@reticlehq/server 2.5.0 → 2.7.0

package/dist/tools/invoke-tool.d.ts

@@ -7,10 +7,4 @@ export declare const SESSION_BOUND_TOOLS: ReadonlySet<string>;
  * `sessionId`-bearing tool to be classified into exactly one set (bound XOR exempt).
  */
 export declare const SESSION_EXEMPT_TOOLS: ReadonlySet<string>;
-/**
- * The single entry point both the MCP server and the programmatic invoker call instead of
- * `tool.handler` directly. Runs the handler, then — for a live-session tool returning a plain
- * object that did not already include `session` — splices the health envelope on. Idempotent
- * (handlers that already add health are left untouched) and never alters non-object results.
- */
 export declare function runTool(tool: ToolDef, deps: ToolDeps, args: Record<string, unknown>): Promise<unknown>;

package/dist/tools/invoke-tool.js

@@ -1,5 +1,5 @@
 import { healthEnvelope } from '../session/session-health.js';
-import { TelemetryActor, TelemetryEventKind } from '@reticlehq/core';
+import { TelemetryActor, TelemetryEventKind, TRANSPORT_LIMITS, } from '@reticlehq/core';
 import { getSessionMetrics } from '../telemetry/session-metrics.js';
 import { getTelemetry } from '../telemetry/telemetry.js';
 import { takeUpdateNudge } from '../update/update-nudge.js';
@@ -8,10 +8,12 @@ import { noteToolCall } from '../daemon/daemon-usefulness.js';
 import { bugsInResult } from '../telemetry/bug-found.js';
 import { verificationOf } from '../telemetry/verification-of.js';
 import { asString } from './tools-helpers.js';
+import { EnvelopeKey } from './tool-kit.js';
 import { ReticleTool } from './tool-names.js';
 import { takeFeedbackPrompt } from './feedback-tools.js';
 import { takeFeedbackUndelivered } from '../telemetry/feedback-delivery.js';
 import { span } from '../trace.js';
+import { frictionOf, inviteFor } from './feedback-invite.js';
 /**
  * The live-session tools whose result MUST carry the
  * session-health envelope. Owned in ONE place — not retrofitted per handler — so a throttled tab
@@ -138,6 +140,27 @@ function reportBugsFound(toolName, result) {
  * object that did not already include `session` — splices the health envelope on. Idempotent
  * (handlers that already add health are left untouched) and never alters non-object results.
  */
+/**
+ * The name of the first argument whose string value exceeds the transport bound, if any.
+ *
+ * One level of nesting is searched because that is where callers put payloads — `reticle_act`'s
+ * `args.value`, and the fuzz's own `huge-string` shape. Deeper recursion would cost more than it
+ * buys: the failure is a single huge string, not a deeply nested structure of small ones.
+ */
+function firstOversizedArg(args) {
+    const tooLong = (value) => 'string' === typeof value && value.length > TRANSPORT_LIMITS.MAX_STRING_LENGTH;
+    for (const [key, value] of Object.entries(args)) {
+        if (tooLong(value))
+            return key;
+        if ('object' === typeof value && null !== value && !Array.isArray(value)) {
+            for (const [inner, nested] of Object.entries(value)) {
+                if (tooLong(nested))
+                    return `${key}.${inner}`;
+            }
+        }
+    }
+    return undefined;
+}
 export async function runTool(tool, deps, args) {
     // Both dispatch paths (MCP + programmatic) pass through here — the one place "which tool is mostly
     // used" can be counted. This used to EMIT an event per call; it now increments an in-process counter
@@ -150,6 +173,28 @@ export async function runTool(tool, deps, args) {
     // closure carries this call's own identity, which is also what makes peak-concurrency measurable.
     // A daemon that has served even one tool call is doing a job for somebody; see daemon-usefulness.
     noteToolCall();
+    // An oversized argument is refused BEFORE the handler deserialises it.
+    //
+    // tool-fuzz failed CI on the invariant that matters most — `every tool answers every hostile call`
+    // — for a ~100KB string. Seen twice with different tools; the earlier one came back
+    // `-32001 sse_aborted` with the tool surface dropping 48→17, which is a RESTARTED daemon. The
+    // shape is: oversized argument → the daemon dies → the proxy answers the in-flight call with
+    // transport loss, and the next assertion is talking to a different daemon.
+    //
