@reticlehq/server 2.4.0 → 2.6.0

package/dist/tools/error-recovery.js

@@ -8,6 +8,51 @@
  * on stable, human-authored substrings of the thrown messages (see session-manager.ts, the flow/
  * baseline stores). No clock, no IO — unit-testable in isolation.
  */
+import { TRANSPORT_LIMITS } from '@reticlehq/core';
+import { SELF_RECOVERING_MARKER } from '../session/no-session-diagnosis.js';
+/** Marks where an over-long message had its middle removed. Named so tests assert it by value. */
+const ELISION = '… [elided] …';
+/**
+ * Bound a message before the agent ever sees it, keeping BOTH ends.
+ *
+ * Error messages interpolate caller-supplied values — a ref, a baseline name, a selector — and
+ * nothing upstream bounds them, so a 100KB argument came back as a 100KB tool result. Capping here
+ * rather than at each throw site is deliberate: this is the single funnel every tool error passes
+ * through, so it also covers the tools nobody has written yet.
+ *
+ * The middle is what goes. Head-truncation would be simpler and wrong: the recovery table keys off
+ * substrings that sit at the END of the message ("… no longer resolves to an element"), so cutting
+ * the tail severs the part that makes an error recognizable and re-labels a known condition as a
+ * possible Reticle defect.
+ */
+function capMessage(message) {
+    const max = TRANSPORT_LIMITS.MAX_ERROR_LENGTH;
+    if (max >= message.length)
+        return message;
+    const keep = Math.floor((max - ELISION.length) / 2);
+    return `${message.slice(0, keep)}${ELISION}${message.slice(message.length - keep)}`;
+}
+/**
+ * Messages that already spell out the exact retry, so any hint we add can only argue with them.
+ *
+ * The no-session diagnosis is one (it inspected the machine and named the cause). The native-input
+ * refusal is the other: it names the tool, the argument and the follow-up call to make instead — and
+ * it was still collecting the defect ask, inviting a bug report about a refusal that had just told
+ * the agent precisely what to do.
+ */
+const SELF_RECOVERING = [
+    /cannot drive native input/i,
+    // The unknown-session refusal now lists the live ids inline. Appending "call reticle_sessions for
+    // the current ids" to a message that just named them contradicts the shorter path it offers, and
+    // a contradicted instruction is how one agent came to retry a dead id twelve times.
+    /Connected right now:/i,
+];
+/** A message that already carries its own concrete next action, so nothing should be appended. */
+function isSelfRecovering(message) {
+    if (message.includes(SELF_RECOVERING_MARKER))
+        return true;
+    return SELF_RECOVERING.some((pattern) => pattern.test(message));
+}
 /** The recovery hints, named so they are not free strings and can be asserted in tests. */
 export const RECOVERY = {
     NO_SESSION: 'No app is connected to Reticle. Ask the human to start their app in dev with @reticlehq/browser ' +
@@ -21,18 +66,71 @@ export const RECOVERY = {
         'the tab to the front, or run `reticle drive <url>` for a guaranteed scriptable context.',
     MISSING_BASELINE: 'That baseline does not exist yet. Call reticle_baseline { action: "list" } to see saved names, or ' +
         'reticle_baseline { action: "save", name } to capture one before diffing against it.',
-    MISSING_RECORDING: 'No recording by that name is in progress. Start one with reticle_record { action: "start", name } ' +
-        'before annotating, stopping, or saving it.',
+    MISSING_RECORDING: 
+    // `recordingName`, not `name`. The hint said `name`, so an agent that followed it exactly got
+    // InvalidParams — a recovery that fails is worse than none, because it spends the retry.
