@nodetool-ai/agents 0.7.0-rc.23 → 0.8.1

package/dist/evals/codeact-api-surfaces.js

@@ -0,0 +1,2191 @@
+/**
+ * CodeAct eval cases for the `nodetool.*` object model — data and creative
+ * namespaces: collections, memory, shared, web, documents, email, apps,
+ * games, timelines, sketches, scripts, storyboards.
+ *
+ * The fakes here are named EXACTLY like the belt tools those namespaces wrap
+ * (`NODETOOL_API_NAMESPACE_TOOLS`), so the real guest prelude loads and every
+ * `nodetool.<ns>.<method>()` call lands on one of them with the argument names
+ * the prelude sends. The world behind them is small, in-memory, deterministic
+ * and internally consistent: an indexed chunk becomes findable, a voiced line
+ * becomes assemblable, a rendered still unblocks a clip, a planted validation
+ * issue is fixable by an edit. Cases score the outcome, never one working
+ * order.
+ */
+import { applyGameOps, createTopDownRoomGame, gameDocumentOp } from "@nodetool-ai/game-runtime";
+import { RecordingTool } from "./codeact-cases.js";
+import { isFiniteNumber, isNumber, isRecord, isString } from "../utils/type-guards.js";
+// ---------------------------------------------------------------------------
+// param coercion
+// ---------------------------------------------------------------------------
+const str = (value, fallback = "") => isString(value) ? value : fallback;
+const num = (value, fallback) => isFiniteNumber(value) ? value : fallback;
+const rec = (value) => isRecord(value)
+    ? value
+    : {};
+const recList = (value) => Array.isArray(value) ? value.map(rec) : [];
+const strList = (value) => {
+    if (isString(value))
+        return [value];
+    if (!Array.isArray(value))
+        return [];
+    return value.filter((entry) => typeof entry === "string");
+};
+/** Words, with a dotted run kept whole so "4.2" and "example.dev" survive. */
+const words = (text) => text.toLowerCase().match(/[a-z0-9]+(?:\.[a-z0-9]+)*/g) ?? [];
+/** Fake relevance: how many distinct query words the text repeats. */
+const overlap = (query, text) => {
+    const haystack = new Set(words(text));
+    let score = 0;
+    for (const word of new Set(words(query))) {
+        if (haystack.has(word))
+            score++;
+    }
+    return score;
+};
+const clone = (value) => JSON.parse(JSON.stringify(value));
+const obj = (properties, required = []) => required.length > 0
+    ? { type: "object", properties, required }
+    : { type: "object", properties };
+const S = { type: "string" };
+const N = { type: "number" };
+const B = { type: "boolean" };
+const ANY_ARRAY = { type: "array" };
+const HANDBOOK = [
+    {
+        id: "policy-refunds",
+        text: "The refund window is 30 days. A refund is issued to the original card."
+    },
+    {
+        id: "policy-shipping",
+        text: "Standard shipping takes 5 business days after dispatch."
+    },
+    {
+        id: "policy-warranty",
+        text: "Hardware carries a 24 month warranty against manufacturing defects."
+    }
+];
+const WEB_PAGES = [
+    {
+        url: "https://example.dev/nodetool/release-notes",
+        title: "NodeTool 4.2 release notes",
+        snippet: "Everything that shipped in the NodeTool 4.2 release.",
+        text: "NodeTool 4.2 release notes. Agent actions now run inside a QuickJS " +
+            "sandbox. The release ships 58 packages and one desktop app."
+    },
+    {
+        url: "https://example.dev/nodetool/pricing",
+        title: "NodeTool pricing",
+        snippet: "Plans and per-seat prices.",
+        text: "The team plan costs 20 euro per seat per month."
+    },
+    {
+        url: "https://example.dev/blog/codeact",
+        title: "Why CodeAct",
+        snippet: "Acting by writing code instead of JSON tool calls.",
+        text: "CodeAct lets one action chain many tool calls in a loop."
+    }
+];
+const PDF_TEXT = "Q3 revenue reached 4.2 million euro across 1,280 accounts. " +
+    "Churn held at 1.1 percent.";
+/** Fresh, isolated world state — one per `createSurfaceApiTools` call. */
+function createWorld() {
+    return {
+        indexed: [],
+        memories: [
+            {
+                memory_id: "mem_seed_1",
+                title: "workspace layout",
+                content: "Reports live under reports/ in the workspace.",
+                kind: "note",
+                resources: []
+            }
+        ],
+        memorySeq: 1,
+        // Two past conversations. What was actually generated is only in a tool
+        // call's arguments, which the thread listing summarizes away — so the
+        // case has to read the message itself rather than the transcript.
+        threads: [
+            {
+                id: "th_banner",
+                title: "Launch banner images",
+                updated_at: "2026-02-10T09:00:00.000Z"
+            },
+            {
+                id: "th_refunds",
+                title: "Refund policy wording",
+                updated_at: "2026-01-05T09:00:00.000Z"
+            }
+        ],
+        chatMessages: [
+            {
+                id: "msg_refunds_1",
+                thread_id: "th_refunds",
+                role: "user",
+                text: "How long is the refund window?",
+                created_at: "2026-01-05T08:59:00.000Z"
+            },
+            {
+                id: "msg_refunds_2",
+                thread_id: "th_refunds",
+                role: "assistant",
+                text: "Thirty days from delivery.",
+                created_at: "2026-01-05T09:00:00.000Z"
+            },
+            {
+                id: "msg_banner_1",
+                thread_id: "th_banner",
+                role: "user",
+                text: "Make the launch banner.",
+                created_at: "2026-02-10T08:59:00.000Z"
+            },
+            {
+                id: "msg_banner_2",
+                thread_id: "th_banner",
+                role: "assistant",
+                text: "Generated the banner.",
+                created_at: "2026-02-10T09:00:00.000Z",
+                tool_calls: [
+                    {
+                        id: "call_1",
+                        name: "generate_image",
+                        args: { model: "fal_ai/flux/dev", prompt: "launch banner" }
+                    }
+                ]
+            }
+        ],
+        // Run-scoped agent memory, seeded with what upstream steps left behind:
+        // the numbers are only here, so the case has to read them.
+        shared: [
+            {
+                key: "task:pricing",
+                kind: "task_result",
+                title: "Pricing research",
+                value: { tier: "pro", monthly_usd: 49 }
+            },
+            {
+                key: "step:draft",
+                kind: "step_result",
+                title: "Announcement draft",
+                value: "The pro tier is our best value."
+            }
+        ],
+        emails: [
+            {
+                message_id: "msg_1",
+                subject: "Invoice 8841 due",
+                from: "billing@vendor.example",
+                hours_ago: 3,
+                labels: [],
+                archived: false
+            },
+            {
+                message_id: "msg_2",
+                subject: "Invoice 8842 due",
+                from: "billing@vendor.example",
+                hours_ago: 11,
+                labels: [],
+                archived: false
+            },
+            {
+                message_id: "msg_3",
+                subject: "Team lunch on Friday",
+                from: "office@acme.example",
+                hours_ago: 5,
+                labels: [],
+                archived: false
+            },
+            {
+                message_id: "msg_4",
+                subject: "Invoice 8840 paid",
+                from: "billing@vendor.example",
+                hours_ago: 100,
+                labels: [],
+                archived: false
+            }
+        ],
+        apps: new Set(["app_notes"]),
+        timelines: new Map([
+            [
+                "tl_promo",
+                {
+                    id: "tl_promo",
+                    name: "Promo cut",
+                    tracks: [{ id: "track_video", type: "video", name: "Video" }],
+                    clips: [
+                        {
+                            id: "clip_shot",
+                            trackId: "track_video",
+                            name: "shot",
+                            startMs: 0,
+                            durationMs: 6000,
+                            animations: []
+                        },
+                        // Planted: this clip names a track the document does not have.
