@reticlehq/server 2.4.0 → 2.6.0

package/dist/tools/browser-tools.js

@@ -1,4 +1,8 @@
 import { z } from 'zod';
+import { navigateResult } from './navigate-result.js';
+import { awaitArrival } from './navigate-arrival.js';
+import { reloadResult } from './reload-result.js';
+import { waitForReconnect, RELOAD_RECONNECT_TIMEOUT_MS } from '../session/session-reconnect.js';
 import { ReticleCommand } from '@reticlehq/core';
 import { ReticleTool } from './tool-names.js';
 import { asString } from './tools-helpers.js';
@@ -7,7 +11,7 @@ export const BROWSER_TOOLS = [
     {
         name: ReticleTool.NAVIGATE,
         example: { url: '/settings' },
-        description: 'Navigate the connected browser tab to a URL, or reload it in place with { reload: true } (add { hard: true } to bypass the cache). The SDK reconnects automatically after the page loads. Use reticle_sessions to confirm the new tab is connected before acting.',
+        description: 'Navigate the connected browser tab to a URL, or reload it in place with { reload: true } (add { hard: true } to bypass the cache). `ok` means the navigation was DISPATCHED — the SDK is torn down by the navigation, so the page itself cannot report on it. The daemon then waits briefly for the SDK to reconnect: `confirmed:true` with a new `sessionId` means the page arrived and you can act immediately. `confirmed:false` means it did not arrive within the window — the page may be slow, uninstrumented, or not there; check reticle_sessions before acting.',
         inputSchema: {
             url: z.string().optional().describe('The URL to navigate to. Omit when using reload.'),
             reload: z
@@ -24,17 +28,38 @@ export const BROWSER_TOOLS = [
             ok: z.boolean(),
             url: z.string().optional(),
             reason: z.string().optional(),
+            // Present on a dispatched navigation: nothing here can see the new document, so arrival is
+            // reported as unconfirmed rather than implied by `ok`. See navigate-result.ts.
+            confirmed: z.boolean().optional(),
+            /** The session the SDK reconnected as, when arrival was confirmed — it is a NEW id. */
+            sessionId: z.string().optional(),
+            note: z.string().optional(),
         },
         handler: async (deps, args) => {
             // reload:true is the absorbed reticle_refresh — same command, one fewer advertised tool.
-            if (args['reload'] === true) {
+            if (true === args['reload']) {
+                const before = deps.sessions.resolve(asString(args['sessionId']));
                 await commandOrThrow(deps, asString(args['sessionId']), ReticleCommand.REFRESH, {
-                    hard: args['hard'] === true,
+                    hard: true === args['hard'],
                 });
-                return { ok: true };
+                // WAIT for the page to come back, rather than telling the agent to. The id survives the
+                // reload, but the seconds between dispatch and the new HELLO are seconds in which every call
+                // lands in the old, disconnected session — measured as reticle_run failing 5 of 5 on a page
+                // that was healthy immediately afterwards. Returns on the first poll in the common case.
+                const back = await waitForReconnect({
+                    current: () => deps.sessions.get(before.id),
+                    previous: before,
+                    timeoutMs: RELOAD_RECONNECT_TIMEOUT_MS,
+                    now: deps.now,
+                    sleep: (ms) => new Promise((resolve) => setTimeout(resolve, ms)),
+                });
+                // Not a bare `{ ok: true }`. The URL branch below already discloses that `ok` means
+                // DISPATCHED — the reload branch had identical semantics and said nothing, on the path most
+                // likely to need it. See reload-result.
+                return reloadResult(back);
             }
             const url = asString(args['url']);
-            if (url === undefined || url.length === 0)
+            if (url === undefined || 0 === url.length)
                 return { ok: false, reason: 'url required' };
             // Record navigate as an action. Its window is usually empty (the page unloads and the SDK
             // reconnects), but the action record itself — "navigated to X" — is the causal fact worth keeping.
@@ -42,11 +67,11 @@ export const BROWSER_TOOLS = [
             session.beginAction(ReticleTool.NAVIGATE, { url });
             try {
                 const result = (await commandOrThrow(deps, asString(args['sessionId']), ReticleCommand.NAVIGATE, { url }));
-                return {
-                    ok: result.ok === true,
-                    ...(typeof result.url === 'string' ? { url: result.url } : {}),
-                    ...(typeof result.reason === 'string' ? { reason: result.reason } : {}),
-                };
+                // `ok` is the browser accepting the instruction, not the page arriving — see navigate-result.
+                // The daemon is the only party that CAN see arrival (the SDK reconnects to it), so it looks,
+                // briefly, instead of telling the agent to go poll reticle_sessions itself.
+                const arrival = true === result.ok ? await awaitArrival(deps.sessions, url) : null;
+                return navigateResult(result, arrival);
             }
             finally {
                 session.finishAction();
@@ -68,7 +93,7 @@ export const BROWSER_TOOLS = [
         },
         handler: async (deps, args) => {
             await commandOrThrow(deps, asString(args['sessionId']), ReticleCommand.REFRESH, {
-                hard: args['hard'] === true,
+                hard: true === args['hard'],
             });
             return { ok: true };
         },