+    'No recording by that name is in progress. Start one with ' +
+        'reticle_record { action: "start", recordingName } before annotating, stopping, or saving it.',
     STALE_REF: 'That ref is stale: refs are invalidated whenever the DOM re-renders, so any action that ' +
         'navigated, opened a modal, re-sorted a list or changed the page invalidates every ref taken ' +
         'before it. Reticle refuses here rather than clicking whatever now occupies that slot. Call ' +
         'reticle_query again for a fresh ref and retry the action — and prefer reticle_act_and_wait ' +
         '{ until } when an action changes the page, so the next ref is taken after it settles.',
+    INVALID_NAME: 'That name is not a usable one. A flow, baseline or recording name must be a single safe path ' +
+        'segment — letters, digits, dash and underscore only, starting with a letter or digit, at most ' +
+        '64 characters. No slashes, no "..", no leading dot: the name becomes a filename under ' +
+        '.reticle/, so anything that could escape that directory is refused. Retry with a slug like ' +
+        '"checkout-happy-path". This is an invalid call, not a Reticle defect: there is nothing to report.',
+    NO_SUCH_OPTION: 'That <select> has no option with the value you asked for, and the message above lists the ones ' +
+        'it does have. Reticle refuses rather than assigning it: an unmatched value deselects everything, ' +
+        'and the change event that follows has been observed persisting an empty setting into the app. ' +
+        'Retry with one of the listed values — note they are option VALUES, which often differ from the ' +
+        'visible labels shown in brackets. This is an invalid call, not a Reticle defect.',
     BAD_ARGUMENTS: "That call did not match the tool's schema — the message above names the tool and the exact " +
         "argument it wanted. Re-read that tool's parameters and retry with the missing or corrected " +
         'argument. This is an invalid call, not a Reticle defect: there is nothing to report.',
+    WRONG_TARGET: 'That action does not apply to this element — the ref resolved, the element just cannot take it ' +
+        '(you cannot fill a <button>, submit outside a form, or upload to anything but <input ' +
+        'type="file">). Call reticle_snapshot or reticle_query to find the control that does, or ' +
+        'reticle_inspect on this ref to see what it actually is. This is an invalid call, not a Reticle ' +
+        'defect: there is nothing to report.',
+    NOT_EDITABLE: 'The field is disabled or readonly, so a user could not edit it either — forcing the value would ' +
+        'put the app in a state nobody can reach, which is why Reticle refuses. This is usually the app ' +
+        'telling you something: drive whatever ENABLES the field (open the editor, satisfy the ' +
+        'precondition) and act again. If the field should have been editable, that is a bug in the app, ' +
+        'not in Reticle.',
+    CONFIRM_DANGEROUS: 'Reticle blocked a potentially destructive control (delete, remove, revoke…) on purpose. If you ' +
+        'mean it, retry the same action with args.confirmDangerous set to true. This is a deliberate ' +
+        'refusal, not a defect: there is nothing to report.',
+    UNSUPPORTED_SURFACE: 'That surface is not supported yet, and Reticle refuses rather than pretending — a rich-text ' +
+        'editor keeps its own document model, so writing to the DOM would look right and submit the old ' +
+        'content. Drive the same outcome another way (a plain input, a command, a keyboard action) or ' +
+        'assert on the result instead of typing it. Known gap, already reported: no need to file it.',
     TOKEN_REQUIRED: 'The bridge binds beyond localhost and requires a pairing token. Set the same token in the SDK ' +
         'init (@reticlehq/core) and the Reticle server config, then reconnect.',
+    SCOPE_MISMATCH: 'A tab IS connected — it just belongs to a different project than the daemon is scoped to, so ' +
+        'nothing is wrong with the app and there is nothing to report. The message above lists each ' +
+        'connected session with its sessionId: pass that sessionId to this tool to target it, or ' +
+        "restart the daemon from that app's directory so it becomes the default. Call reticle_sessions " +
+        'for the same list at any time.',
+    SESSION_GONE: 'The tab this call was targeting disconnected while the command was in flight, so the command ' +
+        'never completed — do not assume it applied or that it did not. Call reticle_sessions: a tab ' +
