@reticlehq/server 3.5.0 → 3.7.0

package/dist/features/harness/drive-runs.d.ts

@@ -0,0 +1,120 @@
+/**
+ * The Harness drives this daemon is running, and the ones it finished, by run id.
+ *
+ * A drive takes minutes and an MCP client gives a tool call about one. Awaited inside the call, a
+ * drive outlived the client's timeout: the agent got no run id, the daemon kept driving and
+ * spending, and a retry started a second paid drive. So a drive is started here and the call only
+ * waits on it for a while; whoever asks again by run id (the same agent, another one, the HUD)
+ * reads the same entry.
+ *
+ * No I/O and no tool surface: what to run and where to write it arrive as functions, so this stays
+ * a sink the surface hands things to.
+ */
+import { type ImpactSnapshot } from '@reticlehq/core';
+export declare const DriveStatus: {
+    readonly RUNNING: "running";
+    readonly DONE: "done";
+    readonly STOPPED: "stopped";
+    readonly BROKEN: "broken";
+    /** Refused before it drove anything (switched off, no credits, locked): not a broken drive. */
+    readonly REFUSED: "refused";
+};
+export type DriveStatus = (typeof DriveStatus)[keyof typeof DriveStatus];
+/**
+ * Thrown when the platform or the gate refuses a drive before it drives. Releasing a tab reported
+ * such a run as "broken", which reads as a fault in the Harness; a refusal is a refusal.
+ */
+export declare class DriveRefusedError extends Error {
+}
+/** Who asked for the drive. */
+export declare const DriveOrigin: {
+    readonly AGENT: "agent";
+    readonly HUD: "hud";
+    readonly CHAT: "chat";
+    readonly TRY: "try";
+};
+export type DriveOrigin = (typeof DriveOrigin)[keyof typeof DriveOrigin];
+/**
+ * One drive, as it is written to `.reticle/runs/<run>.drive.json` and answered to a poll.
+ *
+ * The id is `harnessRun`, never `runId`: the sync uploads every file in that directory carrying a
+ * `runId` as a verification run, and this record would then collide with the drive's own run.
+ */
+export interface DriveRecord {
+    harnessRun: string;
+    status: DriveStatus;
+    origin: DriveOrigin;
+    sessionId?: string;
+    startedAt: number;
+    endedAt?: number;
+    steps: number;
+    lastLine: string;
+    /** The tool's full answer, once the drive has ended. */
+    result?: Record<string, unknown>;
+    /** Why it broke, when it threw instead of answering. */
+    error?: string;
+}
+/** What a running drive reads from whoever started it. */
+export interface DriveControl {
+    /** True once somebody asked it to stop. */
+    stopped: () => boolean;
+    /** Text somebody said to it since it last asked, oldest first. */
+    takeSaid: () => string[];
+}
+export interface StartDrive {
+    harness: string;
+    runId: string;
+    origin: DriveOrigin;
+    sessionId?: string;
+    now: () => number;
+    persist: (record: DriveRecord) => Promise<void>;
+    run: (control: DriveControl) => Promise<{
+        status: DriveStatus;
+        result: Record<string, unknown>;
+    }>;
+}
+export declare function startDrive(start: StartDrive): DriveRecord;
+/** The drive this run id names, while this daemon still holds it. */
+export declare function driveRecord(runId: string): DriveRecord | undefined;
+/**
+ * The drive running on this session, if any. An unnamed session matches any running drive, and a
+ * drive started without one matches any session: either way it is the tab a second start would use.
+ */
+export declare function runningDrive(sessionId: string | undefined): DriveRecord | undefined;
+/** Every drive still running, for the HUD. */
+export declare function runningDrives(): DriveRecord[];
+/**
+ * Wait until the drive ends, `ms` passes (no limit when undefined), or `signal` aborts — whichever
+ * is first. Ending the wait never ends the drive. `onChange` hears every step while waiting.
+ */
+export declare function awaitDrive(runId: string, ms: number | undefined, signal?: AbortSignal, onChange?: (record: DriveRecord) => void): Promise<void>;
+/** Ask a running drive to stop. False when there is no such drive running. */
+export declare function stopDrive(runId: string): boolean;
+/** Queue text for the driver's next turn. False when there is no such drive running. */
+export declare function sayToDrive(runId: string, text: string): boolean;
+/** One tool call a drive made. `harness` may be a lane's id (`<harness>-L2`). */
+export declare function noteDriveStep(harness: string, tool: string): void;
+/** A line the drive narrated to the HUD. */
+export declare function noteDriveLine(harness: string, line: string): void;
+/** Hear every step a drive takes, so a panel showing its count can repaint. Returns the unsubscribe. */
+export declare function onDriveStep(listener: (record: DriveRecord) => void): () => void;
+/** Hear every drive start and end, from any origin. Returns the unsubscribe. */
+export declare function onDriveChange(listener: (record: DriveRecord) => void): () => void;
+/**
+ * The notes the agent has not heard: a drive started or finished, from any origin. Taken once.