+    // `TRANSPORT_LIMITS.MAX_STRING_LENGTH` already bounds what crosses the bridge; a tool argument
+    // arrives over MCP stdio and never met it. Here, because this is the one place every invocation
+    // routes through, so a tool added later inherits the bound instead of having to remember it.
+    //
+    // A refusal is a normal answer. An UNANSWERED call is a hung agent, and that is the failure this
+    // prevents: the agent learns its argument was too big, which is something it can fix.
+    const oversized = firstOversizedArg(args);
+    if (oversized !== undefined) {
+        return {
+            error: `argument '${oversized}' is larger than ${String(TRANSPORT_LIMITS.MAX_STRING_LENGTH)} ` +
+                'characters, which is the most Reticle moves in one call. Nothing ran. Send a smaller ' +
+                'value — a selector or an id rather than a document, or a slice of the text you need.',
+        };
+    }
     // An ACTION is what gets abandoned. Counted here, against verifications, so "the agent drove the
     // page and then wandered off" becomes a number instead of an impression.
     if (ACTION_TOOLS.has(tool.name))
@@ -218,12 +263,12 @@ export async function runTool(tool, deps, args) {
         ? raw
         : {
             ...raw,
-            ...(prompt !== undefined ? { feedback_prompt: prompt } : {}),
-            ...(update !== undefined ? { update_available: update } : {}),
-            ...(skew !== undefined ? { version_skew: skew } : {}),
+            ...(prompt !== undefined ? { [EnvelopeKey.FEEDBACK_PROMPT]: prompt } : {}),
+            ...(update !== undefined ? { [EnvelopeKey.UPDATE_AVAILABLE]: update } : {}),
+            ...(skew !== undefined ? { [EnvelopeKey.VERSION_SKEW]: skew } : {}),
             ...(undelivered !== undefined
                 ? {
-                    feedback_undelivered: `your earlier report did NOT send: ${undelivered}. Tell the human what you found so it is not lost.`,
+                    [EnvelopeKey.FEEDBACK_UNDELIVERED]: `your earlier report did NOT send: ${undelivered}. Tell the human what you found so it is not lost.`,
                 }
                 : {}),
         };
@@ -242,9 +287,33 @@ export async function runTool(tool, deps, args) {
     // one-time pool-lease reminder or the age cleanup nudge. Splice them regardless.
     const lease = resolved.takeSessionLease();
     if (lease !== undefined)
-        envelope['session_lease'] = lease;
+        envelope[EnvelopeKey.SESSION_LEASE] = lease;
     const warning = resolved.ageWarning();
     if (warning !== undefined)
-        envelope['session_age_warning'] = warning;
+        envelope[EnvelopeKey.SESSION_AGE_WARNING] = warning;
+    // Ask for a verdict when the agent has driven the page and not asked for one. Almost every
+    // verdict-less session in the field never called a verdict-producing tool ONCE;
+    // the counter behind this already existed and was reported only to us. One-shot per abandoned
+    // run, same discipline as the pool lease — a hint on every call is noise that gets tuned out.
+    const unverified = getSessionMetrics().takeUnverifiedNudge();
+    if (unverified !== undefined)
+        envelope[EnvelopeKey.VERIFY_NEXT] = unverified;
+    // Invite feedback at the moment of friction, worded for what just happened. Feedback produced
+    // ~11 of this release's fixes from 14 reports; the handshake instruction is what got those 14,
+    // and this closes the gap it leaves — the agent mid-problem, not thinking about feedback tooling.
+    // Counted so `feedback_submitted / feedbackPrompted` can tell us whether it works at all.
+    const friction = frictionOf({
+        unknownTool: false,
+        verifiedUnknown: 'unknown' === result.verified,
+        repeatRun: getSessionMetrics().currentRun,
+        errored: 'error' in result,
+        // A verdict is progress. Without this, three successful act_and_wait calls in a row — the exact
+        // loop this release exists to encourage — get told they are stuck.
+        producedVerdict: 'verified' in result,
+    });
+    if (friction !== undefined) {
+        envelope[EnvelopeKey.FEEDBACK_INVITE] = inviteFor(friction);
+        getSessionMetrics().recordFeedbackPrompt();
+    }
     return Object.keys(envelope).length > 0 ? { ...result, ...envelope } : result;
 }