+        'that reloaded comes back under a NEW sessionId (every ref taken before it is gone — re-query), ' +
+        'and an empty list means the page was closed and the human has to reopen it.',
+    COMMAND_TIMEOUT: 'The page did not answer within the command window. That is a fact about the page, not a Reticle ' +
+        'failure: check reticle_sessions for `throttled`/`stale` on this session — a backgrounded tab is ' +
+        'throttled and may never answer, so ask the human to bring it to the front or drive your own ' +
+        'browser with reticle_lease. If the tab is in front, the page is busy or blocked (a long ' +
+        'synchronous task, an alert/confirm dialog); reticle_console usually shows what it hit. Retry ' +
+        'once the cause is addressed.',
+    FLOW_STEP_MISSING: 'A recorded step points at an element that is not on the page now. Either the app changed under ' +
+        'the flow — run reticle_flow_heal to re-anchor it, then reticle_flow_verify — or an earlier step ' +
+        'did not do what it used to, leaving the run on the wrong page: check reticle_snapshot to see ' +
+        'where it actually is. This is app/flow drift, which is exactly what the flow exists to catch.',
+    NO_POOL: 'Reticle cannot launch its own browser here: the daemon-managed pool is absent, or Playwright ' +
+        'has no Chromium installed (the message above says which). Run Reticle via `reticle mcp`, and ' +
+        'ask the human for `npx playwright install chromium` if that is what is missing. Meanwhile drive ' +
+        'a tab the human already has open — reticle_sessions lists them.',
 };
 /** Ordered match rules; the first hit wins. Substrings track the thrown messages they recover. */
 const RULES = [
@@ -43,9 +141,51 @@ const RULES = [
     { match: /no baseline named/i, hint: RECOVERY.MISSING_BASELINE },
     { match: /no (?:active|compiled) recording named/i, hint: RECOVERY.MISSING_RECORDING },
     { match: /pairing token is required/i, hint: RECOVERY.TOKEN_REQUIRED },
+    // Three conditions the daemon understands perfectly and still asked for a bug report about. Each
+    // needs its own rule: none of them contains "no browser session connected" (the scope miss is
+    // "no browser session FOR project 'x'", which is the opposite claim — sessions exist).
+    {
+        match: /scope mismatch|session\(s\) ARE connected under a different project/i,
+        hint: RECOVERY.SCOPE_MISMATCH,
+    },
+    { match: /^session disconnected$|session .* never connected/i, hint: RECOVERY.SESSION_GONE },
+    { match: /^command '[^']*' timed out after/i, hint: RECOVERY.COMMAND_TIMEOUT },
+    // Flow replay: the recorded anchor is gone. Ordinary drift, and the tool that fixes it exists.
+    { match: /did not resolve in current page/i, hint: RECOVERY.FLOW_STEP_MISSING },
+    {
+        match: /browser pool (?:unavailable|is shut down)|Chromium is not installed/i,
+        hint: RECOVERY.NO_POOL,
+    },
     // The commonest post-action condition there is. Unmatched, it fell through to FEEDBACK_ASK and
     // told the agent a successful click's aftermath might be a bug in Reticle.
     { match: /no longer resolves to an element/i, hint: RECOVERY.STALE_REF },
+    { match: /no <option> with value/i, hint: RECOVERY.NO_SUCH_OPTION },
+    // Every guard in the browser's executeAction. These are Reticle-authored refusals of an invalid
+    // call, and ALL of them used to fall through to the feedback ask — including the destructive-action
+    // block, which carries its own retry instruction and was still telling the agent Reticle might be
+    // broken. Order matters: contenteditable and disabled/readonly are `cannot <verb> …` messages too,
+    // so they must be tested before the general wrong-target rule.
+    { match: /contenteditable/i, hint: RECOVERY.UNSUPPORTED_SURFACE },
+    { match: /cannot \w+ a (disabled|readonly) </i, hint: RECOVERY.NOT_EDITABLE },
+    // Three spellings ship — "action", "native action", "WebMCP tool" — and the rule matched one, so
+    // two thirds of the same deliberate refusal still read as a possible defect.
+    {
+        match: /potentially destructive (?:\w+ )*(?:action|tool) blocked/i,
+        hint: RECOVERY.CONFIRM_DANGEROUS,
+    },
+    // Split in two because `^` binds to the FIRST alternative only. As one pattern this read as though
+    // the anchor governed all four spellings; it never did, and the next person to add a fifth would
+    // have had a coin-flip's chance of getting the behaviour they intended. The anchored spelling is
+    // one Reticle authors at the start of its own message; the other three appear mid-message.