+ * `except` is the run the agent is already reading, which needs no note about itself.
+ */
+export declare function takeDriveNotes(except?: string): string[];
+/** Tests only: every drive and note forgotten. */
+export declare function forgetDrives(): void;
+/**
+ * How long one poll waits on a drive before answering `running`. Under the 60s tool timeout most
+ * MCP clients use, with room for the answer to travel; a poll's `wait` overrides it.
+ */
+export declare const POLL_WAIT_DEFAULT_S = 50;
+/** The exact call that reads a drive: the same call while it runs and once it has ended. */
+export declare function pollCall(runId: string): string;
+/** What a poll answering `running` tells the agent to do, so it stays rather than walks away. */
+export declare function stillRunning(runId: string, steps: number, elapsedS: number): string;
+/** The impact snapshot, with the drive running on this tab when there is one. */
+export declare function withRunningDrive(snapshot: ImpactSnapshot | undefined, sessionId: string): ImpactSnapshot | undefined;

package/dist/features/harness/drive-runs.js

@@ -0,0 +1,289 @@
+/**
+ * The Harness drives this daemon is running, and the ones it finished, by run id.
+ *
+ * A drive takes minutes and an MCP client gives a tool call about one. Awaited inside the call, a
+ * drive outlived the client's timeout: the agent got no run id, the daemon kept driving and
+ * spending, and a retry started a second paid drive. So a drive is started here and the call only
+ * waits on it for a while; whoever asks again by run id (the same agent, another one, the HUD)
+ * reads the same entry.
+ *
+ * No I/O and no tool surface: what to run and where to write it arrive as functions, so this stays
+ * a sink the surface hands things to.
+ */
+import { ReticleTool } from '@reticlehq/core';
+export const DriveStatus = {
+    RUNNING: 'running',
+    DONE: 'done',
+    STOPPED: 'stopped',
+    BROKEN: 'broken',
+    /** Refused before it drove anything (switched off, no credits, locked): not a broken drive. */
+    REFUSED: 'refused',
+};
+/**
+ * Thrown when the platform or the gate refuses a drive before it drives. Releasing a tab reported
+ * such a run as "broken", which reads as a fault in the Harness; a refusal is a refusal.
+ */
+export class DriveRefusedError extends Error {
+}
+/** Who asked for the drive. */
+export const DriveOrigin = { AGENT: 'agent', HUD: 'hud', CHAT: 'chat', TRY: 'try' };
+/** Finished drives kept for polling; older ones are read back from disk. */
+const MAX_FINISHED = 50;
+const LANE_SUFFIX = /-L\d+$/;
+const entries = new Map();
+const byHarness = new Map();
+/** One pending note per drive, latest state wins: what the agent has not heard yet. */
+const notes = new Map();
+const changeListeners = new Set();
+const stepListeners = new Set();
+export function startDrive(start) {
+    const record = {
+        harnessRun: start.runId,
+        status: DriveStatus.RUNNING,
+        origin: start.origin,
+        ...(start.sessionId === undefined ? {} : { sessionId: start.sessionId }),
+        startedAt: start.now(),
+        steps: 0,
+        lastLine: 'Harness is starting',
+    };
+    const entry = {
+        record,
+        harness: start.harness,
+        abort: new AbortController(),
+        inbox: [],
+        waiters: new Set(),
+    };
+    entries.set(start.runId, entry);
+    byHarness.set(start.harness, entry);
+    const save = () => {
+        // A record that cannot be written is still a drive the registry answers for.
+        void start.persist({ ...entry.record }).catch(() => undefined);
+    };
+    save();
+    changed(entry, true);
+    const control = {
+        stopped: () => entry.abort.signal.aborted,
+        takeSaid: () => entry.inbox.splice(0),
+    };
+    void start
+        .run(control)
+        .then(({ status, result }) => {
+        entry.record = { ...entry.record, status, result };
+    }, (error) => {
+        entry.record = {
+            ...entry.record,
+            status: error instanceof DriveRefusedError ? DriveStatus.REFUSED : DriveStatus.BROKEN,
+            error: error instanceof Error ? error.message : String(error),
+        };
+    })
+        .then(() => {
+        entry.record.endedAt = start.now();
+        save();
+        changed(entry, true);
+        prune();
+    });
+    return { ...record };
+}
+/** The drive this run id names, while this daemon still holds it. */
+export function driveRecord(runId) {
+    const entry = entries.get(runId);
+    return entry === undefined ? undefined : { ...entry.record };
+}
+/**
+ * The drive running on this session, if any. An unnamed session matches any running drive, and a
+ * drive started without one matches any session: either way it is the tab a second start would use.