+                        {
+                            id: "clip_title",
+                            trackId: "track_text",
+                            name: "Title",
+                            startMs: 0,
+                            durationMs: 2000,
+                            animations: []
+                        }
+                    ]
+                }
+            ]
+        ]),
+        sketches: new Map([
+            [
+                "img_cover",
+                {
+                    id: "img_cover",
+                    name: "Cover",
+                    layers: [
+                        {
+                            id: "layer_bg",
+                            name: "Background",
+                            opacity: 1,
+                            blendMode: "normal",
+                            visible: true
+                        },
+                        {
+                            id: "layer_art",
+                            name: "Art",
+                            opacity: 0.9,
+                            blendMode: "normal",
+                            visible: true
+                        }
+                    ],
+                    // Planted: no layer carries this id.
+                    activeLayerId: "layer_ghost"
+                }
+            ]
+        ]),
+        versions: new Map(),
+        scripts: new Map([
+            [
+                "scr_intro",
+                {
+                    script_id: "scr_intro",
+                    name: "Intro voiceover",
+                    cast: [
+                        {
+                            name: "Narrator",
+                            provider: "openai",
+                            model: "tts-1",
+                            voice: "alloy"
+                        }
+                    ],
+                    lines: [
+                        {
+                            id: "line_1",
+                            speaker: "Narrator",
+                            text: "Welcome to NodeTool.",
+                            status: "voiced",
+                            take: "asset://take_line_1",
+                            duration_seconds: 2
+                        },
+                        {
+                            id: "line_2",
+                            speaker: "Narrator",
+                            text: "Everything here is a graph.",
+                            status: "draft",
+                            take: null,
+                            duration_seconds: 0
+                        },
+                        {
+                            id: "line_3",
+                            speaker: "Narrator",
+                            text: "Let us begin.",
+                            status: "stale",
+                            take: "asset://take_line_3_old",
+                            duration_seconds: 2
+                        }
+                    ]
+                }
+            ]
+        ]),
+        storyboards: new Map([
+            [
+                "sb_lighthouse",
+                {
+                    storyboard_id: "sb_lighthouse",
+                    name: "Lighthouse",
+                    image_model: { provider: "fal_ai", model: "fal-ai/flux/schnell" },
+                    video_model: { provider: "fal_ai", model: "fal-ai/ltx" },
+                    shots: [
+                        {
+                            id: "shot_1",
+                            slug: "wide-lighthouse",
+                            action: "Wide of the lighthouse at dusk",
+                            duration_seconds: 4,
+                            duration_source: "manual",
+                            still: "asset://still_shot_1",
+                            clip: null,
+                            revision: 1
+                        },
+                        {
+                            id: "shot_2",
+                            slug: "keeper-door",
+                            action: "The keeper closes the door",
+                            duration_seconds: 3,
+                            duration_source: "manual",
+                            still: null,
+                            clip: null,
+                            revision: 1
+                        },
+                        {
+                            id: "shot_3",
+                            slug: "beam-sweep",
+                            action: "The beam sweeps across the water",
+                            duration_seconds: 5,
+                            duration_source: "manual",
+                            still: null,
+                            clip: null,
+                            revision: 1
+                        }
+                    ]
+                }
+            ]
+        ]),
+        // NodeTool's own configuration. A setting has a declared shape; a secret
+        // has only a name and whether it exists — which is the whole point of the
+        // split the `settings` capabilities enforce.
+        settingCatalog: [
+            {
+                key: "AUTOSAVE_ENABLED",
+                group: "Autosave",
+                description: "Save workflow versions automatically.",
+                allowed_values: ["true", "false"]
+            },
+            {
+                key: "AUTOSAVE_INTERVAL_MINUTES",
+                group: "Autosave",
+                description: "Minutes between autosaves.",
+                allowed_values: null
+            },
+            {
+                key: "FONT_PATH",
+                group: "Folders",
+                description: "Where image nodes look for fonts.",
+                allowed_values: null
+            }
+        ],
+        settingValues: new Map([["AUTOSAVE_ENABLED", "true"]]),
+        secretCatalog: [
+            { key: "OPENAI_API_KEY", description: "OpenAI API key." },
+            { key: "STRIPE_API_KEY", description: "Stripe API key." }
+        ],
+        configuredSecrets: new Set(["OPENAI_API_KEY"]),
+        /** What `request_secret` was asked for, so a case can score the ask. */
+        secretRequests: []
+    };
+}
+// ---------------------------------------------------------------------------
+// document helpers
+// ---------------------------------------------------------------------------
+const ANIMATION_PRESETS = ["fade", "slide", "zoom", "wipe"];
+const BLEND_MODES = ["normal", "multiply", "screen", "overlay", "add"];
+const slug = (text) => text
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, "_")
+    .replace(/^_|_$/g, "") || "item";
+function needTimeline(world, id) {
+    const doc = world.timelines.get(id);
+    if (!doc)
+        throw new Error(`no timeline sequence with id "${id}"`);
+    return doc;
+}
+function needSketch(world, id) {
+    const doc = world.sketches.get(id);
+    if (!doc)
+        throw new Error(`no image document with id "${id}"`);
+    return doc;
+}
+function timelineIssues(doc) {
+    const issues = [];
+    const trackIds = new Set(doc.tracks.map((t) => t.id));
+    for (const clip of doc.clips) {
+        if (!trackIds.has(clip.trackId)) {
+            issues.push({
+                severity: "error",
+                code: "missing_track",
+                message: `clip "${clip.name}" sits on track "${clip.trackId}", which the document does not have`,
+                clip: clip.id,
+                track: clip.trackId
+            });
+        }
+        if (clip.durationMs <= 0) {
+            issues.push({
+                severity: "error",
+                code: "zero_duration",
+                message: `clip "${clip.name}" has no duration`,
+                clip: clip.id
+            });
+        }
+        for (const animation of clip.animations) {
+            if (!ANIMATION_PRESETS.includes(animation.preset)) {
+                issues.push({
+                    severity: "error",
+                    code: "unknown_preset",
+                    message: `clip "${clip.name}" uses animation preset "${animation.preset}", which does not exist`,
+                    clip: clip.id
+                });