+    { match: /^cannot [\w()]+ (?:a|an|into a|on a) </i, hint: RECOVERY.WRONG_TARGET },
+    {
+        match: /no form to submit|upload target must be|is not an HTMLElement/i,
+        hint: RECOVERY.WRONG_TARGET,
+    },
+    // Authored by Reticle, about the caller's arguments: the message already names the valid answers.
+    { match: /^unknown action '/i, hint: RECOVERY.BAD_ARGUMENTS },
+    { match: /^unsupported query strategy '/i, hint: RECOVERY.BAD_ARGUMENTS },
+    { match: /^\w+ requires a string `\w+`/i, hint: RECOVERY.BAD_ARGUMENTS },
     // A message naming a `reticle_*` tool is one WE authored about our own API — an invalid call, not
     // an unanticipated failure. Left unmatched it collected the feedback ask, which pushed agents to
     // file reports about their own bad arguments. Kept LAST so a specific rule above always wins.
@@ -54,14 +194,108 @@ const RULES = [
         hint: RECOVERY.BAD_ARGUMENTS,
     },
 ];
-/** The actionable next move for a known error message, or undefined when none is recognized. */
+/**
+ * The actionable next move for a known error message, or undefined when none is recognized.
+ *
+ * A message that already carries its own next action gets NO second one. The no-session diagnosis
+ * inspects the machine and says which of three causes this actually is; pairing that with the
+ * generic "ask the human to start their app" produced a result that contradicted itself — the error
+ * saying a server is running and the recovery saying to go start one.
+ */
 export function recoveryFor(message) {
+    if (isSelfRecovering(message))
+        return undefined;
     for (const rule of RULES) {
         if (rule.match.test(message))
             return rule.hint;
     }
     return undefined;
 }
+/**
+ * A schema rejection: the call named a parameter that does not exist, or gave one the wrong type.
+ *
+ * Matched by the SHAPES the validators actually produce (zod's codes, and the MCP layer's wrapper),
+ * because these are our own machinery talking, not the app.
+ *
+ * `did not parse` is the shape `parsePredicate` produces. It was added because removing the raw zod
+ * array from that path ALSO removed the codes this pattern keys on — so a malformed predicate
+ * stopped being recognised as the agent's mistake and started being blamed on Reticle, complete with
+ * a request for a bug report. The e2e battery caught it (`no bad argument is blamed on Reticle`);
+ * no unit test could, which is why one now exists below.
+ */
+/** Max issues rendered — a union rejection can emit one per member, which is how the array became unreadable. */
+const MAX_RENDERED_ISSUES = 3;
+function isZodIssue(value) {
+    if (typeof value !== 'object' || null === value)
+        return false;
+    const v = value;
+    return 'string' === typeof v['code'] && 'string' === typeof v['message'];
+}
+/** `["net","urlContains"]` → `net.urlContains`; an empty path means the object itself. */
+function issueClause(issue) {
+    if ('unrecognized_keys' === issue.code && Array.isArray(issue.keys)) {
+        return `unknown field ${issue.keys.map(String).join(', ')}`;
+    }
+    const path = issue.path ?? [];
+    return `${0 === path.length ? 'the arguments' : path.map(String).join('.')}: ${issue.message}`;
+}
+/**
+ * A serialized zod issue array, rendered as a sentence. Anything else is returned untouched.
+ *
+ * This is the one boundary every agent-visible tool failure crosses, which is why the rewrite lives
+ * here rather than in each tool. `parsePredicate` solved this for the three predicate tools by never
+ * producing the array; it cannot help anywhere else, because the MCP SDK validates a tool's schema
+ * BEFORE our handler runs. So `reticle_session` with no action — the commonest mistake there is —
+ * never reaches the friendly "unknown action … expected: [tune, yield, …]" its handler already
+ * builds, and the agent gets the raw array instead. Driven live against the shipped daemon.
+ *
+ * Conservative by construction: it only rewrites a JSON ARRAY whose every element carries both
+ * `code` and `message`. A tool that legitimately returns JSON is left alone.
