ? {} : { settled: currentSettled }),
+        };
+    }
+    /** Forget one session entirely (settings changed, task replaced, agent disposed). */
+    clear(sessionId) {
+        this.active.delete(sessionId);
+        this.settled.delete(sessionId);
+    }
+    /** Forget every session (settings change, shutdown). */
+    clearAll() {
+        this.active.clear();
+        this.settled.clear();
+    }
+}
+//# sourceMappingURL=verifier-activity.js.map

package/lib/types/verifier-activity.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"verifier-activity.js","sourceRoot":"","sources":["../../src/verifier-activity.ts"],"names":[],"mappings":"AA2EA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,uBAAuB,GAAG,MAAM,CAAA;AAE7C;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAG,OAAO,CAAA;AAE7C;;;;;;;;;GASG;AACH,MAAM,UAAU,sBAAsB,CACpC,OAAe,EACf,QAA2C;IAE3C,IAAI,QAAQ,KAAK,WAAW;QAAE,OAAO,UAAU,CAAA;IAC/C,IAAI,OAAO,KAAK,qBAAqB;QAAE,OAAO,MAAM,CAAA;IACpD,IAAI,OAAO,KAAK,KAAK,IAAI,OAAO,KAAK,mBAAmB;QAAE,OAAO,MAAM,CAAA;IACvE,IAAI,OAAO,CAAC,UAAU,CAAC,yBAAyB,CAAC;QAAE,OAAO,MAAM,CAAA;IAChE,OAAO,QAAQ,CAAA;AACjB,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,mBAAmB,CACjC,WAAgE,EAChE,SAAiB;IAEjB,QAAQ,WAAW,CAAC,KAAK,EAAE,CAAC;QAC1B,KAAK,UAAU;YACb,OAAO,EAAE,OAAO,EAAE,WAAW,CAAC,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,aAAa,EAAE,WAAW,CAAC,aAAa,EAAE,SAAS,EAAE,CAAA;QAC/H,KAAK,SAAS,CAAC;QACf,KAAK,QAAQ,CAAC;QACd,KAAK,OAAO;YACV,OAAO,EAAE,OAAO,EAAE,WAAW,CAAC,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,WAAW,EAAE,WAAW,CAAC,KAAK,EAAE,aAAa,EAAE,WAAW,CAAC,aAAa,EAAE,SAAS,EAAE,CAAA;QAC7J,KAAK,OAAO;YACV,OAAO,EAAE,OAAO,EAAE,WAAW,CAAC,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,aAAa,EAAE,WAAW,CAAC,aAAa,EAAE,SAAS,EAAE,CAAA;QAC7H;YACE,OAAO,SAAS,CAAA;IACpB,CAAC;AACH,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,oBAAoB,CAClC,WAAyD;IAEzD,QAAQ,WAAW,CAAC,KAAK,EAAE,CAAC;QAC1B,KAAK,SAAS,CAAC;QACf,KAAK,QAAQ,CAAC;QACd,KAAK,OAAO;YACV,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,WAAW,EAAE,WAAW,CAAC,KAAK,EAAE,aAAa,EAAE,WAAW,CAAC,aAAa,EAAE,CAAA;QACzG;YACE,OAAO,SAAS,CAAA;IACpB,CAAC;AACH,CAAC;AAED;;;;;;;;;;GAUG;AACH,MAAM,UAAU,kBAAkB,CAChC,WAA8C,EAC9C,OAA+B;IAE/B,OAAO,WAAW,CAAC,KAAK,KAAK,OAAO;QAClC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,OAAO,KAAK,WAAW,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,UAAU,EAAE;QAChF,CAAC,CAAC,SAAS,CAAA;AACf,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,sBAAsB,CACpC,UAA8B,EAC9B,MAAoB,IAAI,CAAC,GAAG;IAE5B,MAAM,GAAG,GAAG,CAAC,KAAkB,EAAU,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,CAAA;IAC5D,OAAO;QACL,KAAK,EAAE,CAAC,KAAK,EAAE,WAAW,EAAE,EAAE;YAC5B,MAAM,QAAQ,GAAG,mBAAmB,CAAC,WAAW,EAAE,GAAG,EAAE,CAAC,CAAA;YACxD,IAAI,QAAQ,KAAK,SAAS;gBAAE,UAAU,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAA;QACpE,CAAC;QACD,QAAQ,EAAE,CAAC,KAAK,EAAE,WAAW,EAAE,EAAE;YAC/B,MAAM,KAAK,GAAG,oBAAoB,CAAC,WAAW,CAAC,CAAA;YAC/C,IAAI,KAAK,KAAK,SAAS;gBAAE,UAAU,CAAC,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,WAAW,CAAC,EAAE,EAAE,KAAK,CAAC,CAAA;QAC/E,CAAC;QACD,OAAO,EAAE,CAAC,KAAK,EAAE,WAAW,EAAE,OAAO,EAAE,EAAE;YACvC,MAAM,OAAO,GAAG,kBAAkB,CAAC,WAAW,EAAE,OAAO,CAAC,CAAA;YACxD,UAAU,CAAC,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,WAAW,CAAC,EAAE,EAAE,OAAO,EAAE,OAAO,EAAE,GAAG,EAAE,CAAC,CAAA;QACxE,CAAC;KACF,CAAA;AACH,CAAC;AAED;;;;;;GAMG;AACH,MAAM,OAAO,kBAAkB;IACZ,MAAM,GAAG,IAAI,GAAG,EAA0B,CAAA;IAC1C,OAAO,GAAG,IAAI,GAAG,EAA2B,CAAA;IAE7D,sCAAsC;IACtC,KAAK,CAAC,SAAiB,EAAE,QAAwB;QAC/C,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,EAAE,QAAQ,CAAC,CAAA;IACtC,CAAC;IAED;;;;;;OAMG;IACH,MAAM,CACJ,SAAiB,EACjB,OAAe,EACf,KAA8F;QAE9F,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QAC1C,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,CAAC,OAAO,KAAK,OAAO;YAAE,OAAO,KAAK,CAAA;QACtE,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,EAAE,EAAE,GAAG,OAAO,EAAE,GAAG,KAAK,EAAE,CAAC,CAAA;QACpD,OAAO,IAAI,CAAA;IACb,CAAC;IAED;;;;;;;;;;;;OAYG;IACH,MAAM,CACJ,SAAiB,EACjB,OAAe,EACf,OAAoC,EACpC,GAAW;QAEX,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QACzC,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QAC3C,MAAM,KAAK,GAAG,MAAM,EAAE,OAAO,KAAK,OAAO;YACvC,CAAC,CAAC,MAAM;YACR,CAAC,CAAC,OAAO,EAAE,OAAO,KAAK,OAAO;gBAC5B,CAAC,CAAC,OAAO;gBACT,CAAC,CAAC,SAAS,CAAA;QACf,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,KAAK,CAAA;QACrC,IAAI,MAAM,EAAE,OAAO,KAAK,OAAO;YAAE,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;QAC9D,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC1B,IAAI,OAAO,EAAE,OAAO,KAAK,OAAO;gBAAE,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;YAChE,OAAO,IAAI,CAAA;QACb,CAAC;QACD,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,SAAS,EAAE;YAC1B,OAAO;YACP,KAAK,EAAE,KAAK,CAAC,KAAK;YAClB,OAAO;YACP,GAAG,CAAC,KAAK,CAAC,UAAU,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,KAAK,CAAC,UAAU,EAAE,CAAC;YAC3E,EAAE,EAAE,GAAG;SACR,CAAC,CAAA;QACF,OAAO,IAAI,CAAA;IACb,CAAC;IAED;;;;OAIG;IACH,IAAI,CAAC,SAAiB,EAAE,GAAW;QACjC,MAAM,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QACzC,IAAI,MAAM,KAAK,SAAS,IAAI,GAAG,GAAG,MAAM,CAAC,SAAS,IAAI,sBAAsB;YAAE,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;QAC3G,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QAC3C,IAAI,OAAO,KAAK,SAAS,IAAI,GAAG,GAAG,OAAO,CAAC,EAAE,IAAI,uBAAuB;YAAE,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;QACxG,MAAM,aAAa,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QAChD,MAAM,cAAc,GAAG,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QAClD,OAAO;YACL,GAAG,CAAC,aAAa,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,aAAa,EAAE,CAAC;YACjE,GAAG,CAAC,cAAc,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,cAAc,EAAE,CAAC;SACrE,CAAA;IACH,CAAC;IAED,qFAAqF;IACrF,KAAK,CAAC,SAAiB;QACrB,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;QAC7B,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;IAChC,CAAC;IAED,wDAAwD;IACxD,QAAQ;QACN,IAAI,CAAC,MAAM,CAAC,KAAK,EAAE,CAAA;QACnB,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,CAAA;IACtB,CAAC;CACF"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-llm-verifier",
-  "version": "0.6.0",
+  "version": "0.6.2",
   "description": "Configurable DSH-native LLM verifier with a Web settings page",
   "repository": {
     "type": "git",