+            }
+        }
+    }
+    return issues;
+}
+function sketchIssues(doc) {
+    const issues = [];
+    const layerIds = new Set(doc.layers.map((l) => l.id));
+    if (!layerIds.has(doc.activeLayerId)) {
+        issues.push({
+            severity: "error",
+            code: "dangling_active_layer",
+            message: `activeLayerId "${doc.activeLayerId}" names no layer in this document`,
+            layer: doc.activeLayerId
+        });
+    }
+    if (layerIds.size !== doc.layers.length) {
+        issues.push({
+            severity: "error",
+            code: "duplicate_layer_id",
+            message: "two layers share one id"
+        });
+    }
+    for (const layer of doc.layers) {
+        if (layer.opacity < 0 || layer.opacity > 1) {
+            issues.push({
+                severity: "error",
+                code: "opacity_out_of_range",
+                message: `layer "${layer.name}" has opacity ${layer.opacity}`,
+                layer: layer.id
+            });
+        }
+        if (!BLEND_MODES.includes(layer.blendMode)) {
+            issues.push({
+                severity: "error",
+                code: "unknown_blend_mode",
+                message: `layer "${layer.name}" uses blend mode "${layer.blendMode}", which no compositor ships`,
+                layer: layer.id
+            });
+        }
+    }
+    return issues;
+}
+function snapshotVersion(world, key, name, saveType, document) {
+    const list = world.versions.get(key) ?? [];
+    const entry = {
+        version: list.length + 1,
+        name,
+        save_type: saveType,
+        document: clone(document)
+    };
+    list.push(entry);
+    world.versions.set(key, list);
+    return entry;
+}
+function findClip(doc, target) {
+    const found = doc.clips.find((c) => c.id === target || c.name === target);
+    if (!found) {
+        throw new Error(`no clip matching "${target}" (have ${doc.clips.map((c) => c.id).join(", ")})`);
+    }
+    return found;
+}
+function findLayer(doc, target) {
+    const found = doc.layers.find((l) => l.id === target || l.name === target);
+    if (!found) {
+        throw new Error(`no layer matching "${target}" (have ${doc.layers.map((l) => l.id).join(", ")})`);
+    }
+    return found;
+}
+/** Apply one timeline edit op; throws with the op named when it cannot. */
+function applyTimelineOp(doc, op) {
+    const kind = str(op["op"]);
+    switch (kind) {
+        case "get_state":
+            return;
+        case "add_track": {
+            const type = str(op["type"], "video");
+            let id = str(op["id"]) || `track_${type}`;
+            let n = 2;
+            while (doc.tracks.some((t) => t.id === id))
+                id = `track_${type}_${n++}`;
+            doc.tracks.push({ id, type, name: str(op["name"], type) });
+            return;
+        }
+        case "add_text_clip": {
+            const track = doc.tracks.find((t) => t.type === "text");
+            if (!track)
+                throw new Error("the document has no text track to add to");
+            const text = str(op["text"], "Text");
+            doc.clips.push({
+                id: `clip_${slug(text)}`,
+                trackId: track.id,
+                name: text,
+                startMs: num(op["startMs"], 0),
+                durationMs: num(op["durationMs"], 2000),
+                animations: []
+            });
+            return;
+        }
+        case "split_clip": {
+            const clip = findClip(doc, str(op["target"]));
+            const atMs = num(op["atMs"], Math.round(clip.durationMs / 2));
+            const offset = atMs - clip.startMs;
+            if (offset <= 0 || offset >= clip.durationMs) {
+                throw new Error(`atMs ${atMs} is outside clip "${clip.id}"`);
+            }
+            const tail = {
+                id: `${clip.id}_b`,
+                trackId: clip.trackId,
+                name: `${clip.name} (b)`,
+                startMs: atMs,
+                durationMs: clip.durationMs - offset,
+                animations: []
+            };
+            clip.durationMs = offset;
+            doc.clips.push(tail);
+            return;
+        }
+        case "animate_clip": {
+            const clip = findClip(doc, str(op["target"]));
+            clip.animations = recList(op["animations"]).map((a) => ({
+                role: str(a["role"], "in"),
+                preset: str(a["preset"], "fade")
+            }));
+            return;
+        }
+        case "move_clip": {
+            const clip = findClip(doc, str(op["target"]));
+            clip.startMs = num(op["startMs"], clip.startMs);
+            return;
+        }
+        case "remove_clip": {
+            const clip = findClip(doc, str(op["target"]));
+            doc.clips = doc.clips.filter((c) => c.id !== clip.id);
+            return;
+        }
+        default:
+            throw new Error(`unknown op "${kind}" — try get_state, add_track, add_text_clip, ` +
+                "split_clip, animate_clip, move_clip, remove_clip");
+    }
+}
+function applySketchOp(doc, op) {
+    const kind = str(op["op"]);
+    switch (kind) {
+        case "get_state":
+            return;
+        case "add_layer": {
+            const name = str(op["name"], "Layer");
+            let id = str(op["id"]) || `layer_${slug(name)}`;
+            let n = 2;
+            while (doc.layers.some((l) => l.id === id))
+                id = `layer_${slug(name)}_${n++}`;
+            doc.layers.push({
+                id,
+                name,
+                opacity: num(op["opacity"], 1),
+                blendMode: str(op["blendMode"], "normal"),
+                visible: true
+            });
+            return;
+        }
+        case "set_layer_props": {
+            const layer = findLayer(doc, str(op["target"]));
+            if (op["opacity"] !== undefined)
+                layer.opacity = num(op["opacity"], layer.opacity);
+            if (op["blendMode"] !== undefined)
+                layer.blendMode = str(op["blendMode"], layer.blendMode);
+            if (op["visible"] !== undefined)
+                layer.visible = op["visible"] !== false;
+            if (op["name"] !== undefined)
+                layer.name = str(op["name"], layer.name);
+            return;
+        }
+        case "set_active_layer": {
+            const layer = findLayer(doc, str(op["target"]));
+            doc.activeLayerId = layer.id;
+            return;
+        }
+        case "remove_layer": {
+            const layer = findLayer(doc, str(op["target"]));
+            doc.layers = doc.layers.filter((l) => l.id !== layer.id);
+            return;
+        }
+        default:
+            throw new Error(`unknown op "${kind}" — try get_state, add_layer, set_layer_props, ` +
+                "set_active_layer, remove_layer");
+    }
+}
+function applyScriptOp(doc, op) {
+    const kind = str(op["op"]);
+    switch (kind) {
+        case "get_state":
+            return;
+        case "add_speaker": {
+            const name = str(op["name"], "Speaker");
+            doc.cast.push({
+                name,
+                provider: str(op["provider"], "openai"),
+                model: str(op["model"], "tts-1"),
+                voice: str(op["voice"], "alloy")
+            });
+            return;
+        }
+        case "add_line": {
+            doc.lines.push({
+                id: `line_${doc.lines.length + 1}`,
+                speaker: str(op["speaker"], doc.cast[0]?.name ?? "Narrator"),
+                text: str(op["text"]),
+                status: "draft",
+                take: null,
+                duration_seconds: 0
+            });
+            return;
+        }
+        case "set_line_text": {
+            const target = str(op["target"]);
+            const line = doc.lines.find((l) => l.id === target);
+            if (!line)
+                throw new Error(`no line with id "${target}"`);
+            line.text = str(op["text"], line.text);
+            if (line.status === "voiced")
+                line.status = "stale";
+            return;
+        }