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

@@ -1,7 +1,2 @@
 import type { ToolDef } from './tools.js';
-/**
- * The capability-contract tools. `reticle_capabilities` reads the live session, or the
- * git-checked `.reticle/contract.json` when `{ fromDisk:true }`; `reticle_contract_save` persists the
- * live registry to that file (pretty-printed, stable key order — diffable in PRs).
- */
 export declare const CONTRACT_TOOLS: ToolDef[];

package/dist/tools/contract-tools.js

@@ -9,6 +9,31 @@ import { reticleDirPaths, readContract, writeContract } from '../project/reticle
  * git-checked `.reticle/contract.json` when `{ fromDisk:true }`; `reticle_contract_save` persists the
  * live registry to that file (pretty-printed, stable key order — diffable in PRs).
  */
+/**
+ * Refuse to persist one project's surface into another project's repo.
+ *
+ * `writeContract` writes to the DAEMON's `.reticle/`, which is the directory the daemon was started
+ * in — not the session's project. A daemon started above several apps therefore saves app A's
+ * testids into app B's checkout, reports `saved: true`, and hands back a path that looks right. A
+ * contract is git-checked and diffable, so that is a change somebody commits.
+ *
+ * Refuse rather than redirect: a session advertises a projectId, not a project DIRECTORY, so there
+ * is no honest way to work out where the correct `.reticle/` lives. Naming both ids and stopping is
+ * the whole of what can be said truthfully.
+ *
+ * Silence is not a mismatch. An unstamped build carries no projectId, and a daemon above an
+ * uninitialised directory has none either; neither is evidence of two projects.
+ */
+function assertSameProject(deps, sessionId) {
+    const mine = deps.projectId;
+    const theirs = deps.sessions.resolve(sessionId).projectId;
+    if (mine === undefined || theirs === undefined || mine === theirs)
+        return;
+    throw new Error(`refusing to save: this session belongs to project '${theirs}', but this daemon writes to ` +
+        `project '${mine}' (.reticle/ in the directory it was started in). Saving would put one ` +
+        "app's testable surface into another app's git-checked contract. Run a daemon from " +
+        `'${theirs}'s own directory and save there.`);
+}
 export const CONTRACT_TOOLS = [
     {
         name: ReticleTool.CAPABILITIES,
@@ -29,7 +54,7 @@ export const CONTRACT_TOOLS = [
             governance: z.unknown().optional(),
         },
         handler: async (deps, args) => {
-            if (args[FROM_DISK_ARG] === true) {
+            if (true === args[FROM_DISK_ARG]) {
                 const r = await readContract(deps.fs, deps.reticleRoot);
                 if (!r.ok)
                     throw new Error(r.reason === ContractReadError.MISSING
@@ -54,6 +79,7 @@ export const CONTRACT_TOOLS = [
         handler: async (deps, args) => {
             const res = await commandOrThrow(deps, asString(args['sessionId']), ReticleCommand.CAPABILITIES, {});
             const caps = CapabilitiesSchema.parse(res);
+            assertSameProject(deps, asString(args['sessionId']));
             await writeContract(deps.fs, deps.reticleRoot, caps, deps.now);
             return {
                 saved: true,