+ */
+export function runningDrive(sessionId) {
+    for (const entry of entries.values()) {
+        if (DriveStatus.RUNNING !== entry.record.status)
+            continue;
+        const own = entry.record.sessionId;
+        if (sessionId === undefined || own === undefined || own === sessionId)
+            return { ...entry.record };
+    }
+    return undefined;
+}
+/** Every drive still running, for the HUD. */
+export function runningDrives() {
+    return [...entries.values()]
+        .filter((entry) => DriveStatus.RUNNING === entry.record.status)
+        .map((entry) => ({ ...entry.record }));
+}
+/**
+ * Wait until the drive ends, `ms` passes (no limit when undefined), or `signal` aborts — whichever
+ * is first. Ending the wait never ends the drive. `onChange` hears every step while waiting.
+ */
+export function awaitDrive(runId, ms, signal, onChange) {
+    const entry = entries.get(runId);
+    if (entry === undefined || DriveStatus.RUNNING !== entry.record.status)
+        return Promise.resolve();
+    return new Promise((resolve) => {
+        const waiter = (record) => {
+            onChange?.(record);
+            if (DriveStatus.RUNNING !== record.status)
+                done();
+        };
+        const done = () => {
+            if (timer !== undefined)
+                clearTimeout(timer);
+            entry.waiters.delete(waiter);
+            signal?.removeEventListener('abort', done);
+            resolve();
+        };
+        const timer = ms === undefined ? undefined : setTimeout(done, ms);
+        entry.waiters.add(waiter);
+        signal?.addEventListener('abort', done, { once: true });
+        if (true === signal?.aborted)
+            done();
+    });
+}
+/** Ask a running drive to stop. False when there is no such drive running. */
+export function stopDrive(runId) {
+    const entry = entries.get(runId);
+    if (entry === undefined || DriveStatus.RUNNING !== entry.record.status)
+        return false;
+    entry.abort.abort();
+    return true;
+}
+/** Queue text for the driver's next turn. False when there is no such drive running. */
+export function sayToDrive(runId, text) {
+    const entry = entries.get(runId);
+    if (entry === undefined || DriveStatus.RUNNING !== entry.record.status)
+        return false;
+    entry.inbox.push(text);
+    return true;
+}
+/** One tool call a drive made. `harness` may be a lane's id (`<harness>-L2`). */
+export function noteDriveStep(harness, tool) {
+    const entry = entryFor(harness);
+    if (entry === undefined)
+        return;
+    entry.record.steps += 1;
+    entry.record.lastLine = `step ${String(entry.record.steps)} · ${tool}`;
+    changed(entry, false);
+    const record = { ...entry.record };
+    for (const listener of stepListeners) {
+        try {
+            listener(record);
+        }
+        catch {
+            /* a listener never breaks the drive it listens to */
+        }
+    }
+}
+/** A line the drive narrated to the HUD. */
+export function noteDriveLine(harness, line) {
+    const entry = entryFor(harness);
+    if (entry === undefined)
+        return;
+    entry.record.lastLine = line;
+    changed(entry, false);
+}
+/** Hear every step a drive takes, so a panel showing its count can repaint. Returns the unsubscribe. */
+export function onDriveStep(listener) {
+    stepListeners.add(listener);
+    return () => stepListeners.delete(listener);
+}
+/** Hear every drive start and end, from any origin. Returns the unsubscribe. */
+export function onDriveChange(listener) {
+    changeListeners.add(listener);
+    return () => changeListeners.delete(listener);
+}
+/**
+ * The notes the agent has not heard: a drive started or finished, from any origin. Taken once.
+ * `except` is the run the agent is already reading, which needs no note about itself.