+        default:
+            throw new Error(`unknown op "${kind}" — try get_state, add_speaker, add_line, set_line_text`);
+    }
+}
+/** The two sources a shot's duration may come from; anything else is an error. */
+function durationSource(value, fallback) {
+    if (value === undefined && fallback !== undefined)
+        return fallback;
+    if (value === "audio" || value === "manual")
+        return value;
+    throw new Error('duration_source must be "audio" or "manual"');
+}
+function applyStoryboardOp(doc, op) {
+    const kind = str(op["op"]);
+    switch (kind) {
+        case "get_state":
+            return;
+        case "add_shot": {
+            const action = str(op["action"], "Shot");
+            doc.shots.push({
+                id: `shot_${doc.shots.length + 1}`,
+                slug: slug(action).slice(0, 24),
+                action,
+                duration_seconds: num(op["duration_seconds"], 3),
+                duration_source: durationSource(op["duration_source"], "manual"),
+                still: null,
+                clip: null,
+                revision: 1
+            });
+            return;
+        }
+        case "reorder_shot": {
+            const target = str(op["target"]);
+            const from = doc.shots.findIndex((s) => s.id === target || s.slug === target);
+            if (from < 0)
+                throw new Error(`no shot matching "${target}"`);
+            const [shot] = doc.shots.splice(from, 1);
+            doc.shots.splice(Math.max(0, num(op["index"], 0)), 0, shot);
+            return;
+        }
+        case "update_shot": {
+            const target = str(op["target"]);
+            const shot = doc.shots.find((s) => s.id === target || s.slug === target);
+            if (!shot)
+                throw new Error(`no shot matching "${target}"`);
+            if (op["action"] !== undefined)
+                shot.action = str(op["action"], shot.action);
+            if (op["duration_seconds"] !== undefined) {
+                shot.duration_seconds = num(op["duration_seconds"], shot.duration_seconds);
+            }
+            if (op["duration_source"] !== undefined) {
+                shot.duration_source = durationSource(op["duration_source"]);
+            }
+            return;
+        }
+        case "remove_shot": {
+            const target = str(op["target"]);
+            const at = doc.shots.findIndex((s) => s.id === target || s.slug === target);
+            if (at < 0)
+                throw new Error(`no shot matching "${target}"`);
+            doc.shots.splice(at, 1);
+            return;
+        }
+        default:
+            throw new Error(`unknown op "${kind}" — try get_state, add_shot, update_shot, ` +
+                "remove_shot, reorder_shot");
+    }
+}
+/** Resolve `targets` (ids, slugs, or indexes) against a shot list. */
+function resolveShots(doc, targets) {
+    const wanted = Array.isArray(targets) ? targets : [];
+    if (wanted.length === 0)
+        return [];
+    return wanted.map((entry) => {
+        if (isNumber(entry)) {
+            const shot = doc.shots[entry];
+            if (!shot)
+                throw new Error(`no shot at index ${entry}`);
+            return shot;
+        }
+        const key = String(entry);
+        const shot = doc.shots.find((s) => s.id === key || s.slug === key);
+        if (!shot)
+            throw new Error(`no shot matching "${key}"`);
+        return shot;
+    });
+}
+// ---------------------------------------------------------------------------
+// the belt
+// ---------------------------------------------------------------------------
+/**
+ * Instrumented fakes for the data + creative belt tools, over a fresh world.
+ * `web_search` is the one search entry point — the real tool routes across
+ * its configured backends host-side, so the belt carries no per-provider
+ * search duplicates.
+ */
+export function createSurfaceApiTools(recorder) {
+    const world = createWorld();
+    const games = new Map();
+    const tool = (name, description, schema, impl) => new RecordingTool(name, description, schema, recorder, impl);
+    return [
+        // -- collections ------------------------------------------------------
+        tool("list_collections", "List vector collections.", obj({}), () => ({
+            collections: [
+                { name: "handbook", count: HANDBOOK.length },
+                { name: "scratch", count: 0 }
+            ]
+        })),
+        tool("query_collection", "Semantic search over a named collection.", obj({ collection: S, query: S, n_results: N }, ["collection"]), (params) => {
+            const name = str(params["collection"]);
+            if (name !== "handbook") {
+                throw new Error(`no collection named "${name}" — try handbook`);
+            }
+            const query = str(params["query"]);
+            const limit = num(params["n_results"], 5);
+            const ranked = HANDBOOK.map((doc) => ({
+                id: doc.id,
+                text: doc.text,
+                score: overlap(query, doc.text)
+            })).sort((a, b) => b.score - a.score || a.id.localeCompare(b.id));
+            return { collection: name, results: ranked.slice(0, limit) };
+        }),
+        tool("vector_index", "Index one chunk of text under a source id.", obj({ text: S, source_id: S, metadata: obj({}) }, ["text", "source_id"]), (params) => {
+            const sourceId = str(params["source_id"]);
+            const text = str(params["text"]);
+            if (!sourceId || !text) {
+                throw new Error("vector_index needs both text and source_id");
+            }
+            world.indexed = world.indexed.filter((c) => c.source_id !== sourceId);
+            world.indexed.push({
+                source_id: sourceId,
+                text,
+                metadata: rec(params["metadata"])
+            });
+            return { indexed: 1, source_id: sourceId, total: world.indexed.length };
+        }),
+        tool("vector_batch_index", "Index many chunks at once.", obj({ chunks: ANY_ARRAY, base_metadata: obj({}) }, ["chunks"]), (params) => {
+            const chunks = recList(params["chunks"]);
+            for (const chunk of chunks) {
+                const sourceId = str(chunk["source_id"]);
+                if (!sourceId)
+                    throw new Error("every chunk needs a source_id");
+                world.indexed = world.indexed.filter((c) => c.source_id !== sourceId);
+                world.indexed.push({
+                    source_id: sourceId,
+                    text: str(chunk["text"]),
+                    metadata: {
+                        ...rec(params["base_metadata"]),
+                        ...rec(chunk["metadata"])
+                    }
+                });
+            }
+            return { indexed: chunks.length, total: world.indexed.length };
+        }),
+        tool("vector_text_search", "Semantic search over the indexed chunks.", obj({ text: S, n_results: N }, ["text"]), (params) => {
+            const query = str(params["text"]);
+            const limit = num(params["n_results"], 5);
+            const ranked = world.indexed
+                .map((chunk) => ({
+                source_id: chunk.source_id,
+                text: chunk.text,
+                metadata: chunk.metadata,
+                score: overlap(query, chunk.text)
+            }))
+                .filter((hit) => hit.score > 0)
+                .sort((a, b) => b.score - a.score || a.source_id.localeCompare(b.source_id));
+            return { query, results: ranked.slice(0, limit) };
+        }),
+        tool("vector_hybrid_search", "Semantic + keyword search fused by reciprocal rank.", obj({ text: S, n_results: N, k_constant: N }, ["text"]), (params) => {
+            const query = str(params["text"]);
+            const limit = num(params["n_results"], 5);
+            const ranked = world.indexed
+                .map((chunk) => ({
+                source_id: chunk.source_id,
+                text: chunk.text,
+                // Keyword half nudges exact-substring hits up.