package/dist/tools/coverage-identity.d.ts

@@ -0,0 +1,34 @@
+/**
+ * Which controls has the agent actually driven — across re-renders.
+ *
+ * Coverage was counted by REF, and a ref is invalidated whenever the DOM changes. Reported from a
+ * sweep of a Next app-router project: `total: 34, exercised: 0` after four successful acts. App
+ * router replaces a whole route segment on navigation, so every control the agent drove is a new
+ * element by the time coverage is asked. The number is then 0 forever on any framework that replaces
+ * nodes rather than reconciling in place — and an agent told it has exercised none of 34 controls
+ * re-drives ground it already covered, which is the exact opposite of what this tool is for.
+ *
+ * Bench-app hid it: it reconciles in place, so refs survive there and the number looked right.
+ *
+ * The LABEL is the identity that survives a re-render — `button "Deploy"` names the same control
+ * before and after — and it is already parsed for the untouched list.
+ *
+ * One caveat, stated because it is a real limitation: two controls sharing a label are
+ * indistinguishable here, so driving one marks both. That errs toward "covered", which is the
+ * unsafe direction for a coverage number — but it is bounded to same-labelled controls, and the
+ * alternative it replaces reported ZERO coverage on entire frameworks.
+ */
+interface CoverageControl {
+    ref: string;
+    label: string;
+}
+interface CoverageTally {
+    /** Controls on the page the agent has driven, by ref or by surviving label. */
+    exercised: number;
+    /** Controls the agent drove that are on the page no longer, and no label explains. */
+    droveGone: number;
+    /** The controls still untouched, in document order. */
+    untouched: CoverageControl[];
+}
+export declare function exercisedCount(controls: readonly CoverageControl[], actedRefs: ReadonlySet<string>, actedLabels: ReadonlySet<string>): CoverageTally;
+export {};

package/dist/tools/coverage-identity.js

@@ -0,0 +1,46 @@
+/**
+ * Which controls has the agent actually driven — across re-renders.
+ *
+ * Coverage was counted by REF, and a ref is invalidated whenever the DOM changes. Reported from a
+ * sweep of a Next app-router project: `total: 34, exercised: 0` after four successful acts. App
+ * router replaces a whole route segment on navigation, so every control the agent drove is a new
+ * element by the time coverage is asked. The number is then 0 forever on any framework that replaces
+ * nodes rather than reconciling in place — and an agent told it has exercised none of 34 controls
+ * re-drives ground it already covered, which is the exact opposite of what this tool is for.
+ *
+ * Bench-app hid it: it reconciles in place, so refs survive there and the number looked right.
+ *
+ * The LABEL is the identity that survives a re-render — `button "Deploy"` names the same control
+ * before and after — and it is already parsed for the untouched list.
+ *
+ * One caveat, stated because it is a real limitation: two controls sharing a label are
+ * indistinguishable here, so driving one marks both. That errs toward "covered", which is the
+ * unsafe direction for a coverage number — but it is bounded to same-labelled controls, and the
+ * alternative it replaces reported ZERO coverage on entire frameworks.
+ */
+export function exercisedCount(controls, actedRefs, actedLabels) {
+    // An empty label is not an identity — anonymous controls would otherwise collapse into one bucket
+    // and a single act would mark every one of them driven.
+    // Matched on the whole label (`button "Save"`) OR on a testid the label contains. The snapshot
+    // renders a control's testid inside its label, and a testid is the identity most likely to survive
+    // a re-render — matching only the full label meant a control with a testid but no accessible name
+    // was never recognised, and coverage read `exercised: 0` after real work.
+    const droveLabel = (label) => {
+        if (0 === label.length)
+            return false;
+        if (actedLabels.has(label))
+            return true;
+        for (const acted of actedLabels) {
+            if (acted.length > 0 && !acted.includes('"') && label.includes(acted))
+                return true;
+        }
+        return false;
+    };
+    const untouched = controls.filter((control) => !actedRefs.has(control.ref) && !droveLabel(control.label));
+    // A driven ref that is not on the page is only "gone" if no present control carries its label —
+    // otherwise it is the same control after a re-render, which is the case this exists to fix.
+    const presentRefs = new Set(controls.map((control) => control.ref));
+    const explainedByLabel = new Set(controls.filter((control) => droveLabel(control.label)).map((control) => control.label));
+    const droveGone = [...actedRefs].filter((ref) => !presentRefs.has(ref) && 0 === explainedByLabel.size).length;
+    return { exercised: controls.length - untouched.length, droveGone, untouched };
+}