package/dist/tools/navigate-arrival.d.ts

@@ -0,0 +1,26 @@
+/**
+ * Wait, briefly, for the SDK to reconnect after a navigation — so `confirmed` can mean something.
+ *
+ * `window.location.assign(url)` destroys the SDK that would report on the new document, so the
+ * navigation itself cannot be confirmed from inside the page. But the SDK reconnects TO THE DAEMON,
+ * which makes the daemon the only party that can see arrival happen. Before this, it did not look,
+ * and told the agent to poll `reticle_sessions` instead — a tool call on every navigation, on the
+ * least reliable tool we ship.
+ *
+ * Bounded and best-effort by construction: a navigation to a page that is not instrumented, or not
+ * there at all, must still return promptly with `confirmed:false` rather than hanging. The clock and
+ * the sleep are injected so this is testable without waiting on a real one.
+ */
+import type { SessionManager } from '../session/session-manager.js';
+import type { NavigateArrival } from './navigate-result.js';
+/** Long enough for a dev server to serve a page and the SDK to dial back; short enough to not hang. */
+export declare const ARRIVAL_TIMEOUT_MS = 5000;
+export interface ArrivalClock {
+    now: () => number;
+    sleep: (ms: number) => Promise<void>;
+}
+/**
+ * Poll until a session is present at `target`, or the budget runs out. Returns `null` on timeout —
+ * never throws, because a failure to confirm is a legitimate answer, not an error.
+ */
+export declare function awaitArrival(sessions: SessionManager, target: string, timeoutMs?: number, clock?: ArrivalClock): Promise<NavigateArrival | null>;

package/dist/tools/navigate-arrival.js

@@ -0,0 +1,60 @@
+/**
+ * Wait, briefly, for the SDK to reconnect after a navigation — so `confirmed` can mean something.
+ *
+ * `window.location.assign(url)` destroys the SDK that would report on the new document, so the
+ * navigation itself cannot be confirmed from inside the page. But the SDK reconnects TO THE DAEMON,
+ * which makes the daemon the only party that can see arrival happen. Before this, it did not look,
+ * and told the agent to poll `reticle_sessions` instead — a tool call on every navigation, on the
+ * least reliable tool we ship.
+ *
+ * Bounded and best-effort by construction: a navigation to a page that is not instrumented, or not
+ * there at all, must still return promptly with `confirmed:false` rather than hanging. The clock and
+ * the sleep are injected so this is testable without waiting on a real one.
+ */
+/** Long enough for a dev server to serve a page and the SDK to dial back; short enough to not hang. */
+export const ARRIVAL_TIMEOUT_MS = 5_000;
+const POLL_MS = 100;
+/**
+ * Compare only origin + pathname.
+ *
+ * The app is entitled to add or rewrite the query and hash on arrival — a redirect to
+ * `?redirect=%2F`, a router normalising a trailing slash, an auth guard appending a reason. Matching
+ * the whole URL string would report `confirmed:false` for a navigation that plainly succeeded, which
+ * is the same lie in the opposite direction.
+ */
+function samePage(a, b) {
+    try {
+        const x = new URL(a);
+        const y = new URL(b);
+        return x.origin === y.origin && x.pathname.replace(/\/$/, '') === y.pathname.replace(/\/$/, '');
+    }
+    catch {
+        return false;
+    }
+}
+function findArrival(sessions, target) {
+    for (const s of sessions.all()) {
+        if (samePage(s.url, target))
+            return { sessionId: s.id };
+    }
+    return null;
+}
+const REAL_CLOCK = {
+    now: () => Date.now(),
+    sleep: (ms) => new Promise((resolve) => setTimeout(resolve, ms)),
+};
+/**
+ * Poll until a session is present at `target`, or the budget runs out. Returns `null` on timeout —
+ * never throws, because a failure to confirm is a legitimate answer, not an error.
+ */
+export async function awaitArrival(sessions, target, timeoutMs = ARRIVAL_TIMEOUT_MS, clock = REAL_CLOCK) {
+    const deadline = clock.now() + timeoutMs;
+    for (;;) {
+        const found = findArrival(sessions, target);
+        if (found !== null)
+            return found;
+        if (clock.now() >= deadline)
+            return null;
+        await clock.sleep(POLL_MS);
+    }
+}