+                score: overlap(query, chunk.text) +
+                    (chunk.text.toLowerCase().includes(query.toLowerCase()) ? 1 : 0)
+            }))
+                .filter((hit) => hit.score > 0)
+                .sort((a, b) => b.score - a.score || a.source_id.localeCompare(b.source_id));
+            return { query, results: ranked.slice(0, limit) };
+        }),
+        // -- memory -----------------------------------------------------------
+        tool("memory_save", "Remember something durably in this conversation.", obj({ content: S, title: S, kind: S, resources: ANY_ARRAY }, ["content"]), (params) => {
+            const content = str(params["content"]);
+            if (!content)
+                throw new Error("memory_save needs content");
+            const entry = {
+                memory_id: `mem_${++world.memorySeq}`,
+                title: str(params["title"], content.slice(0, 40)),
+                content,
+                kind: str(params["kind"], "note"),
+                resources: Array.isArray(params["resources"])
+                    ? params["resources"]
+                    : []
+            };
+            world.memories.push(entry);
+            return { saved: true, memory_id: entry.memory_id };
+        }),
+        tool("memory_list", "List this conversation's memories.", obj({ limit: N }), (params) => ({
+            memories: world.memories.slice(0, num(params["limit"], 50)).map(clone)
+        })),
+        tool("memory_update", "Update a memory by id.", obj({ memory_id: S, content: S, title: S, resources: ANY_ARRAY }, [
+            "memory_id"
+        ]), (params) => {
+            const id = str(params["memory_id"]);
+            const entry = world.memories.find((m) => m.memory_id === id);
+            if (!entry)
+                throw new Error(`no memory with id "${id}"`);
+            if (params["content"] !== undefined)
+                entry.content = str(params["content"], entry.content);
+            if (params["title"] !== undefined)
+                entry.title = str(params["title"], entry.title);
+            if (Array.isArray(params["resources"]))
+                entry.resources = params["resources"];
+            return { updated: true, memory_id: id };
+        }),
+        tool("memory_delete", "Delete a memory by id.", obj({ memory_id: S }, ["memory_id"]), (params) => {
+            const id = str(params["memory_id"]);
+            const before = world.memories.length;
+            world.memories = world.memories.filter((m) => m.memory_id !== id);
+            if (world.memories.length === before) {
+                throw new Error(`no memory with id "${id}"`);
+            }
+            return { deleted: true, memory_id: id };
+        }),
+        // -- threads ----------------------------------------------------------
+        tool("list_threads", "List past chat threads, most recently updated first.", obj({ limit: N, workflow_id: S, cursor: S, preview: B }), (params) => {
+            const wanted = num(params["limit"], 20);
+            const rows = [...world.threads]
+                .sort((a, b) => (a.updated_at < b.updated_at ? 1 : -1))
+                .slice(0, wanted);
+            return {
+                threads: rows.map((thread) => {
+                    const last = [...world.chatMessages]
+                        .filter((m) => m.thread_id === thread.id)
+                        .sort((a, b) => (a.created_at < b.created_at ? 1 : -1))[0];
+                    return {
+                        ...clone(thread),
+                        last_message: params["preview"] === false || !last
+                            ? undefined
+                            : { id: last.id, role: last.role, text: last.text }
+                    };
+                })
+            };
+        }),
+        tool("get_thread", "Read one thread and a page of its messages.", obj({
+            thread_id: S,
+            limit: N,
+            newest_first: B,
+            cursor: S,
+            max_chars: N
+        }, ["thread_id"]), (params) => {
+            const id = str(params["thread_id"]);
+            const thread = world.threads.find((t) => t.id === id);
+            if (!thread)
+                throw new Error(`no thread with id "${id}"`);
+            const ordered = world.chatMessages
+                .filter((m) => m.thread_id === id)
+                .sort((a, b) => (a.created_at < b.created_at ? -1 : 1));
+            const rows = params["newest_first"] === true ? ordered.reverse() : ordered;
+            return {
+                ...clone(thread),
+                // Tool calls come back named but without their arguments — the
+                // summary the real capability returns.
+                messages: rows.slice(0, num(params["limit"], 50)).map((m) => ({
+                    id: m.id,
+                    role: m.role,
+                    text: m.text,
+                    created_at: m.created_at,
+                    tool_calls: m.tool_calls?.map((c) => ({ id: c.id, name: c.name }))
+                }))
+            };
+        }),
+        tool("get_message", "Read one chat message in full, arguments and all.", obj({ message_id: S }, ["message_id"]), (params) => {
+            const id = str(params["message_id"]);
+            const message = world.chatMessages.find((m) => m.id === id);
+            if (!message)
+                throw new Error(`no message with id "${id}"`);
+            return clone(message);
+        }),
+        // -- shared -----------------------------------------------------------
+        tool("list_shared", "List entries in shared agent memory (metadata only).", obj({ kind: ANY_ARRAY, key_prefix: S, sources: ANY_ARRAY }), (params) => {
+            const kinds = strList(params["kind"]);
+            const prefix = str(params["key_prefix"]);
+            const entries = world.shared
+                .filter((e) => kinds.length === 0 || kinds.includes(e.kind))
+                .filter((e) => !prefix || e.key.startsWith(prefix))
+                .map((e) => ({
+                key: e.key,
+                kind: e.kind,
+                title: e.title,
+                valueBytes: JSON.stringify(e.value).length
+            }));
+            return { total: entries.length, returned: entries.length, entries };
+        }),
+        tool("read_shared", "Read full values from shared agent memory by key.", obj({ keys: ANY_ARRAY }, ["keys"]), (params) => {
+            const entries = {};
+            const missing = [];
+            for (const key of strList(params["keys"])) {
+                // A bare suffix falls back to the `shared:` namespace, the way the
+                // real capability does.
+                const hit = world.shared.find((e) => e.key === key) ??
+                    (key.includes(":")
+                        ? undefined
+                        : world.shared.find((e) => e.key === `shared:${key}`));
+                if (hit)
+                    entries[key] = clone(hit);
+                else
+                    missing.push(key);
+            }
+            return { entries, missing };
+        }),
+        tool("share_result", "Publish a value under the `shared:` namespace.", obj({ key: S, value: {}, title: S, description: S }, ["key", "value"]), (params) => {
+            const raw = str(params["key"]);
+            if (!raw)
+                throw new Error("share_result needs a key");
+            const suffix = raw.startsWith("shared:")
+                ? raw.slice("shared:".length)
+                : raw;
+            const key = `shared:${suffix}`;
+            const entry = {
+                key,
+                kind: "shared",
+                title: str(params["title"], suffix),
+                value: params["value"]
+            };
+            const existing = world.shared.findIndex((e) => e.key === key);
+            if (existing >= 0)
+                world.shared[existing] = entry;
+            else
+                world.shared.push(entry);
+            return { ok: true, key, kind: "shared" };
+        }),
+        // -- web --------------------------------------------------------------
+        tool("web_search", "Search the web (routes across configured backends host-side). " +
+            "search_type picks web, news, or images.", obj({ query: S, backend: S, search_type: S, max_results: N }, ["query"]), (params) => {
+            const query = str(params["query"]);
+            // One tool, three search types — the shape the real capability has
+            // since the google_news / google_images wrappers folded into it.