package/dist/tools/coverage-tools.js

@@ -2,6 +2,7 @@ import { z } from 'zod';
 import { ReticleCommand, SnapshotMode } from '@reticlehq/core';
 import { ReticleTool } from './tool-names.js';
 import { asString } from './tools-helpers.js';
+import { exercisedCount } from './coverage-identity.js';
 import { commandOrThrow, sessionIdShape } from './tool-kit.js';
 /**
  * `reticle_coverage` — which interactive controls this session has driven, and which it has not.
@@ -28,7 +29,7 @@ export function parseControls(tree) {
     const seen = new Set();
     for (const line of tree.split('\n')) {
         const match = LINE_WITH_REF.exec(line);
-        if (match === null)
+        if (null === match)
             continue;
         const [, label, ref] = match;
         if (label === undefined || ref === undefined || seen.has(ref))
@@ -71,21 +72,16 @@ export const COVERAGE_TOOLS = [
                 mode: SnapshotMode.INTERACTIVE,
             });
             const tree = asString(result['tree']) ?? '';
-            const acted = session.actedRefs();
-            const controls = parseControls(tree);
-            const present = new Set(controls.map((control) => control.ref));
-            const untouched = controls.filter((control) => !acted.has(control.ref));
-            // Drives that landed on controls no longer in the DOM. Without this the number is biased
-            // against the actions that WORKED: archive, delete, submit and navigate all remove their own
-            // control, so clicking one and asking again reported `exercised: 0` — measured on the Electron
-            // demo right after a successful archive. An agent reading "0" concludes it has done nothing
-            // and re-drives ground it already covered.
-            const drovenGone = [...acted].filter((ref) => !present.has(ref)).length;
+            // Matched by ref AND by label — see coverage-identity. A ref dies with the next re-render, so
+            // on a framework that replaces nodes this reported `exercised: 0` however much work was done.
+            // `droveGone` keeps the other honesty: archive/delete/submit remove their own control, so a
+            // drive that WORKED must not read as no coverage at all.
+            const { exercised, droveGone, untouched } = exercisedCount(parseControls(tree), session.actedRefs(), session.actedLabels());
             return {
-                total: controls.length,
-                exercised: controls.length - untouched.length,
+                total: parseControls(tree).length,
+                exercised,
                 untouched,
-                ...(drovenGone > 0 ? { alsoDroveGone: drovenGone } : {}),
+                ...(droveGone > 0 ? { alsoDroveGone: droveGone } : {}),
             };
         },
     },