+ */
+export function takeDriveNotes(except) {
+    if (except !== undefined)
+        notes.delete(except);
+    const taken = [...notes.values()];
+    notes.clear();
+    return taken;
+}
+/** Tests only: every drive and note forgotten. */
+export function forgetDrives() {
+    entries.clear();
+    byHarness.clear();
+    notes.clear();
+}
+function entryFor(harness) {
+    return byHarness.get(harness) ?? byHarness.get(harness.replace(LANE_SUFFIX, ''));
+}
+/** Waiters hear every change; an agent and the HUD only hear a start or an end. */
+function changed(entry, startOrEnd) {
+    const record = { ...entry.record };
+    for (const waiter of [...entry.waiters])
+        waiter(record);
+    if (!startOrEnd)
+        return;
+    notes.set(record.harnessRun, noteFor(record));
+    for (const listener of changeListeners) {
+        try {
+            listener(record);
+        }
+        catch {
+            /* a listener never breaks the drive it listens to */
+        }
+    }
+}
+/**
+ * How long one poll waits on a drive before answering `running`. Under the 60s tool timeout most
+ * MCP clients use, with room for the answer to travel; a poll's `wait` overrides it.
+ */
+export const POLL_WAIT_DEFAULT_S = 50;
+/** How much of a finished drive's own summary rides in the note: its first line, bounded. */
+const NOTE_SUMMARY_MAX = 200;
+/** The exact call that reads a drive: the same call while it runs and once it has ended. */
+export function pollCall(runId) {
+    return `${ReticleTool.VERIFY} {action:"explore", runId:"${runId}"}`;
+}
+/** What a poll answering `running` tells the agent to do, so it stays rather than walks away. */
+export function stillRunning(runId, steps, elapsedS) {
+    return (`Still driving (${String(steps)} steps, ${String(elapsedS)}s). Call ${pollCall(runId)} again ` +
+        `now: it waits up to ${String(POLL_WAIT_DEFAULT_S)}s and answers the moment the drive ends. If ` +
+        `you stop polling, the drive keeps running and this runId still reads it.`);
+}
+const FROM = {
+    [DriveOrigin.AGENT]: '',
+    [DriveOrigin.HUD]: ' (started from the HUD)',
+    [DriveOrigin.CHAT]: ' (started from the Reticle dashboard)',
+    [DriveOrigin.TRY]: ' (started by `reticle try`)',
+};
+function noteFor(record) {
+    const id = record.harnessRun;
+    const tab = record.sessionId === undefined ? '' : ` on tab ${record.sessionId}`;
+    if (DriveStatus.RUNNING === record.status)
+        return (`Harness drive ${id}${FROM[record.origin]} is running${tab}. Follow it: ${pollCall(id)}. ` +
+            `Each call waits up to ${String(POLL_WAIT_DEFAULT_S)}s and answers the moment it ends; if you ` +
+            `stop polling, the drive keeps running and this runId still reads it.`);
+    return `Harness drive ${id}${FROM[record.origin]} finished (${summaryOf(record)}). Full result: ${pollCall(id)}.`;
+}
+function summaryOf(record) {
+    if (record.error !== undefined)
+        return `${record.status}: ${record.error}`;
+    const checks = record.result?.['checks'];
+    if ('object' !== typeof checks || null === checks)
+        return record.status;
+    const count = (key) => {
+        const n = checks[key];
+        return String('number' === typeof n ? n : 0);
+    };
+    const said = record.result?.['summary'];
+    const first = 'string' === typeof said ? (said.split('\n')[0] ?? '').slice(0, NOTE_SUMMARY_MAX) : '';
+    return `${record.status}: ${count('held')} held, ${count('failed')} failed, ${count('undecided')} undecided${0 === first.length ? '' : `. ${first}`}`;
+}
+function prune() {
+    const finished = [...entries.entries()].filter(([, entry]) => DriveStatus.RUNNING !== entry.record.status);
+    for (const [runId, entry] of finished.slice(0, Math.max(0, finished.length - MAX_FINISHED))) {
+        entries.delete(runId);
+        byHarness.delete(entry.harness);
+    }
+}
+/** The impact snapshot, with the drive running on this tab when there is one. */
+export function withRunningDrive(snapshot, sessionId) {
+    const running = runningDrive(sessionId);
+    if (running === undefined || snapshot === undefined)
+        return snapshot;
+    const drive = { runId: running.harnessRun, steps: running.steps };
+    return { ...snapshot, harnessDrive: drive };
+}

package/dist/features/harness/drivers.d.ts

@@ -1,19 +1,15 @@
 /**
- * The names of the drivers the harness can be asked for.
+ * The Harness's driver names.
  *
- * A leaf module on purpose. These used to be re-exported from `harness-explore.ts`, which sits in
- * the tool surface and is reached, through the toolset and the tool table, from the very tool that
- * needs to declare them as an enum — a cycle that left the list `[undefined, undefined]` at the
- * moment zod read it, and failed inside a JSON-schema parser rather than anywhere near the cause.