+            if (str(params["search_type"]) === "news") {
+                return {
+                    query,
+                    results: [
+                        {
+                            title: "NodeTool 4.2 lands the CodeAct sandbox",
+                            url: "https://news.example.dev/nodetool-4-2",
+                            published: "2026-07-02"
+                        }
+                    ]
+                };
+            }
+            const ranked = WEB_PAGES.map((page) => ({
+                title: page.title,
+                url: page.url,
+                snippet: page.snippet,
+                score: overlap(query, `${page.title} ${page.snippet} ${page.url}`)
+            }))
+                .sort((a, b) => b.score - a.score || a.url.localeCompare(b.url))
+                .slice(0, num(params["max_results"], 5));
+            return { query, results: ranked };
+        }),
+        tool("http_request", "One HTTP request; returns the body as text.", obj({ url: S, method: S, headers: obj({}), body: S }, ["url"]), (params) => {
+            const url = str(params["url"]);
+            const page = WEB_PAGES.find((p) => p.url === url);
+            return page
+                ? { status: 200, url, body: page.text }
+                : { status: 404, url, body: "" };
+        }),
+        tool("browser", "Fetch a page and return its readable text.", obj({ url: S }, ["url"]), (params) => {
+            const url = str(params["url"]);
+            const page = WEB_PAGES.find((p) => p.url === url);
+            if (!page)
+                throw new Error(`nothing served at ${url}`);
+            return { url, title: page.title, text: page.text };
+        }),
+        tool("download_file", "Save a URL's bytes into the workspace.", obj({ url: S, output_file: S }, ["url", "output_file"]), (params) => {
+            const url = str(params["url"]);
+            const page = WEB_PAGES.find((p) => p.url === url);
+            if (!page)
+                throw new Error(`nothing served at ${url}`);
+            return {
+                output_file: str(params["output_file"]),
+                bytes: page.text.length
+            };
+        }),
+        // -- documents --------------------------------------------------------
+        tool("extract_pdf_text", "Extract text from a PDF in the workspace.", obj({ path: S, start_page: N, end_page: N }, ["path"]), (params) => {
+            const path = str(params["path"]);
+            if (path !== "reports/q3.pdf") {
+                throw new Error(`no such file: ${path} (the workspace has reports/q3.pdf)`);
+            }
+            return { path, pages: 1, text: PDF_TEXT };
+        }),
+        tool("extract_pdf_tables", "Extract a PDF's tables to a JSON file.", obj({ path: S, output_file: S }, ["path", "output_file"]), (params) => {
+            const path = str(params["path"]);
+            if (path !== "reports/q3.pdf")
+                throw new Error(`no such file: ${path}`);
+            return {
+                path,
+                output_file: str(params["output_file"], "tables.json"),
+                tables: 1,
+                rows: 3
+            };
+        }),
+        tool("convert_pdf_to_markdown", "Convert a PDF to markdown.", obj({ input_file: S, output_file: S }, ["input_file", "output_file"]), (params) => {
+            const input = str(params["input_file"]);
+            if (input !== "reports/q3.pdf")
+                throw new Error(`no such file: ${input}`);
+            const output = str(params["output_file"], "out.md");
+            return { output_file: output, characters: PDF_TEXT.length + 8 };
+        }),
+        tool("convert_markdown_to_pdf", "Convert markdown to a PDF.", obj({ input_file: S, output_file: S }, ["input_file", "output_file"]), (params) => ({
+            output_file: str(params["output_file"], "out.pdf"),
+            pages: 1
+        })),
+        tool("convert_document", "Convert a document from one format to another.", obj({
+            input_file: S,
+            output_file: S,
+            from_format: S,
+            to_format: S,
+            extra_args: ANY_ARRAY
+        }, ["input_file", "output_file"]), (params) => ({
+            input_file: str(params["input_file"]),
+            output_file: str(params["output_file"]),
+            converted: true
+        })),
+        // -- settings ---------------------------------------------------------
+        tool("list_settings", "List NodeTool's configuration settings.", obj({ group: S }), (params) => {
+            const group = str(params["group"]).toLowerCase();
+            const settings = world.settingCatalog
+                .filter((def) => !group || def.group.toLowerCase() === group)
+                .map((def) => ({
+                ...def,
+                value: world.settingValues.get(def.key) ?? null
+            }));
+            return { settings };
+        }),
+        tool("get_setting", "Read one setting's value.", obj({ key: S }, ["key"]), (params) => {
+            const key = str(params["key"]);
+            const def = world.settingCatalog.find((d) => d.key === key);
+            if (!def) {
+                return {
+                    ok: false,
+                    error: `no setting named "${key}"`,
+                    error_kind: "unknown_setting"
+                };
+            }
+            return { ...def, value: world.settingValues.get(key) ?? null };
+        }),
+        tool("set_setting", "Change one setting.", obj({ key: S, value: S }, ["key", "value"]), (params) => {
+            const key = str(params["key"]);
+            const value = str(params["value"]);
+            if (world.secretCatalog.some((sec) => sec.key === key)) {
+                return {
+                    ok: false,
+                    error: `${key} holds a credential — call request_secret instead`,
+                    error_kind: "is_secret"
+                };
+            }
+            const def = world.settingCatalog.find((d) => d.key === key);
+            if (!def) {
+                return {
+                    ok: false,
+                    error: `no setting named "${key}"`,
+                    error_kind: "unknown_setting"
+                };
+            }
+            if (def.allowed_values && !def.allowed_values.includes(value)) {
+                return {
+                    ok: false,
+                    error: `${key} accepts only: ${def.allowed_values.join(", ")}`,
+                    error_kind: "invalid_value"
+                };
+            }
+            world.settingValues.set(key, value);
+            return { ok: true, key, value };
+        }),
+        tool("list_secrets", "Which credentials this install holds. Never their values.", obj({}), () => ({
+            secrets: world.secretCatalog.map((sec) => ({
+                ...sec,
+                configured: world.configuredSecrets.has(sec.key)
+            }))
+        })),
+        tool("request_secret", "Ask the user to enter a credential in a dialog. Takes no value.", obj({ key: S, reason: S, help_url: S }, ["key"]), (params) => {
+            const key = str(params["key"]);
+            // The fake user always saves. What is worth scoring is that the model
+            // asked rather than trying to write the value itself — there is no
+            // parameter here it could have written it into.
+            world.secretRequests.push({ key, reason: str(params["reason"]) });
+            world.configuredSecrets.add(key);
+            return { ok: true, key, configured: true };
+        }),
+        // -- email ------------------------------------------------------------
+        tool("search_email", "Search the mailbox.", obj({
+            subject: S,
+            text: S,
+            from: S,
+            since_hours_ago: N,
+            max_results: N,
+            include_archived: B
+        }), (params) => {
+            const subject = str(params["subject"]).toLowerCase();
+            const text = str(params["text"]).toLowerCase();
+            const from = str(params["from"]).toLowerCase();
+            const since = num(params["since_hours_ago"], Number.POSITIVE_INFINITY);
+            const includeArchived = params["include_archived"] === true;
+            const matched = world.emails.filter((message) => {
+                if (!includeArchived && message.archived)
+                    return false;
+                if (message.hours_ago > since)
+                    return false;
+                const haystack = `${message.subject} ${message.from}`.toLowerCase();
+                if (subject && !message.subject.toLowerCase().includes(subject))
+                    return false;
+                if (from && !message.from.toLowerCase().includes(from))
+                    return false;
+                if (text && !haystack.includes(text))
+                    return false;
+                return true;
+            });
+            return {
+                messages: matched
+                    .slice(0, num(params["max_results"], 25))
+                    .map((message) => ({
+                    message_id: message.message_id,
+                    subject: message.subject,