+ */
+export function zodArrayAsSentence(message) {
+    const text = message.trim();
+    if (!text.startsWith('['))
+        return message;
+    let parsed;
+    try {
+        parsed = JSON.parse(text);
+    }
+    catch {
+        return message; // not JSON at all — e.g. prose that happens to open with a bracket
+    }
+    if (!Array.isArray(parsed) || 0 === parsed.length || !parsed.every(isZodIssue))
+        return message;
+    const shown = parsed.slice(0, MAX_RENDERED_ISSUES).map(issueClause).join('; ');
+    const rest = parsed.length - MAX_RENDERED_ISSUES;
+    return rest > 0 ? `${shown} (and ${String(rest)} more)` : shown;
+}
+/**
+ * `unknown field` and `<path>: Required` are the shapes `zodArrayAsSentence` produces.
+ *
+ * Added for the same reason `did not parse` was: rendering the array as a sentence REMOVES the very
+ * codes (`invalid_type`, `unrecognized_keys`) this pattern matched on, so a schema rejection fell
+ * through to the generic branch and came back as "may be a defect in Reticle" — worse than the dump
+ * it replaced, because it also asks for a bug report about the agent's own malformed call. This repo
+ * has now made that mistake twice; the second time a test caught it before CI did.
+ *
+ * Widened a THIRD time for PR #182, which formats the MCP SDK's validation error at the argument
+ * boundary — "Missing required parameter for reticle_session: action (one of: …)". That is better
+ * prose than the field-level sentence below (it names the TOOL and the allowed VALUES) and it was
+ * authored before this formatter existed, so it must not be punished for arriving first.
+ */
+const ARGUMENT_REJECTION = /Unrecognized key|unrecognized_keys|invalid_type|did not parse|unknown field|: Required\b|Missing required parameter|Invalid parameter for|Invalid arguments for tool|Unknown parameters? for|^Required |Expected \w+, received/i;
+/**
+ * What to say instead of asking for a bug report. The schema already stated the problem precisely;
+ * this only says who can fix it and where to look.
+ */
+const ARGUMENT_RECOVERY = "That call did not match the tool's schema — the message above names the parameter. Nothing ran, " +
+    'so no result is affected. Call reticle_tools { names: ["<tool>"] } for its exact parameters and ' +
+    'retry.';
+/**
+ * `invalid <thing> name: <value>` — the shape every name validator in the server throws. Matched on
+ * the phrasing rather than per-caller so a new store that validates a name is covered by default.
+ */
+const INVALID_NAME_REJECTION = /^invalid (?:[a-z]+ )*name: /i;
 /**
  * The ask attached to errors we do NOT recognize. An unrecognized error is, by definition, the case
  * where Reticle failed in a way nobody anticipated — the single highest-value moment to hear from the
@@ -79,7 +313,24 @@ export const FEEDBACK_ASK = 'This error is not one Reticle recognizes, which mea
  * known; the feedback ask is added when it is NOT — the two are mutually exclusive by design, so the
  * agent always gets exactly one next move.
  */
-export function buildErrorPayload(message) {
+export function buildErrorPayload(rawMessage) {
+    // Render the zod array BEFORE capping: the cap would otherwise truncate the JSON mid-issue and
+    // leave an unparseable fragment as the agent's error text — the worst of both shapes.
+    const message = capMessage(zodArrayAsSentence(rawMessage));
+    // A self-diagnosing message gets NEITHER a generic recovery nor the feedback ask. It already
+    // inspected the machine and named the cause; a second, contradictory hint is noise, and inviting a
+    // bug report about a condition Reticle diagnosed itself is exactly backwards.
+    if (isSelfRecovering(message))
+        return { error: message };
+    // An argument rejection is the agent's to fix, not a Reticle defect. Reported from the field: a
+    // clean `unrecognized_keys: ["value"]` came back wrapped in "may be a defect in Reticle", which
+    // spends the agent's turn and fills the feedback channel with reports about malformed calls.