package/src/client-i18n.ts

@@ -228,6 +228,7 @@ export const zh = {
   'recent.decisionPrompt': '提示词',
   'recent.decisionOutput': '原始回答',
   'recent.details': '查看详情',
+  'recent.detailsHide': '收起详情',
   'recent.detail.criterion': '判据',
   'recent.detail.score': '分数',
   'recent.detail.threshold': '阈值',
@@ -256,6 +257,15 @@ export const zh = {
   'models.tokens': '{tokens} Token',
   'models.empty': '暂无模型调用',
 
+  // Chat chip for the automatic stages that give no other sign of life
+  'activity.route.classifying': '正在识别需要独立复核的对象…',
+  'activity.route.compare': '正在比较候选方案（约 {n} 次调用）…',
+  'activity.route.select': '正在为多个候选排名（约 {n} 次调用）…',
+  'activity.route.track': '正在复核任务进度（约 {n} 次调用）…',
+  'activity.final.accepting': '正在最终验收（约 {n} 次调用）…',
+  'activity.final.accepted': '最终验收通过',
+  'activity.final.rejected': '最终验收未通过',
+
   // Process selection (P06) chat chip: the reply is being buffered, so this is the only sign of life
   'process.generating': '正在过程选优：生成 {n} 份候选…',
   'process.comparing': '正在过程选优：裁判比较 {n} 份候选…',
@@ -563,6 +573,7 @@ export const en: I18nDict = {
   'recent.decisionPrompt': 'Prompt',
   'recent.decisionOutput': 'Raw answer',
   'recent.details': 'Details',
+  'recent.detailsHide': 'Hide details',
   'recent.detail.criterion': 'Criterion',
   'recent.detail.score': 'Score',
   'recent.detail.threshold': 'Threshold',
@@ -591,6 +602,15 @@ export const en: I18nDict = {
   'models.tokens': '{tokens} Tokens',
   'models.empty': 'No model calls',
 
+  // Chat chip for the automatic stages that give no other sign of life
+  'activity.route.classifying': 'Identifying what needs independent review…',
+  'activity.route.compare': 'Comparing the candidate replies (about {n} calls)…',
+  'activity.route.select': 'Ranking the candidate set (about {n} calls)…',
+  'activity.route.track': 'Reviewing task progress (about {n} calls)…',
+  'activity.final.accepting': 'Running the final acceptance (about {n} calls)…',
+  'activity.final.accepted': 'Final acceptance passed',
+  'activity.final.rejected': 'Final acceptance did not pass',
+
   // Process selection (P06) chat chip: the reply is being buffered, so this is the only sign of life
   'process.generating': 'Selecting the best reply: generating {n} candidates…',
   'process.comparing': 'Selecting the best reply: judging {n} candidates…',
@@ -694,40 +714,59 @@ export function tFormat(template: string, params?: Record<string, string | numbe
   return template.replace(/\{(\w+)\}/g, (match, key) => (key in params ? String(params[key]) : match))
 }
 