- * Nothing here imports anything, so nothing can be half-initialised when it is read.
+ * A leaf module on purpose: the tool table reads these while the tool surface is still loading, and
+ * a module that imports nothing cannot be read half-initialised.
+ *
+ * There is one driver. Every Harness decision, its models and its planning run on the Reticle
+ * platform, on the workspace's Harness credits; this machine executes what it is told.
  */
-/** Generates its tool calls as text, and names the journeys it records. The default. */
-export declare const ANTHROPIC_DRIVER_NAME = "anthropic";
-/** Answers typed questions; picks from the elements on the page and cannot invent one. */
-export declare const JEV_DRIVER_NAME = "jev";
-/** Generates its tool calls, over Chat Completions. Here so an A/B has a third arm. */
-export declare const OPENAI_DRIVER_NAME = "openai";
-/** What a caller may name. Order is the order a reader sees them in the tool description. */
-export declare const DRIVER_NAMES: readonly ["anthropic", "jev", "openai"];
+/** The platform's Harness: it decides on its side, this machine executes. See server-driver.ts. */
+export declare const SERVER_DRIVER = "server";
 /** An injected driver: not selectable, because only a caller in-process can supply one. */
 export declare const CUSTOM_DRIVER_NAME = "custom";
+/** What a drive needs, said once for every place that tells somebody. */
+export declare const EXPLORE_NEEDS = "Runs on the Reticle platform: needs a linked project (`reticle connect`); Free has credits.";

package/dist/features/harness/drivers.js

@@ -1,19 +1,15 @@
 /**
- * The names of the drivers the harness can be asked for.
+ * The Harness's driver names.
  *
- * A leaf module on purpose. These used to be re-exported from `harness-explore.ts`, which sits in
- * the tool surface and is reached, through the toolset and the tool table, from the very tool that
- * needs to declare them as an enum — a cycle that left the list `[undefined, undefined]` at the
- * moment zod read it, and failed inside a JSON-schema parser rather than anywhere near the cause.
- * Nothing here imports anything, so nothing can be half-initialised when it is read.
+ * A leaf module on purpose: the tool table reads these while the tool surface is still loading, and
+ * a module that imports nothing cannot be read half-initialised.
+ *
+ * There is one driver. Every Harness decision, its models and its planning run on the Reticle
+ * platform, on the workspace's Harness credits; this machine executes what it is told.
  */
-/** Generates its tool calls as text, and names the journeys it records. The default. */
-export const ANTHROPIC_DRIVER_NAME = 'anthropic';
-/** Answers typed questions; picks from the elements on the page and cannot invent one. */
-export const JEV_DRIVER_NAME = 'jev';
-/** Generates its tool calls, over Chat Completions. Here so an A/B has a third arm. */
-export const OPENAI_DRIVER_NAME = 'openai';
-/** What a caller may name. Order is the order a reader sees them in the tool description. */
-export const DRIVER_NAMES = [ANTHROPIC_DRIVER_NAME, JEV_DRIVER_NAME, OPENAI_DRIVER_NAME];
+/** The platform's Harness: it decides on its side, this machine executes. See server-driver.ts. */
+export const SERVER_DRIVER = 'server';
 /** An injected driver: not selectable, because only a caller in-process can supply one. */
 export const CUSTOM_DRIVER_NAME = 'custom';
+/** What a drive needs, said once for every place that tells somebody. */
+export const EXPLORE_NEEDS = 'Runs on the Reticle platform: needs a linked project (`reticle connect`); Free has credits.';

package/dist/features/harness/explore/console-pages.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Pages of the real console, read in full by a live persona drive (persona 21), trimmed. The
+ * exploratory Harness's tests replay its live misses against these rather than against a tidy tree.