+    if (ARGUMENT_REJECTION.test(message))
+        return { error: message, recovery: ARGUMENT_RECOVERY };
+    // Reticle's own name validators rejecting a caller's value: an argument rejection that the schema
+    // cannot express (the shape is a runtime pattern, not a type), so it needs its own arm.
+    if (INVALID_NAME_REJECTION.test(message))
+        return { error: message, recovery: RECOVERY.INVALID_NAME };
     const recovery = recoveryFor(message);
     return recovery !== undefined
         ? { error: message, recovery }

package/dist/tools/feedback-invite.d.ts

@@ -0,0 +1,48 @@
+/**
+ * The one line that tells an agent it can push back — worded for what just happened.
+ *
+ * Feedback is the highest-yield channel this product has: **14 reports over two days produced
+ * roughly 11 of the fixes in 2.6.0**, against ~6,100 numeric events that produced a funnel which
+ * turned out to be broken. Prose from an agent at the moment of friction beats any structured field
+ * we could add.
+ *
+ * It reached those 14 through a standing instruction in the MCP handshake — paid once at connect.
+ * The gap that leaves is the agent who hits something mid-task and is concentrating on the problem
+ * rather than on remembering that a feedback tool exists.
+ *
+ * So: invite at the moment of friction, and vary the wording. A fixed string repeated forty times
+ * is invisible by call five; a line that names what just went wrong is read because it is about the
+ * thing the agent is already looking at. ~12 tokens against a ~4,546 tok/turn surface.
+ *
+ * Every invite is counted (`feedbackPrompted`). If `feedback_submitted ÷ feedbackPrompted` stays
+ * flat, this is decoration and should be deleted — instrumenting our own nudge is the difference
+ * between a designed system and a hope.
+ */
+/** What just happened, as far as the chokepoint can tell. Ordered most specific first. */
+export declare const FrictionKind: {
+    /** A tool that does not exist — the agent named a capability it expected us to have. */
+    readonly UNKNOWN_TOOL: "unknown_tool";
+    /** A verdict came back `unknown`: we drove the app and could not tell what happened. */
+    readonly UNKNOWN_VERDICT: "unknown_verdict";
+    /** The same tool, several times in a row. */
+    readonly REPEATING: "repeating";
+    /** The tool refused or errored. */
+    readonly REFUSED: "refused";
+};
+export type FrictionKind = (typeof FrictionKind)[keyof typeof FrictionKind];
+/**
+ * Which friction, if any, this result represents.
+ *
+ * Deliberately reads only what the chokepoint already has: no handler changes, and nothing here can
+ * see app data.
+ */
+export declare function frictionOf(facts: {
+    unknownTool: boolean;
+    verifiedUnknown: boolean;
+    repeatRun: number;
+    errored: boolean;
+    /** Did this call produce a verdict? A verdict is progress, and progress is not being stuck. */
+    producedVerdict: boolean;
+}): FrictionKind | undefined;
+/** The line for a friction kind. */
+export declare function inviteFor(kind: FrictionKind): string;

package/dist/tools/feedback-invite.js

@@ -0,0 +1,71 @@
+/**
+ * The one line that tells an agent it can push back — worded for what just happened.
+ *
+ * Feedback is the highest-yield channel this product has: **14 reports over two days produced
+ * roughly 11 of the fixes in 2.6.0**, against ~6,100 numeric events that produced a funnel which
+ * turned out to be broken. Prose from an agent at the moment of friction beats any structured field
+ * we could add.
+ *
+ * It reached those 14 through a standing instruction in the MCP handshake — paid once at connect.
+ * The gap that leaves is the agent who hits something mid-task and is concentrating on the problem
+ * rather than on remembering that a feedback tool exists.
+ *
+ * So: invite at the moment of friction, and vary the wording. A fixed string repeated forty times
+ * is invisible by call five; a line that names what just went wrong is read because it is about the
+ * thing the agent is already looking at. ~12 tokens against a ~4,546 tok/turn surface.
+ *
+ * Every invite is counted (`feedbackPrompted`). If `feedback_submitted ÷ feedbackPrompted` stays
+ * flat, this is decoration and should be deleted — instrumenting our own nudge is the difference
+ * between a designed system and a hope.