package/dist/tools/navigate-result.d.ts

@@ -10,8 +10,22 @@
  * `reticle_act` already draws this line — `dispatched` (sent) versus `settled` (a frame flushed).
  * This gives navigate the same honesty without changing what it does.
  */
+/** The session the SDK reconnected as after the page moved, when one was observed in the window. */
+export interface NavigateArrival {
+    sessionId: string;
+}
 export declare function navigateResult(result: {
     ok?: unknown;
     url?: unknown;
     reason?: unknown;
-}): Record<string, unknown>;
+}, 
+/**
+ * Arrival, if the daemon saw the SDK come back. `null` keeps the original honest envelope.
+ *
+ * Measured on three fixtures: `confirmed` was `false` on every navigation on every app, because
+ * nothing ever set it. A boolean that never varies teaches an agent nothing, and an agent that
+ * learns to ignore it will keep ignoring it when it starts meaning something. The daemon is the
+ * one place that CAN answer this — the SDK reconnects to it — so making it answer removes a
+ * `reticle_sessions` poll from every single navigation.
+ */
+arrival?: NavigateArrival | null): Record<string, unknown>;

package/dist/tools/navigate-result.js

@@ -1,16 +1,14 @@
+export function navigateResult(result, 
 /**
- * The honest envelope for a navigation nobody can confirm.
+ * Arrival, if the daemon saw the SDK come back. `null` keeps the original honest envelope.
  *
- * `window.location.assign(url)` returns before the page moves, and the SDK that would report on the
- * new document is destroyed by the navigation itself. So a successful NAVIGATE means the browser
- * accepted the instruction — not that anything arrived. Driven against a dead URL it returned
- * `{"ok":true}` while the session died, which is the worst combination available: the agent believes
- * it navigated, is looking at nothing, and has lost Reticle.
- *
- * `reticle_act` already draws this line — `dispatched` (sent) versus `settled` (a frame flushed).
- * This gives navigate the same honesty without changing what it does.
+ * Measured on three fixtures: `confirmed` was `false` on every navigation on every app, because
+ * nothing ever set it. A boolean that never varies teaches an agent nothing, and an agent that
+ * learns to ignore it will keep ignoring it when it starts meaning something. The daemon is the
+ * one place that CAN answer this — the SDK reconnects to it — so making it answer removes a
+ * `reticle_sessions` poll from every single navigation.
  */
-export function navigateResult(result) {
+arrival = null) {
     const ok = true === result.ok;
     const base = {
         ok,
@@ -20,6 +18,15 @@ export function navigateResult(result) {
     // A refusal is conclusive: the page never moved, so there is nothing unconfirmed to report.
     if (!ok)
         return base;
+    if (arrival !== null) {
+        return {
+            ...base,
+            confirmed: true,
+            // The SDK that reconnects after a navigation is a NEW session with a new id. Returning it here
+            // is the difference between a confirmation and a confirmation the agent can act on.
+            sessionId: arrival.sessionId,
+        };
+    }
     return {
         ...base,
         confirmed: false,