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

@@ -1,6 +1,7 @@
 import type { ToolDef } from './tools.js';
+import { type ToolSurfaceOrigin } from './tool-surface.js';
 /**
  * Build the two dynamic meta-tools over the full tool table. `reticle_run` dispatches through the same
  * `runTool` chokepoint as a direct call, so session-health splicing and every other invariant hold.
  */
-export declare function buildDynamicTools(allTools: ToolDef[]): ToolDef[];
+export declare function buildDynamicTools(allTools: ToolDef[], profile?: ToolSurfaceOrigin): ToolDef[];

package/dist/tools/dynamic-tools.js

@@ -1,6 +1,9 @@
 import { z } from 'zod';
 import { runTool } from './invoke-tool.js';
+import { buildErrorPayload } from './error-recovery.js';
+import { mergedNameRedirect, mergedNameMessage } from './merged-name-redirect.js';
 import { ReticleTool } from './tool-names.js';
+import { TOOL_PROFILE_ENV } from './tool-surface.js';
 import { getSessionMetrics } from '../telemetry/session-metrics.js';
 /**
  * On-demand tool loading for MCP — the answer to the per-turn tool-definition tax.
@@ -36,12 +39,39 @@ function paramInfo(shape) {
         description: schema.description ?? '',
     }));
 }
+/**
+ * The one wording for "you named a parameter that does not exist", shared by both checks in
+ * `reticle_run` — the one over the WRAPPED tool's `args`, and the one over reticle_run's own
+ * top-level keys. Two spellings of the same refusal is how one of them ends up not existing.
+ */
+function unknownParamsError(toolName, unknown) {
+    return `unknown ${1 === unknown.length ? 'parameter' : 'parameters'} for ${toolName}: ${unknown.join(', ')} — NOT applied, so any result would be an answer to a different question`;
+}
+/** The keys not declared by `shape`, in call order. */
+function unknownKeys(args, shape) {
+    const declared = new Set(Object.keys(shape));
+    return Object.keys(args).filter((key) => !declared.has(key));
+}
 /**
  * Build the two dynamic meta-tools over the full tool table. `reticle_run` dispatches through the same
  * `runTool` chokepoint as a direct call, so session-health splicing and every other invariant hold.
  */
-export function buildDynamicTools(allTools) {
+export function buildDynamicTools(allTools, profile) {
     const byName = new Map(allTools.map((t) => [t.name, t]));
+    // The profile is a DAEMON-startup decision, so an agent that exported RETICLE_TOOL_PROFILE into its
+    // own environment sees no change and has, until now, no way to tell. Reported with the catalog.
+    const profileBlock = profile === undefined
+        ? {}
+        : {
+            profile: {
+                ...profile,
+                // NOT "every tool is reachable via reticle_run under any profile" — measured false:
+                // `full` advertises all 46 tools DIRECTLY and carries no meta-tools at all, so
+                // reticle_run does not exist there. Saying otherwise sends the agent to a tool the
+                // profile does not have.
+                note: `The profile is read once at daemon startup — change ${TOOL_PROFILE_ENV} and restart the daemon, or it has no effect. Under this profile every unadvertised tool is reachable through reticle_run; under \`full\` there are no meta-tools because every tool is advertised directly.`,
+            },
+        };
     const reticleTools = {
         name: ReticleTool.TOOLS,
         description: 'Discover Reticle tools on demand. Call with no arguments to list every tool (name + one-line summary); call with names:["reticle_network", …] to load full descriptions and parameters for specific tools. Then invoke them with reticle_run. This avoids paying for every tool definition on every turn. To make a verification REUSABLE (record once, replay free forever), the flow workflow lives here: reticle_record{action:"start"} → act → reticle_flow_save → reticle_flow_verify (and reticle_flow_heal on drift). Load those names when you want to save or re-run a flow.',
@@ -53,11 +83,12 @@ export function buildDynamicTools(allTools) {
         },
         handler: (_deps, args) => {
             const names = Array.isArray(args['names'])
-                ? args['names'].filter((n) => typeof n === 'string')
+                ? args['names'].filter((n) => 'string' === typeof n)
                 : undefined;
-            if (names === undefined || names.length === 0) {
+            if (names === undefined || 0 === names.length) {
                 return Promise.resolve({
                     tools: allTools.map((t) => ({ name: t.name, summary: firstSentence(t.description) })),
+                    ...profileBlock,
                     next: 'Load params with reticle_tools { names:[…] }, then call reticle_run { tool, args }.',
                 });
             }
@@ -71,35 +102,81 @@ export function buildDynamicTools(allTools) {
             });
         },
     };
+    const runShape = {
+        tool: z.string().describe('Tool name to invoke, e.g. reticle_network.'),
+        args: z.record(z.unknown()).optional().describe('Arguments object for that tool.'),
+        // Accepted AND FORWARDED, not merely tolerated. reticle_run is the only way to reach an
+        // unadvertised tool, so on a machine running several projects it has to be aimable — and
+        // `sessionId` is the shape an agent already uses on every other tool. Reported across 6 of 6
+        // apps: it took this key, dropped it, resolved by the daemon's cwd project, and failed with
+        // "no browser session for project X" while naming the very session it had been given.
+        sessionId: z
+            .string()
+            .optional()
+            .describe('Target tab for the invoked tool. Forwarded to it — an explicit args.sessionId wins.'),
+    };
     const reticleRun = {
         name: ReticleTool.RUN,
         description: "Invoke any Reticle tool by name (discover names/params first with reticle_tools). On an unknown tool or bad arguments it returns the available names or the tool's params, so you can correct and retry.",