+ */
+/** Impact (/overview), signed in: the rail, the banner and the top of the page. */
+export declare const CONSOLE_OVERVIEW = "- link \"Skip to content\" (ref=e68)\n- complementary\n  - link \"Impact\" (ref=e95)\n    - img \"Reticle\" (ref=e96)\n  - navigation \"Main\" (ref=e97)\n    - link \"Impact\" (ref=e77)\n      - text \"Impact\"\n    - link \"Chat\" (ref=e98)\n      - text \"Chat\"\n    - link \"Inbox\" (ref=e99)\n      - text \"Inbox\"\n    - link \"Issues\" (ref=e100)\n      - text \"Issues\"\n    - link \"Runs\" (ref=e101)\n      - text \"Runs\"\n    - link \"Flows\" (ref=e102)\n      - text \"Flows\"\n  - region \"Chats\" (ref=e103)\n    - paragraph [text=\"Your chats will be listed here.\"]\n  - button \"Account: Bea Final21, Bea-Labs F302964 \u00B7 Pro\" (ref=e104)\n    - text \"Bea Final21\"\n    - text \"Bea-Labs F302964 \u00B7 Pro\"\n- button \"Resize sidebar\" (ref=e105)\n- banner\n  - button \"Hide sidebar (\u2318B)\" (ref=e106) [pressed]\n  - button \"Bea-Labs F302964\" (ref=e107)\n    - text \"Bea-Labs F302964\"\n  - button \"All projects\" (ref=e108)\n    - text \"All projects\"\n  - button \"Search, \u2318K\" (ref=e109)\n    - text \"Search\"\n    - text \"\u2318K\"\n  - link \"14 days left \u00B7 500 trial credits left\" (ref=e73)\n    - text \"14 days left\"\n    - text \"\u00B7 500 trial credits left\"\n  - generic \"p21f-bea-302964@example.com\" (ref=e110) [grid-cols:20px]\n  - generic \"p21f-ana-302964@example.com\" (ref=e111) [grid-cols:20px]\n  - generic \"p21f-raj-302964@example.com\" (ref=e112) [grid-cols:20px]\n  - link \"Invite\" (ref=e113)\n- main\n  - heading \"What Reticle did for your team\" (ref=e91)\n  - paragraph [text=\"Last 30 days\"]\n  - paragraph [text=\"Reticle tested the work of your team 1 time in the last 30 days and found nothin\u2026\"]\n  - link \"Runs that failed 0 of 1 run\" (ref=e114)\n    - text \"Runs that failed\"\n    - text \"0\"\n    - text \"of 1 run\"\n  - link \"Runs that passed 1 of 1 run\" (ref=e115)\n    - text \"Runs that passed\"\n    - text \"1\"\n    - text \"of 1 run\"\n  - link \"Runs that partly passed 0\" (ref=e116)\n    - text \"Runs that partly passed\"\n    - text \"0\"\n  - link \"Runs not proved 0\" (ref=e117)\n    - text \"Runs not proved\"\n    - text \"0\"\n  - link \"Issues found and fixed 0 An issue that a later check passes.\" (ref=e118)\n    - text \"Issues found and fixed\"\n    - text \"0\"";
+/** Flows (/flows): Memory is a link in the page's own "Flows" navigation, not in the rail. */
+export declare const CONSOLE_FLOWS = "- link \"Skip to content\" (ref=e68)\n- complementary\n  - link \"Impact\" (ref=e95)\n    - img \"Reticle\" (ref=e96)\n  - navigation \"Main\" (ref=e97)\n    - link \"Impact\" (ref=e77)\n      - text \"Impact\"\n    - link \"Chat\" (ref=e98)\n      - text \"Chat\"\n    - link \"Inbox\" (ref=e99)\n      - text \"Inbox\"\n    - link \"Issues\" (ref=e100)\n      - text \"Issues\"\n    - link \"Runs\" (ref=e101)\n      - text \"Runs\"\n    - link \"Flows\" (ref=e102) [focused]\n      - text \"Flows\"\n  - region \"Chats\" (ref=e103)\n    - paragraph [text=\"Your chats will be listed here.\"]\n  - button \"Account: Bea Final21, Bea-Labs F302964 \u00B7 Pro\" (ref=e104)\n    - text \"Bea Final21\"\n    - text \"Bea-Labs F302964 \u00B7 Pro\"\n- button \"Resize sidebar\" (ref=e105)\n- banner\n  - button \"Hide sidebar (\u2318B)\" (ref=e106) [pressed]\n  - button \"Bea-Labs F302964\" (ref=e107)\n    - text \"Bea-Labs F302964\"\n  - button \"All projects\" (ref=e108)\n    - text \"All projects\"\n  - button \"Search, \u2318K\" (ref=e109)\n    - text \"Search\"\n    - text \"\u2318K\"\n  - link \"14 days left \u00B7 500 trial credits left\" (ref=e73)\n    - text \"14 days left\"\n    - text \"\u00B7 500 trial credits left\"\n  - generic \"p21f-bea-302964@example.com\" (ref=e110) [grid-cols:20px]\n  - generic \"p21f-ana-302964@example.com\" (ref=e111) [grid-cols:20px]\n  - generic \"p21f-raj-302964@example.com\" (ref=e112) [grid-cols:20px]\n  - link \"Invite\" (ref=e113)\n- main\n  - navigation \"Flows\" (ref=e144)\n    - link \"Flows\" (ref=e148)\n    - link \"Plans\" (ref=e149)\n    - link \"Memory\" (ref=e150)\n  - heading \"Flows\" (ref=e151)\n  - paragraph [text=\"Saved paths through your app. Reticle can replay each flow.\"]\n  - group \"View\" (ref=e152)\n    - button \"Map\" (ref=e153)\n    - button \"List\" (ref=e154) [pressed]\n  - text \"Flows\"\n  - text \"1\"\n  - text \"Pages they reach\"\n  - text \"1\"\n  - text \"Failing now\"\n  - text \"0\"\n  - text \"Not replayed yet\"\n  - text \"1\"\n  - heading \"Starts at /\" (ref=e146)";
+/**
+ * Impact with the workspace switcher open, as the live drive left it: the menu makes the rest of the
+ * page inert, so a read sees the menu and nothing else.