package/dist/tools/observe-tools.js

@@ -10,6 +10,7 @@ import { ReticleTool } from './tool-names.js';
 import { buildReactionReport } from '../events/reaction.js';
 import { findContradictions } from '../events/contradictions.js';
 import { evaluatePredicate, waitForPredicate, PredicateSchema } from '../events/predicate.js';
+import { parsePredicate } from '../events/predicate-parse.js';
 import { matchNet, matchConsole, isConsoleEvent, eventMatchesFilters, netEmptyHint, consoleEmptyHint, reconcileNet, projectNetCall, projectConsoleLog, } from '../events/event-filters.js';
 import { applyEventBudget, costHint, withSizeCost, DEFAULT_QUERY_LIMIT, } from '../session/output-budget.js';
 import { healthEnvelope, bufferEnvelope } from '../session/session-health.js';
@@ -209,6 +210,10 @@ export const OBSERVE_TOOLS = [
             pass: z.boolean(),
             evidence: z.unknown().optional(),
             failureReason: z.string().optional(),
+            observationLost: z
+                .boolean()
+                .optional()
+                .describe('The tab disconnected mid-wait, so this was never observed — the verdict is UNKNOWN, not a failure of the app.'),
             session: z
                 .object({ lastSeenMs: z.number(), throttled: z.boolean(), focused: z.boolean() })
                 .optional(),
@@ -233,7 +238,7 @@ export const OBSERVE_TOOLS = [
         handler: async (deps, args) => {
             const session = deps.sessions.resolve(asString(args['sessionId']));
             // `until` is act_and_wait's name for this — see alias-args.ts.
-            const predicate = PredicateSchema.parse(aliasParam(args, 'predicate', ['until'])['predicate']);
+            const predicate = parsePredicate(aliasParam(args, 'predicate', ['until'])['predicate']);
             // Honesty: explicit since wins; else default to the last act's cursor; else the whole buffer.
             const since = asNumber(args['since']) ?? session.lastAct.cursor() ?? 0;
             const verdict = await waitForPredicate(session, predicate, asNumber(args['timeout_ms']) ?? DEFAULT_ASSERT_TIMEOUT_MS, since);
@@ -252,7 +257,16 @@ export const OBSERVE_TOOLS = [
         example: { predicate: { kind: PredicateKind.SIGNAL, name: 'todos:loaded' } },
         description: 'Evaluate a predicate (optionally waiting up to timeout_ms). Returns { pass, evidence, failureReason? }. The end of every verify loop. Prefer a { signal } or { net } consequence over { element }/{ text } presence — a passing presence-only assertion returns `advice` because a wrong/healed element can fake it. By default it only counts events since your last act, so a stale buffered signal can never fake a pass; pass `since` (an observe/act cursor) to set the window explicitly.',
         inputSchema: {
-            predicate: PredicateSchema.optional().describe('Predicate to evaluate: { signal }, { net }, { element } or a combination.'),
+            predicate: PredicateSchema.optional().describe(
+            // Every kind, each with the field that carries its argument. Five of the nine were
+            // undocumented here, including `route` — and "did submitting the login form navigate away"
+            // is the most common thing an agent wants to assert. A field report reached us from an agent
+            // that guessed `urlContains` on route (net's spelling) and got unrecognized_keys.
+            'Predicate to evaluate. Kinds: { signal, name } { net, urlContains|method|status|count } ' +
+                '{ state, path|equals } { route, pathname (exact) | contains (path+query+hash) } ' +
+                '{ element, testid|role|text } { text } { console, level|absent } { animation, name } ' +
+                '{ settled } — combine with { allOf | anyOf | not }. Prefer a signal/net/state consequence ' +
+                'over element/text presence.'),
             until: PredicateSchema.optional().describe("Alias for `predicate` (act_and_wait's name)."),
             timeout_ms: z
                 .number()
@@ -268,6 +282,10 @@ export const OBSERVE_TOOLS = [
             pass: z.boolean(),
             evidence: z.unknown().optional(),
             failureReason: z.string().optional(),
+            observationLost: z
+                .boolean()
+                .optional()
+                .describe('The tab disconnected mid-wait, so this was never observed — the verdict is UNKNOWN, not a failure of the app.'),
             advice: z
                 .string()
                 .optional()
@@ -304,7 +322,7 @@ export const OBSERVE_TOOLS = [
         handler: async (deps, args) => {
             const session = deps.sessions.resolve(asString(args['sessionId']));
             // `until` is act_and_wait's name for this — see alias-args.ts.
-            const predicate = PredicateSchema.parse(aliasParam(args, 'predicate', ['until'])['predicate']);
+            const predicate = parsePredicate(aliasParam(args, 'predicate', ['until'])['predicate']);
             const timeout = asNumber(args['timeout_ms']) ?? 0;
             // Honesty: explicit since wins; else default to the last act's cursor; else the whole buffer.
             const since = asNumber(args['since']) ?? session.lastAct.cursor() ?? 0;
@@ -323,7 +341,7 @@ export const OBSERVE_TOOLS = [
                     : assertsDerivedIpcStatus(predicate)
                         ? { advice: DERIVED_IPC_STATUS_ADVICE }
                         : {};
-            const { decision, contradictions, coverage } = await assertVerdict(session, predicate, verdict.pass, since, verdict.inconclusive);
+            const { decision, contradictions, coverage } = await assertVerdict(session, predicate, verdict.pass, since, verdict.inconclusive, verdict.observationLost);
             return withControl(session, {
                 ...decision,
                 ...verdict,