+                    from: message.from,
+                    hours_ago: message.hours_ago,
+                    labels: [...message.labels]
+                }))
+            };
+        }),
+        tool("add_label_to_email", "Add one label to one message.", obj({ message_id: S, label: S }, ["message_id", "label"]), (params) => {
+            const id = str(params["message_id"]);
+            const label = str(params["label"]);
+            const message = world.emails.find((m) => m.message_id === id);
+            if (!message)
+                throw new Error(`no message with id "${id}"`);
+            if (!label)
+                throw new Error("add_label_to_email needs a label");
+            if (!message.labels.includes(label))
+                message.labels.push(label);
+            return { message_id: id, labels: [...message.labels] };
+        }),
+        tool("archive_email", "Archive messages by id.", obj({ message_ids: ANY_ARRAY }, ["message_ids"]), (params) => {
+            const ids = strList(params["message_ids"]);
+            let archived = 0;
+            for (const id of ids) {
+                const message = world.emails.find((m) => m.message_id === id);
+                if (!message)
+                    throw new Error(`no message with id "${id}"`);
+                if (!message.archived) {
+                    message.archived = true;
+                    archived++;
+                }
+            }
+            return { archived, message_ids: ids };
+        }),
+        // -- apps -------------------------------------------------------------
+        tool("debug_app", "Validate and simulate a saved app.", obj({ application_id: S, params: obj({}), interact: ANY_ARRAY, run: B }, [
+            "application_id"
+        ]), (params) => {
+            const id = str(params["application_id"]);
+            if (!world.apps.has(id))
+                throw new Error(`no application with id "${id}"`);
+            return {
+                application_id: id,
+                verdict: { ok: true, issues: [] },
+                ran: params["run"] !== false,
+                widgets: [
+                    { id: "Input-1", type: "TextInput" },
+                    { id: "Button-1", type: "Button" },
+                    { id: "Text-1", type: "Text" }
+                ]
+            };
+        }),
+        // -- games ------------------------------------------------------------
+        tool("create_native_game", "Create a native game draft.", obj({ name: S }, ["name"]), (params) => {
+            const id = `game_${games.size + 1}`;
+            const saved = { name: str(params["name"]), document: createTopDownRoomGame(id), draft_updated_at: "draft_0" };
+            games.set(id, saved);
+            return { game: { id, name: saved.name, revision: saved.document.revision }, document: clone(saved.document), draft_updated_at: saved.draft_updated_at };
+        }),
+        tool("edit_native_game", "Apply validated game document operations atomically.", obj({ game_id: S, ops: ANY_ARRAY, base_updated_at: S }, ["game_id", "ops"]), (params) => {
+            const id = str(params["game_id"]);
+            const saved = games.get(id);
+            if (!saved) {
+                throw new Error(`no game with id "${id}"`);
+            }
+            if (params["base_updated_at"] !== undefined && params["base_updated_at"] !== saved.draft_updated_at) {
+                return { error: "conflict", message: "Game draft was modified concurrently" };
+            }
+            const ops = recList(params["ops"]).map((op) => gameDocumentOp.parse(op));
+            saved.document = applyGameOps(saved.document, ops);
+            saved.draft_updated_at = `draft_${Number(saved.draft_updated_at.slice(6)) + 1}`;
+            return { game: { id, name: saved.name, revision: saved.document.revision }, document: clone(saved.document), draft_updated_at: saved.draft_updated_at };
+        }),
+        tool("get_native_game", "Read a native game draft.", obj({ game_id: S, view: S }, ["game_id"]), (params) => {
+            const id = str(params["game_id"]);
+            const saved = games.get(id);
+            if (!saved) {
+                throw new Error(`no game with id "${id}"`);
+            }
+            return { game: { id, name: saved.name, revision: saved.document.revision }, document: clone(saved.document), draft_updated_at: saved.draft_updated_at };
+        }),
+        // -- timelines --------------------------------------------------------
+        tool("list_timelines", "List saved timeline sequences.", obj({ limit: N }), () => ({
+            timelines: [...world.timelines.values()].map((doc) => ({
+                timeline_id: doc.id,
+                name: doc.name,
+                tracks: doc.tracks.length,
+                clips: doc.clips.length
+            }))
+        })),
+        tool("validate_timeline", "Statically check a timeline sequence.", obj({ timeline_id: S, document: obj({}) }), (params) => {
+            const id = str(params["timeline_id"]);
+            const inline = rec(params["document"]);
+            const doc = id
+                ? needTimeline(world, id)
+                : {
+                    id: str(inline["id"], "inline"),
+                    name: str(inline["name"], "inline"),
+                    tracks: recList(inline["tracks"]).map((track) => ({
+                        id: str(track["id"]),
+                        type: str(track["type"]),
+                        name: str(track["name"])
+                    })),
+                    clips: recList(inline["clips"]).map((clip) => ({
+                        id: str(clip["id"]),
+                        trackId: str(clip["trackId"]),
+                        name: str(clip["name"]),
+                        startMs: num(clip["startMs"], 0),
+                        durationMs: num(clip["durationMs"], 0),
+                        animations: recList(clip["animations"]).map((animation) => ({
+                            role: str(animation["role"]),
+                            preset: str(animation["preset"])
+                        }))
+                    }))
+                };
+            const issues = timelineIssues(doc);
+            return { ok: issues.length === 0, issues, warnings: [] };
+        }),
+        tool("edit_timeline", "Apply document edits to a saved sequence.", obj({ timeline_id: S, ops: ANY_ARRAY }, ["timeline_id", "ops"]), (params) => {
+            const doc = needTimeline(world, str(params["timeline_id"]));
+            const ops = recList(params["ops"]);
+            if (ops.length === 0)
+                throw new Error("edit_timeline needs at least one op");
+            ops.forEach((op, index) => {
+                try {
+                    applyTimelineOp(doc, op);
+                }
+                catch (e) {
+                    throw new Error(`edit_timeline: op ${index + 1} (${str(op["op"], "?")}): ` +
+                        (e instanceof Error ? e.message : String(e)));
+                }
+            });
+            return { ok: true, applied: ops.length, document: clone(doc) };
+        }),
+        tool("list_timeline_versions", "List a sequence's snapshot history.", obj({ timeline_id: S, save_type: S, limit: N }, ["timeline_id"]), (params) => {
+            const id = str(params["timeline_id"]);
+            needTimeline(world, id);
+            return {
+                timeline_id: id,
+                versions: (world.versions.get(`timeline:${id}`) ?? []).map((v) => ({
+                    version: v.version,
+                    name: v.name,
+                    save_type: v.save_type
+                }))
+            };
+        }),
+        tool("get_timeline_version", "Read one snapshot without restoring it.", obj({ timeline_id: S, version: N }, ["timeline_id", "version"]), (params) => {
+            const id = str(params["timeline_id"]);
+            const version = num(params["version"], 0);
+            const found = (world.versions.get(`timeline:${id}`) ?? []).find((v) => v.version === version);
+            if (!found)
+                throw new Error(`no version ${version} for timeline "${id}"`);
+            return { timeline_id: id, version, document: clone(found.document) };
+        }),
+        tool("create_timeline_version", "Snapshot a sequence's current document.", obj({ timeline_id: S, name: S }, ["timeline_id"]), (params) => {
+            const id = str(params["timeline_id"]);
+            const doc = needTimeline(world, id);
+            const entry = snapshotVersion(world, `timeline:${id}`, str(params["name"], "manual save"), "manual", doc);
+            return { timeline_id: id, version: entry.version, name: entry.name };
+        }),
+        tool("restore_timeline_version", "Restore a snapshot, snapshotting the current state first.", obj({ timeline_id: S, version: N }, ["timeline_id", "version"]), (params) => {
+            const id = str(params["timeline_id"]);
+            const doc = needTimeline(world, id);
+            const version = num(params["version"], 0);
+            const found = (world.versions.get(`timeline:${id}`) ?? []).find((v) => v.version === version);
+            if (!found)