+ */
+/** What just happened, as far as the chokepoint can tell. Ordered most specific first. */
+export const FrictionKind = {
+    /** A tool that does not exist — the agent named a capability it expected us to have. */
+    UNKNOWN_TOOL: 'unknown_tool',
+    /** A verdict came back `unknown`: we drove the app and could not tell what happened. */
+    UNKNOWN_VERDICT: 'unknown_verdict',
+    /** The same tool, several times in a row. */
+    REPEATING: 'repeating',
+    /** The tool refused or errored. */
+    REFUSED: 'refused',
+};
+/**
+ * Runs of one tool before it counts as being stuck.
+ *
+ * Three, matching the verdict nudge: `reticle_act` led the field data at 110 consecutive-repeat
+ * runs, and a run of three is where a deliberate retry becomes a loop.
+ */
+const REPEAT_INVITE_AT = 3;
+const LINES = {
+    [FrictionKind.UNKNOWN_TOOL]: 'no such tool — reticle_feedback tells us what you expected it to do, which is how it gets built',
+    [FrictionKind.UNKNOWN_VERDICT]: 'Reticle could not tell what happened here. If you expected otherwise, reticle_feedback — an ' +
+        'unknown verdict is our defect, not yours',
+    [FrictionKind.REPEATING]: 'stuck on the same call? reticle_feedback — say what you were trying to do and it gets fixed',
+    [FrictionKind.REFUSED]: 'if that diagnosis was wrong or unhelpful, reticle_feedback — error messages are a product ' +
+        'surface and we fix the ones that fail',
+};
+/**
+ * Which friction, if any, this result represents.
+ *
+ * Deliberately reads only what the chokepoint already has: no handler changes, and nothing here can
+ * see app data.
+ */
+export function frictionOf(facts) {
+    if (facts.unknownTool)
+        return FrictionKind.UNKNOWN_TOOL;
+    if (facts.verifiedUnknown)
+        return FrictionKind.UNKNOWN_VERDICT;
+    // Repetition is only friction when it is repetition WITHOUT PROGRESS. Found by driving the Tauri
+    // smoke app: three consecutive SUCCESSFUL act_and_wait calls, each producing a verdict, were
+    // answered with "stuck on the same call?" — nagging the exact loop this release exists to
+    // encourage. A verdict is progress.
+    if (facts.repeatRun >= REPEAT_INVITE_AT && !facts.producedVerdict)
+        return FrictionKind.REPEATING;
+    if (facts.errored)
+        return FrictionKind.REFUSED;
+    return undefined;
+}
+/** The line for a friction kind. */
+export function inviteFor(kind) {
+    return LINES[kind];
+}

package/dist/tools/feedback-tools.d.ts

@@ -12,7 +12,18 @@ export declare const FEEDBACK_PROMPT: {
     readonly ask: "Ask the human, in your own words and only once: how did this verification go, and what would have made Reticle more useful? If they answer, file it with `reticle feedback` (they run it) or relay it yourself.";
     readonly command: "reticle feedback";
 };
-/** Tools whose completion means "a verification just happened" — the only moment worth asking at. */
+/**
+ * Tools whose completion means "a verification just happened" — the only moment worth asking at,
+ * and the set that decides what `verification_completed` counts.
+ *
+ * ACT_AND_WAIT belongs here and was missing for a long time. It declares `verified` in its
+ * outputSchema, returns the whole verdict block, and its own description calls it "one hop for the
+ * act->observe->assert loop" — it IS the verification path most agents take. The guard that should
+ * have caught the omission keyed on the tool NAME (/assert|verify/), which act_and_wait does not
+ * match, so it slipped through in silence. Measured over a day of real telemetry: act_and_wait 14
+ * calls, assert ZERO, verification_completed 2. Agents were verifying the entire time and the metric
+ * could not see it. The contract test now checks the SHAPE instead.
+ */
 export declare const VERIFICATION_TOOLS: ReadonlySet<string>;
 /**
  * The human prompt, when one is due. Resettable for tests.