package/dist/tools/real-input-attempt.js

@@ -27,17 +27,22 @@ function synthetic(reason) {
  */
 export async function tryRealInput(deps, session, ref, action, args) {
     const provider = deps.realInput;
-    if (provider === undefined)
-        return synthetic(); // real input not configured — no diagnostic
+    const inner = asRecord(args['args']);
+    const askedForNative = true === inner[NATIVE_INPUT_ARG];
+    if (provider === undefined) {
+        // Silent by default: with no provider EVERY action is synthetic, and a reason on all of them is
+        // noise on the most-used tool in the product. But an agent that passed native:true asked a
+        // question and got the opposite answer — reported from the field as a silent downgrade that cost
+        // real debugging time, because the tool description promises a reason is "never silent".
+        return askedForNative ? synthetic(InputModeReason.NOT_CONFIGURED) : synthetic();
+    }
     if (!isPointerAction(action))
         return synthetic(InputModeReason.NOT_POINTER); // fill/type stay synthetic
-    const inner = asRecord(args['args']);
     // "Don't click, run the code": a click/dblclick runs the occlusion-honest SYNTHETIC path by default
     // even with a provider configured — no coordinate gesture to be intercepted by the HUD or missed
     // off-screen. Opt into a trusted native click with args.native:true (file pickers, clipboard,
     // isTrusted-gated handlers). hover/drag genuinely need native pointer state, so they stay real.
-    if ((action === ActionType.CLICK || action === ActionType.DBLCLICK) &&
-        inner[NATIVE_INPUT_ARG] !== true) {
+    if ((action === ActionType.CLICK || action === ActionType.DBLCLICK) && !askedForNative) {
         return synthetic(InputModeReason.SYNTHETIC_CLICK_PREFERRED);
     }
     if (!(await provider.isAvailableFor(session.url)))

package/dist/tools/tool-kit.d.ts

@@ -33,6 +33,14 @@ export interface ToolDeps {
     fs: FileSystemPort;
     /** absolute.reticle path (index.ts computes cwd/.reticle). */
     reticleRoot: string;
+    /**
+     * This daemon's OWN project id, derived from the directory it was started in.
+     *
+     * Optional because a daemon above an uninitialised directory has none, and because every existing
+     * test construction of ToolDeps predates it. Absence means "cannot tell", which must never be
+     * treated as a mismatch.
+     */
+    projectId?: string;
     /** injected clock for the contract's generatedAt stamp. */
     now: () => number;
     /**
@@ -76,11 +84,58 @@ export declare const sessionIdShape: {
     sessionId: z.ZodOptional<z.ZodString>;
 };
 /**
- * Fields that `runTool` / `withControl` splice onto EVERY session-bound tool result at runtime
- * (health, pool lease reminder, age cleanup nudge, and the delivered-once human-control envelope).
- * They are declared here and merged into each session-bound tool's outputSchema so a schema-strict
- * client (structuredContent validation) keeps them instead of silently dropping them — the `control`
- * envelope is the human-in-the-loop guidance channel, so losing it is a safety failure, not cosmetic.
+ * Every key `runTool` splices onto a session-bound result, as a closed list.
+ *
+ * These are wire strings and were inlined at the splice sites, which is exactly how four of them
+ * came to be spliced and never declared. `sessionEnvelopeShape` is derived from this object and
+ * `envelope-keys-declared.test.ts` fails if a member is missing from it, so the two cannot drift.
+ *
+ * Each of these is a channel that exists BECAUSE the agent had no other way to learn the fact.
+ * Undeclared, a schema-strict client validating `structuredContent` drops it, and the channel is
+ * indistinguishable from never having been built.
+ */
+export declare const EnvelopeKey: {
+    /** Session health — which tab, how fresh. */
+    readonly SESSION: "session";
+    /** One-shot pool-lease reminder. */
+    readonly SESSION_LEASE: "session_lease";
+    /** One-shot "these sessions are old, clean them up". */
+    readonly SESSION_AGE_WARNING: "session_age_warning";
+    /** The delivered-once human-in-the-loop guidance envelope. Losing it is a safety failure. */
+    readonly CONTROL: "control";
+    /**
+     * Rides alongside `session` whenever the tab is THROTTLED. It was declared on `reticle_act`'s
+     * schema and nowhere else, so on a validating profile every other session-bound tool silently
+     * dropped it — a throttled tab, where drives can no-op, returned a healthy-looking result.
+     */
+    readonly WARNING: "warning";
+    /** One-shot "ask the human how this went". */
+    readonly FEEDBACK_PROMPT: "feedback_prompt";
+    /** One-shot "a newer Reticle exists". */
+    readonly UPDATE_AVAILABLE: "update_available";
+    /**
+     * "You have driven this page N times without asking for a verdict."
+     *
+     * The largest known lever on the metric this product is judged by: verdict-less sessions
+     * overwhelmingly never call a verdict-producing tool at all. It was spliced undeclared, so on a
+     * validating client the nudge was built, fired, and thrown away before any agent saw it.
+     */
+    readonly VERIFY_NEXT: "verify_next";
+    /** The contextual "tell us what just went wrong" invitation. Same story, same release. */
+    readonly FEEDBACK_INVITE: "feedback_invite";
+    /** SDK/daemon version skew — often the one fact that explains everything else in the session. */
+    readonly VERSION_SKEW: "version_skew";
+    /** A feedback report that was accepted and then failed to send. Only the reporter can act on it. */
+    readonly FEEDBACK_UNDELIVERED: "feedback_undelivered";
+};
+export type EnvelopeKey = (typeof EnvelopeKey)[keyof typeof EnvelopeKey];
+/**
+ * Fields that `runTool` / `withControl` splice onto EVERY session-bound tool result at runtime.
+ * Merged into each session-bound tool's outputSchema so a schema-strict client (structuredContent
+ * validation) keeps them instead of silently dropping them.
+ *
+ * Derived from `EnvelopeKey` rather than re-listed: this shape and the splice sites disagreed for a
+ * whole release, and every key that went missing was a channel nobody could see was missing.
  */
 export declare const sessionEnvelopeShape: z.ZodRawShape;
 /** Unwrap a browser command result or throw its error so the agent sees a clean failure. */