+ */
+export declare const CONSOLE_SWITCHER_OPEN: string;
+/** Memory (/memory): the shared project's records. */
+export declare const CONSOLE_MEMORY: string;

package/dist/features/harness/explore/console-pages.js

@@ -0,0 +1,146 @@
+/**
+ * Pages of the real console, read in full by a live persona drive (persona 21), trimmed. The
+ * exploratory Harness's tests replay its live misses against these rather than against a tidy tree.
+ */
+/** Impact (/overview), signed in: the rail, the banner and the top of the page. */
+export const CONSOLE_OVERVIEW = `- link "Skip to content" (ref=e68)
+- complementary
+  - link "Impact" (ref=e95)
+    - img "Reticle" (ref=e96)
+  - navigation "Main" (ref=e97)
+    - link "Impact" (ref=e77)
+      - text "Impact"
+    - link "Chat" (ref=e98)
+      - text "Chat"
+    - link "Inbox" (ref=e99)
+      - text "Inbox"
+    - link "Issues" (ref=e100)
+      - text "Issues"
+    - link "Runs" (ref=e101)
+      - text "Runs"
+    - link "Flows" (ref=e102)
+      - text "Flows"
+  - region "Chats" (ref=e103)
+    - paragraph [text="Your chats will be listed here."]
+  - button "Account: Bea Final21, Bea-Labs F302964 · Pro" (ref=e104)
+    - text "Bea Final21"
+    - text "Bea-Labs F302964 · Pro"
+- button "Resize sidebar" (ref=e105)
+- banner
+  - button "Hide sidebar (⌘B)" (ref=e106) [pressed]
+  - button "Bea-Labs F302964" (ref=e107)
+    - text "Bea-Labs F302964"
+  - button "All projects" (ref=e108)
+    - text "All projects"
+  - button "Search, ⌘K" (ref=e109)
+    - text "Search"
+    - text "⌘K"
+  - link "14 days left · 500 trial credits left" (ref=e73)
+    - text "14 days left"
+    - text "· 500 trial credits left"
+  - generic "p21f-bea-302964@example.com" (ref=e110) [grid-cols:20px]
+  - generic "p21f-ana-302964@example.com" (ref=e111) [grid-cols:20px]
+  - generic "p21f-raj-302964@example.com" (ref=e112) [grid-cols:20px]
+  - link "Invite" (ref=e113)
+- main
+  - heading "What Reticle did for your team" (ref=e91)
+  - paragraph [text="Last 30 days"]
+  - paragraph [text="Reticle tested the work of your team 1 time in the last 30 days and found nothin…"]
+  - link "Runs that failed 0 of 1 run" (ref=e114)
+    - text "Runs that failed"
+    - text "0"
+    - text "of 1 run"
+  - link "Runs that passed 1 of 1 run" (ref=e115)
+    - text "Runs that passed"
+    - text "1"
+    - text "of 1 run"
+  - link "Runs that partly passed 0" (ref=e116)
+    - text "Runs that partly passed"
+    - text "0"
+  - link "Runs not proved 0" (ref=e117)
+    - text "Runs not proved"
+    - text "0"
+  - link "Issues found and fixed 0 An issue that a later check passes." (ref=e118)
+    - text "Issues found and fixed"
+    - text "0"`;
+/** Flows (/flows): Memory is a link in the page's own "Flows" navigation, not in the rail. */
+export const CONSOLE_FLOWS = `- link "Skip to content" (ref=e68)
+- complementary
+  - link "Impact" (ref=e95)
+    - img "Reticle" (ref=e96)
+  - navigation "Main" (ref=e97)
+    - link "Impact" (ref=e77)
+      - text "Impact"
+    - link "Chat" (ref=e98)
+      - text "Chat"
+    - link "Inbox" (ref=e99)
+      - text "Inbox"
+    - link "Issues" (ref=e100)
+      - text "Issues"
+    - link "Runs" (ref=e101)
+      - text "Runs"
+    - link "Flows" (ref=e102) [focused]
+      - text "Flows"
+  - region "Chats" (ref=e103)
+    - paragraph [text="Your chats will be listed here."]