package/dist/tools/feedback-tools.js

@@ -30,6 +30,9 @@ function sessionFacts(deps, sessionId) {
         return {
             ...(session.runtime !== undefined ? { runtime: session.runtime } : {}),
             ...(session.engine !== undefined ? { engine: session.engine } : {}),
+            // The app's own account of what it is. Without this the reported stack came from the daemon's
+            // working directory, so feedback from one app was filed against another's framework.
+            ...(0 < session.adapters.length ? { adapters: session.adapters } : {}),
             driven,
         };
     }
@@ -83,7 +86,10 @@ export const FEEDBACK_TOOLS = [
                 .describe('Only to attach a specific tab s runtime/engine context. Omit it — a report about a session that never connected is still worth filing.'),
         },
         outputSchema: {
+            /** Confirmed delivery. Always false here: the agent path queues and returns. */
             sent: z.boolean(),
+            /** Validated, redacted and queued. This is the one to read. */
+            accepted: z.boolean(),
             reason: z.string().optional(),
             redacted: z.array(z.string()),
             context: z.unknown(),
@@ -106,12 +112,17 @@ export const FEEDBACK_TOOLS = [
                 ...optionalText(args, 'impact'),
                 ...optionalText(args, 'currentApproach'),
                 ...optionalText(args, 'model'),
-            }, { session: sessionFacts(deps, asString(args['sessionId'])) });
+            }, {
+                session: sessionFacts(deps, asString(args['sessionId'])),
+                // The AGENT does not wait for the network. `reticle feedback` (a human, at a terminal,
+                // watching for an answer) still does — see submitFeedback's `background`.
+                background: true,
+            });
             return {
                 ...receipt,
-                note: receipt.sent
-                    ? 'Filed — thank you. Continue your task; do not retry a failing call just to re-report it.'
-                    : 'NOT sent (see reason). Tell the human what you found so it is not lost.',
+                note: receipt.accepted
+                    ? 'Accepted — thank you. It sends in the background so you are not kept waiting; if it fails you will be told on a later tool result. Continue your task; do not retry a failing call just to re-report it.'
+                    : 'NOT accepted (see reason). Tell the human what you found so it is not lost.',
             };
         },
     },
@@ -128,8 +139,20 @@ export const FEEDBACK_PROMPT = {
     ask: 'Ask the human, in your own words and only once: how did this verification go, and what would have made Reticle more useful? If they answer, file it with `reticle feedback` (they run it) or relay it yourself.',
     command: 'reticle feedback',
 };
-/** Tools whose completion means "a verification just happened" — the only moment worth asking at. */
+/**
+ * Tools whose completion means "a verification just happened" — the only moment worth asking at,
+ * and the set that decides what `verification_completed` counts.
+ *
+ * ACT_AND_WAIT belongs here and was missing for a long time. It declares `verified` in its
+ * outputSchema, returns the whole verdict block, and its own description calls it "one hop for the
+ * act->observe->assert loop" — it IS the verification path most agents take. The guard that should
+ * have caught the omission keyed on the tool NAME (/assert|verify/), which act_and_wait does not
+ * match, so it slipped through in silence. Measured over a day of real telemetry: act_and_wait 14
+ * calls, assert ZERO, verification_completed 2. Agents were verifying the entire time and the metric
+ * could not see it. The contract test now checks the SHAPE instead.
+ */
 export const VERIFICATION_TOOLS = new Set([
+    ReticleTool.ACT_AND_WAIT,
     ReticleTool.ASSERT,
     ReticleTool.FLOW_VERIFY,
     ReticleTool.VERIFY_CHANGE,

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

@@ -1,10 +1,4 @@
 import type { ToolDef, ToolDeps } from './tools.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
- * can never return a healthy-looking result from any of these. `runTool` is the single choke point
- * (mcp.ts + tool-invoker.ts) that splices health on; the guard test asserts the set is exhaustive.
- */
 export declare const SESSION_BOUND_TOOLS: ReadonlySet<string>;
 /**
  * Tools that carry a `sessionId` arg but are NOT live-session-health tools — they read/write
@@ -13,10 +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>;