package/dist/tools/tool-kit.js

@@ -14,32 +14,63 @@ export const sessionIdShape = {
         .describe('OMIT THIS unless you mean a specific tab — Reticle scopes to your project, prefers the active one, and refuses rather than guesses when ambiguous. Pass an id from reticle_sessions only to target a particular tab. The previous wording ("omit when only ONE session is open") was wrong and cost real turns: with three sessions connected — an app plus two pool leases — resolution is still unambiguous, but that sentence sent the agent off to list and filter sessions by hand before every single call.'),
 };
 /**
- * Fields that `runTool` / `withControl` splice onto EVERY session-bound tool result at runtime
- * (health, pool lease reminder, age cleanup nudge, and the delivered-once human-control envelope).
- * They are declared here and merged into each session-bound tool's outputSchema so a schema-strict
- * client (structuredContent validation) keeps them instead of silently dropping them — the `control`
- * envelope is the human-in-the-loop guidance channel, so losing it is a safety failure, not cosmetic.
+ * Every key `runTool` splices onto a session-bound result, as a closed list.
+ *
+ * These are wire strings and were inlined at the splice sites, which is exactly how four of them
+ * came to be spliced and never declared. `sessionEnvelopeShape` is derived from this object and
+ * `envelope-keys-declared.test.ts` fails if a member is missing from it, so the two cannot drift.
+ *
+ * Each of these is a channel that exists BECAUSE the agent had no other way to learn the fact.
+ * Undeclared, a schema-strict client validating `structuredContent` drops it, and the channel is
+ * indistinguishable from never having been built.
  */
-export const sessionEnvelopeShape = {
-    session: z.unknown().optional(),
-    session_lease: z.unknown().optional(),
-    session_age_warning: z.unknown().optional(),
-    control: z.unknown().optional(),
-    // `warning` rides alongside `session` whenever the tab is THROTTLED (healthEnvelope splices both).
-    // It was declared on reticle_act's schema but nowhere else, so on a validating profile every other
-    // session-bound tool (observe, assert, wait_for, act_sequence, act_and_wait, snapshot, query, …)
-    // silently dropped it — a throttled tab, where drives can no-op, returned a healthy-looking result.
-    // That is exactly what invoke-tool.ts's health splice exists to prevent, so it belongs in the shared
-    // envelope, not per-tool.
-    warning: z.string().optional(),
-    // The one-shot "ask the human how this went" envelope. Undeclared, a validating profile would strip
-    // it — and a feedback prompt that never reaches the agent is a feedback prompt that never reaches
-    // the person, which is the entire failure mode this channel exists to fix.
-    feedback_prompt: z.unknown().optional(),
-    // One-shot "a newer Reticle exists" notice. Undeclared, a validating profile would strip it and the
-    // agent would never learn an update was waiting.
-    update_available: z.unknown().optional(),
+export const EnvelopeKey = {
+    /** Session health — which tab, how fresh. */
+    SESSION: 'session',
+    /** One-shot pool-lease reminder. */
+    SESSION_LEASE: 'session_lease',
+    /** One-shot "these sessions are old, clean them up". */
+    SESSION_AGE_WARNING: 'session_age_warning',
+    /** The delivered-once human-in-the-loop guidance envelope. Losing it is a safety failure. */
+    CONTROL: 'control',
+    /**