-        inputSchema: {
-            tool: z.string().describe('Tool name to invoke, e.g. reticle_network.'),
-            args: z.record(z.unknown()).optional().describe('Arguments object for that tool.'),
-        },
+        inputSchema: runShape,
         handler: async (deps, args) => {
-            const name = typeof args['tool'] === 'string' ? args['tool'] : '';
-            const callArgs = typeof args['args'] === 'object' && args['args'] !== null
+            // reticle_run's OWN parameters, checked first and with the same wording as the inner check.
+            // It refused an unknown key inside `args` and silently dropped one beside them: an agent that
+            // wrote `reticle_run { tool, args, sessionId }` — the shape it uses on every other tool — had
+            // its sessionId ignored and got a confident answer about whichever session auto-selection
+            // picked. Every other tool on the surface refuses this; the escape hatch has to as well.
+            const strayTop = unknownKeys(args, runShape);
+            if (strayTop.length > 0) {
+                return {
+                    error: unknownParamsError(ReticleTool.RUN, strayTop),
+                    hint: `pass them inside args: reticle_run { tool, args: { ${strayTop.join(', ')}: … } } if the target tool declares them`,
+                };
+            }
+            const name = 'string' === typeof args['tool'] ? args['tool'] : '';
+            const given = 'object' === typeof args['args'] && args['args'] !== null
                 ? args['args']
                 : {};
+            const aimed = args['sessionId'];
             const target = byName.get(name);
             if (target === undefined) {
                 // A non-zero count here means our advertised surface is confusing the agent — it reached for
                 // something it believed existed. That is a docs/naming defect, and it was invisible.
-                getSessionMetrics().recordUnknownTool();
+                //
+                // The NAME goes with it: the count says the surface confused someone, the name says which
+                // capability they expected. It was already in hand here and thrown away, so the one place
+                // the product could learn what agents want was discarding it. Safe — a name from our own
+                // tool namespace, never app data.
+                getSessionMetrics().recordUnknownTool(name);
+                // ...and for 22 of those names the capability DOES exist, it just moved when tools merged.
+                // Answering "unknown tool" there is simply wrong; see merged-name-redirect.
+                const moved = mergedNameRedirect(name);
+                if (moved !== undefined) {
+                    return {
+                        error: mergedNameMessage(name, moved),
+                        tool: moved.tool,
+                        ...(moved.action === undefined ? {} : { action: moved.action }),
+                    };
+                }
                 return { error: `unknown tool '${name}'`, available: allTools.map((t) => t.name) };
             }
+            // The outer aim, forwarded ONLY to a tool that declares a session — injecting it into one that
+            // does not would trip the unknown-key check below and refuse a call the caller got right. An
+            // inner args.sessionId wins as the more specific instruction, and an absent aim stays absent so
+            // auto-selection still applies.
+            const callArgs = 'string' === typeof aimed &&
+                given['sessionId'] === undefined &&
+                undefined !== target.inputSchema['sessionId']
+                ? { ...given, sessionId: aimed }
+                : given;
             // The escape hatch is where a typo is MOST likely, because an unadvertised tool's parameters
             // are not in front of the agent at all — and it is the one path the SDK's own validation never
             // sees, since `reticle_run`'s own args (`tool`, `args`) are perfectly valid. Left unchecked,
             // `reticle_run { tool: "reticle_clock", args: { action: "freeze" } }` returned
             // `{"frozen":false}`: a well-formed answer to a question nobody asked.
-            const declared = new Set(Object.keys(target.inputSchema));
-            const unknown = Object.keys(callArgs).filter((key) => !declared.has(key));
+            const unknown = unknownKeys(callArgs, target.inputSchema);
             if (unknown.length > 0) {
                 return {
-                    error: `unknown ${unknown.length === 1 ? 'parameter' : 'parameters'} for ${name}: ${unknown.join(', ')} — NOT applied, so any result would be an answer to a different question`,
+                    error: unknownParamsError(name, unknown),
                     tool: name,
                     params: paramInfo(target.inputSchema),
                     ...(target.example === undefined ? {} : { example: target.example }),
@@ -110,12 +187,12 @@ export function buildDynamicTools(allTools) {
                 return await runTool(target, deps, callArgs);
             }
             catch (error) {
-                return {
-                    error: error instanceof Error ? error.message : String(error),
-                    tool: name,
-                    params: paramInfo(target.inputSchema),
-                    hint: 'fix the arguments and call reticle_run again',
-                };
+                // The SAME payload the direct tool path builds — recovery included. Answering every failure
+                // with "fix the arguments" threw that away, and under the default profile nearly every tool
+                // is reached through here: a stale ref, a paused session, a missing pairing token all came
+                // back as the agent's arguments being wrong, which is advice that spends the retry.
+                const message = error instanceof Error ? error.message : String(error);
+                return { ...buildErrorPayload(message), tool: name };
             }
         },
     };