-/** The process-selection chip's wire shape (a subset of the server's `ProcessActivityView`). */
-export interface ProcessActivityChipView {
-  active?: { phase?: unknown; candidates?: unknown; alternativeModel?: unknown }
-  settled?: { outcome?: unknown; candidates?: unknown; at?: unknown }
+/** The chat chip's wire shape (a subset of the server's `ActivityView`). */
+export interface ActivityChipView {
+  active?: { stage?: unknown; phase?: unknown; destination?: unknown; candidates?: unknown; expectedCalls?: unknown }
+  settled?: { stage?: unknown; outcome?: unknown }
 }
 
 /** How the chip renders: a busy accent, a settled confirmation, or a failure note. */
-export type ProcessActivityTone = 'busy' | 'ok' | 'error'
+export type ActivityTone = 'busy' | 'ok' | 'error'
+
+/** A count that is safe to print, falling back to the given minimum for a missing/garbled value. */
+function activityCount(value: unknown, minimum: number): number {
+  return typeof value === 'number' && Number.isFinite(value) ? Math.max(minimum, Math.floor(value)) : minimum
+}
 
 /**
- * Render the process-selection chip, or nothing when there is nothing to say.
+ * Render the chat chip, or nothing when there is nothing to say.
  *
  * Pure on purpose (no React, no locale detection): both the chip and its regression tests read the
- * same mapping. An unknown phase counts as generating — a newer host describing a phase this build
- * does not know must still show that work is happening, not disappear.
+ * same mapping. Unknown enum values degrade to the least surprising copy — a newer host describing a
+ * phase or a stage this build does not know must still show that work is happening, not disappear.
  * @param view - the server's activity view, or nothing.
  * @param t - the active dictionary.
  * @returns The tone and copy, or null to render no row at all.
  */
-export function processActivityText(
-  view: ProcessActivityChipView | null | undefined,
+export function verifierActivityText(
+  view: ActivityChipView | null | undefined,
   t: I18nDict,
-): { tone: ProcessActivityTone; text: string } | null {
+): { tone: ActivityTone; text: string } | null {
   const active = view?.active
   if (active !== undefined && active !== null) {
-    const candidates = typeof active.candidates === 'number' && Number.isFinite(active.candidates)
-      ? Math.max(2, Math.floor(active.candidates))
-      : 2
+    if (active.stage === 'route') {
+      if (active.phase !== 'reviewing') return { tone: 'busy', text: t['activity.route.classifying'] }
+      const calls = activityCount(active.expectedCalls, 1)
+      if (active.destination === 'select') return { tone: 'busy', text: tFormat(t['activity.route.select'], { n: calls }) }
+      if (active.destination === 'track') return { tone: 'busy', text: tFormat(t['activity.route.track'], { n: calls }) }
+      return { tone: 'busy', text: tFormat(t['activity.route.compare'], { n: calls }) }
+    }
+    if (active.stage === 'final') {
+      return { tone: 'busy', text: tFormat(t['activity.final.accepting'], { n: activityCount(active.expectedCalls, 1) }) }
+    }
+    // Process selection: the reply is buffered, so this is the only sign of life it has.
+    const candidates = activityCount(active.candidates, 2)
     return active.phase === 'comparing'
       ? { tone: 'busy', text: tFormat(t['process.comparing'], { n: candidates }) }
       : { tone: 'busy', text: tFormat(t['process.generating'], { n: candidates }) }
   }
-  const outcome = view?.settled?.outcome
-  switch (outcome) {
+  const settled = view?.settled
+  if (settled?.stage === 'final') {
+    if (settled.outcome === 'accepted') return { tone: 'ok', text: t['activity.final.accepted'] }
+    if (settled.outcome === 'rejected') return { tone: 'error', text: t['activity.final.rejected'] }
+    return null
+  }
+  switch (settled?.outcome) {
     case 'replaced': return { tone: 'ok', text: t['process.replaced'] }
     case 'kept': return { tone: 'ok', text: t['process.kept'] }
     case 'same': return { tone: 'ok', text: t['process.same'] }