+     * Rides alongside `session` whenever the tab is THROTTLED. It was declared on `reticle_act`'s
+     * schema and nowhere else, so on a validating profile every other session-bound tool silently
+     * dropped it — a throttled tab, where drives can no-op, returned a healthy-looking result.
+     */
+    WARNING: 'warning',
+    /** One-shot "ask the human how this went". */
+    FEEDBACK_PROMPT: 'feedback_prompt',
+    /** One-shot "a newer Reticle exists". */
+    UPDATE_AVAILABLE: 'update_available',
+    /**
+     * "You have driven this page N times without asking for a verdict."
+     *
+     * The largest known lever on the metric this product is judged by: verdict-less sessions
+     * overwhelmingly never call a verdict-producing tool at all. It was spliced undeclared, so on a
+     * validating client the nudge was built, fired, and thrown away before any agent saw it.
+     */
+    VERIFY_NEXT: 'verify_next',
+    /** The contextual "tell us what just went wrong" invitation. Same story, same release. */
+    FEEDBACK_INVITE: 'feedback_invite',
+    /** SDK/daemon version skew — often the one fact that explains everything else in the session. */
+    VERSION_SKEW: 'version_skew',
+    /** A feedback report that was accepted and then failed to send. Only the reporter can act on it. */
+    FEEDBACK_UNDELIVERED: 'feedback_undelivered',
 };
+/**
+ * Fields that `runTool` / `withControl` splice onto EVERY session-bound tool result at runtime.
+ * Merged into each session-bound tool's outputSchema so a schema-strict client (structuredContent
+ * validation) keeps them instead of silently dropping them.
+ *
+ * Derived from `EnvelopeKey` rather than re-listed: this shape and the splice sites disagreed for a
+ * whole release, and every key that went missing was a channel nobody could see was missing.
+ */
+export const sessionEnvelopeShape = Object.fromEntries(Object.values(EnvelopeKey).map((key) => [
+    key,
+    // `warning` is the one with a narrower type; the rest are opaque envelopes the client forwards.
+    EnvelopeKey.WARNING === key ? z.string().optional() : z.unknown().optional(),
+]));
 /** Unwrap a browser command result or throw its error so the agent sees a clean failure. */
 export async function commandOrThrow(deps, sessionId, 
 // ReticleCommand, not string: the union already exists and every caller passes a member. Typed as

package/dist/tools/tool-surface.js

@@ -98,6 +98,19 @@ export const CORE_TOOL_NAMES = new Set([
     ReticleTool.QUERY,
     ReticleTool.ACT,
     ReticleTool.ACT_AND_WAIT,
+    // ACT_SEQUENCE is here because its absence was measurably CAUSING the biggest loop in the data.
+    //
+    // In the field `reticle_act` is called overwhelmingly more often than `reticle_act_sequence`, and
+    // `act` leads the repeat table by a wide margin — inside those sessions the repeated calls are
+    // clicks and fills, a login form driven one round trip at a time, which is exactly the antipattern
+    // SKILL.md warns about and exactly what this tool exists to collapse.
+    //
+    // The repeats are NOT retries: looping sessions have a LOWER error rate than non-looping ones. The
+    // calls succeed and get repeated, because the batching tool was reachable only through
+    // `reticle_run` — so an agent had to already know it existed to use it, and essentially nobody
+    // did. A tool an agent must already know about is a tool that never gets called; the same argument
+    // that put INSPECT and FEEDBACK here.
+    ReticleTool.ACT_SEQUENCE,
     ReticleTool.OBSERVE,
     ReticleTool.NETWORK,
     ReticleTool.CONSOLE,