+  - button "Account: Bea Final21, Bea-Labs F302964 · Pro" (ref=e104)
+    - text "Bea Final21"
+    - text "Bea-Labs F302964 · Pro"
+- button "Resize sidebar" (ref=e105)
+- banner
+  - button "Hide sidebar (⌘B)" (ref=e106) [pressed]
+  - button "Bea-Labs F302964" (ref=e107)
+    - text "Bea-Labs F302964"
+  - button "All projects" (ref=e108)
+    - text "All projects"
+  - button "Search, ⌘K" (ref=e109)
+    - text "Search"
+    - text "⌘K"
+  - link "14 days left · 500 trial credits left" (ref=e73)
+    - text "14 days left"
+    - text "· 500 trial credits left"
+  - generic "p21f-bea-302964@example.com" (ref=e110) [grid-cols:20px]
+  - generic "p21f-ana-302964@example.com" (ref=e111) [grid-cols:20px]
+  - generic "p21f-raj-302964@example.com" (ref=e112) [grid-cols:20px]
+  - link "Invite" (ref=e113)
+- main
+  - navigation "Flows" (ref=e144)
+    - link "Flows" (ref=e148)
+    - link "Plans" (ref=e149)
+    - link "Memory" (ref=e150)
+  - heading "Flows" (ref=e151)
+  - paragraph [text="Saved paths through your app. Reticle can replay each flow."]
+  - group "View" (ref=e152)
+    - button "Map" (ref=e153)
+    - button "List" (ref=e154) [pressed]
+  - text "Flows"
+  - text "1"
+  - text "Pages they reach"
+  - text "1"
+  - text "Failing now"
+  - text "0"
+  - text "Not replayed yet"
+  - text "1"
+  - heading "Starts at /" (ref=e146)`;
+/**
+ * Impact with the workspace switcher open, as the live drive left it: the menu makes the rest of the
+ * page inert, so a read sees the menu and nothing else.
+ */
+export const CONSOLE_SWITCHER_OPEN = [
+    '- menu "Bea-Labs F302964" (ref=e300)',
+    '  - menuitem "Bea-Labs F302964" (ref=e301)',
+    '  - menuitem "New workspace" (ref=e302)',
+    '- region "Notifications alt+T" (ref=e1)',
+].join('\n');
+/** Memory (/memory): the shared project's records. */
+export const CONSOLE_MEMORY = [
+    '- main',
+    '  - navigation "Flows" (ref=e144)',
+    '    - link "Flows" (ref=e148)',
+    '    - link "Plans" (ref=e149)',
+    '    - link "Memory" (ref=e150) [focused]',
+    '  - heading "Memory" (ref=e160)',
+    '  - heading "One record your team disagrees on" (ref=e161)',
+    '  - button "Free shipping threshold 1 saved fact updated today" (ref=e162)',
+    '- region "Notifications alt+T" (ref=e1)',
+].join('\n');

package/dist/features/harness/explore/disparity.d.ts

@@ -0,0 +1,19 @@
+/**
+ * Two-way testing: the coding agent's verdict on a goal beside the Harness's, compared.
+ *
+ * The agent knows where everything is; the Harness starts with one sentence. Agent proved and
+ * Harness could not find is a discoverability finding (where a newcomer gives up), not noise. Only
+ * a DECIDED agent verdict (yes or no) on a claim naming the goal's text is compared: an undecided one
+ * says nothing either way, and a claim about something else is not the same goal.
+ */
+import { Verified, type JournalAction, type RunDisparity } from '@reticlehq/core';
+import type { GoalCheck } from '../goals.js';
+import type { GoalSearch } from './goal-search.js';
+export interface AgentVerdict {
+    claim: string;
+    verified: Verified;
+}
+/** The decided verdicts the connected agent itself made: actions no Harness drove. */
+export declare function agentVerdictsIn(actions: readonly JournalAction[]): AgentVerdict[];
+/** The comparison for one goal, or undefined when the agent made no decided claim about it. */
+export declare function disparityFor(check: GoalCheck, agent: readonly AgentVerdict[], search: GoalSearch | undefined): RunDisparity | undefined;