package/dist/tools/error-recovery.d.ts

@@ -17,11 +17,43 @@ export declare const RECOVERY: {
     readonly MISSING_BASELINE: string;
     readonly MISSING_RECORDING: string;
     readonly STALE_REF: string;
+    readonly INVALID_NAME: string;
+    readonly NO_SUCH_OPTION: string;
     readonly BAD_ARGUMENTS: string;
+    readonly WRONG_TARGET: string;
+    readonly NOT_EDITABLE: string;
+    readonly CONFIRM_DANGEROUS: string;
+    readonly UNSUPPORTED_SURFACE: string;
     readonly TOKEN_REQUIRED: string;
+    readonly SCOPE_MISMATCH: string;
+    readonly SESSION_GONE: string;
+    readonly COMMAND_TIMEOUT: string;
+    readonly FLOW_STEP_MISSING: string;
+    readonly NO_POOL: string;
 };
-/** 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 declare function recoveryFor(message: string): string | undefined;
+/**
+ * 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 declare function zodArrayAsSentence(message: string): string;
 /**
  * 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
@@ -43,5 +75,5 @@ interface ErrorPayload {
  * 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 declare function buildErrorPayload(message: string): ErrorPayload;
+export declare function buildErrorPayload(rawMessage: string): ErrorPayload;
 export {};