+                throw new Error(`no version ${version} for timeline "${id}"`);
+            snapshotVersion(world, `timeline:${id}`, "before restore", "restore", doc);
+            const restored = clone(found.document);
+            world.timelines.set(id, restored);
+            const issues = timelineIssues(restored);
+            return {
+                timeline_id: id,
+                restored: version,
+                validation: { ok: issues.length === 0, issues }
+            };
+        }),
+        tool("delete_timeline_version", "Delete one snapshot. The live sequence is not changed.", obj({ timeline_id: S, version: N }, ["timeline_id", "version"]), (params) => {
+            const id = str(params["timeline_id"]);
+            needTimeline(world, id);
+            const version = num(params["version"], 0);
+            const list = world.versions.get(`timeline:${id}`) ?? [];
+            const idx = list.findIndex((v) => v.version === version);
+            if (idx < 0)
+                throw new Error(`no version ${version} for timeline "${id}"`);
+            list.splice(idx, 1);
+            world.versions.set(`timeline:${id}`, list);
+            return { ok: true, timeline_id: id, deleted_version: version };
+        }),
+        // -- sketches ---------------------------------------------------------
+        tool("list_sketches", "List saved image documents.", obj({ limit: N }), () => ({
+            sketches: [...world.sketches.values()].map((doc) => ({
+                image_document_id: doc.id,
+                name: doc.name,
+                layers: doc.layers.length
+            }))
+        })),
+        tool("create_sketch", "Create a blank sketch (image document).", obj({ name: S, width: N, height: N, id: S }, ["name"]), (params) => {
+            const id = str(params["id"], `sketch_${world.sketches.size + 1}`);
+            if (world.sketches.has(id)) {
+                const existing = world.sketches.get(id);
+                return {
+                    ok: true,
+                    image_document_id: existing.id,
+                    name: existing.name
+                };
+            }
+            const layerId = "layer_1";
+            const doc = {
+                id,
+                name: str(params["name"], "Untitled"),
+                layers: [
+                    {
+                        id: layerId,
+                        name: "Layer 1",
+                        opacity: 1,
+                        blendMode: "normal",
+                        visible: true
+                    }
+                ],
+                activeLayerId: layerId
+            };
+            world.sketches.set(id, doc);
+            return { ok: true, image_document_id: id, name: doc.name };
+        }),
+        tool("validate_sketch", "Statically check an image document.", obj({ image_document_id: S, document: obj({}) }), (params) => {
+            const id = str(params["image_document_id"]);
+            const inline = rec(params["document"]);
+            const raw = id
+                ? needSketch(world, id)
+                : {
+                    id: str(inline["id"], "inline"),
+                    name: str(inline["name"], "inline"),
+                    layers: recList(inline["layers"]).map((layer) => ({
+                        id: str(layer["id"]),
+                        name: str(layer["name"]),
+                        opacity: num(layer["opacity"], 1),
+                        blendMode: str(layer["blendMode"], "normal"),
+                        visible: layer["visible"] !== false
+                    })),
+                    activeLayerId: str(inline["activeLayerId"])
+                };
+            const issues = sketchIssues(raw);
+            return { ok: issues.length === 0, issues, warnings: [] };
+        }),
+        tool("edit_sketch", "Apply layer-structure edits to a saved sketch.", obj({ image_document_id: S, ops: ANY_ARRAY }, [
+            "image_document_id",
+            "ops"
+        ]), (params) => {
+            const doc = needSketch(world, str(params["image_document_id"]));
+            const ops = recList(params["ops"]);
+            if (ops.length === 0)
+                throw new Error("edit_sketch needs at least one op");
+            ops.forEach((op, index) => {
+                try {
+                    applySketchOp(doc, op);
+                }
+                catch (e) {
+                    throw new Error(`edit_sketch: op ${index + 1} (${str(op["op"], "?")}): ` +
+                        (e instanceof Error ? e.message : String(e)));
+                }
+            });
+            return { ok: true, applied: ops.length, document: clone(doc) };
+        }),
+        tool("list_sketch_versions", "List an image document's snapshot history.", obj({ image_document_id: S, limit: N }, ["image_document_id"]), (params) => {
+            const id = str(params["image_document_id"]);
+            needSketch(world, id);
+            return {
+                image_document_id: id,
+                versions: (world.versions.get(`sketch:${id}`) ?? []).map((v) => ({
+                    version: v.version,
+                    name: v.name,
+                    save_type: v.save_type
+                }))
+            };
+        }),
+        tool("get_sketch_version", "Read one sketch snapshot without restoring it.", obj({ image_document_id: S, version: N }, [
+            "image_document_id",
+            "version"
+        ]), (params) => {
+            const id = str(params["image_document_id"]);
+            const version = num(params["version"], 0);
+            const found = (world.versions.get(`sketch:${id}`) ?? []).find((v) => v.version === version);
+            if (!found)
+                throw new Error(`no version ${version} for sketch "${id}"`);
+            return {
+                image_document_id: id,
+                version,
+                document: clone(found.document)
+            };
+        }),
+        tool("create_sketch_version", "Snapshot a sketch's current document.", obj({ image_document_id: S, name: S }, ["image_document_id"]), (params) => {
+            const id = str(params["image_document_id"]);
+            const doc = needSketch(world, id);
+            const entry = snapshotVersion(world, `sketch:${id}`, str(params["name"], "manual save"), "manual", doc);
+            return {
+                image_document_id: id,
+                version: entry.version,
+                name: entry.name
+            };
+        }),
+        tool("restore_sketch_version", "Restore a sketch snapshot, snapshotting the current state first.", obj({ image_document_id: S, version: N }, [
+            "image_document_id",
+            "version"
+        ]), (params) => {
+            const id = str(params["image_document_id"]);
+            const doc = needSketch(world, id);
+            const version = num(params["version"], 0);
+            const found = (world.versions.get(`sketch:${id}`) ?? []).find((v) => v.version === version);
+            if (!found)
+                throw new Error(`no version ${version} for sketch "${id}"`);
+            snapshotVersion(world, `sketch:${id}`, "before restore", "restore", doc);
+            const restored = clone(found.document);
+            world.sketches.set(id, restored);
+            const issues = sketchIssues(restored);
+            return {
+                image_document_id: id,
+                restored: version,
+                validation: { ok: issues.length === 0, issues }
+            };
+        }),
+        tool("delete_sketch_version", "Delete one snapshot. The live document is not changed.", obj({ image_document_id: S, version: N }, [
+            "image_document_id",
+            "version"
+        ]), (params) => {
+            const id = str(params["image_document_id"]);
+            needSketch(world, id);
+            const version = num(params["version"], 0);
+            const list = world.versions.get(`sketch:${id}`) ?? [];
+            const idx = list.findIndex((v) => v.version === version);
+            if (idx < 0)
+                throw new Error(`no version ${version} for sketch "${id}"`);
+            list.splice(idx, 1);
+            world.versions.set(`sketch:${id}`, list);
+            return { ok: true, image_document_id: id, deleted_version: version };
+        }),
+        // -- scripts ----------------------------------------------------------
+        tool("list_scripts", "List saved scripts.", obj({ limit: N }), () => ({
+            scripts: [...world.scripts.values()].map((doc) => ({
+                script_id: doc.script_id,
+                name: doc.name,
+                lines: doc.lines.length,
+                voiced: doc.lines.filter((l) => l.status === "voiced").length
+            }))
+        })),
+        tool("get_script", "One script: cast, lines, and each line's voicing status.", obj({ script_id: S }, ["script_id"]), (params) => {
+            const id = str(params["script_id"]);
+            const doc = world.scripts.get(id);
+            if (!doc)
+                throw new Error(`no script with id "${id}"`);
+            return clone(doc);
+        }),
+        tool("voice_script_lines", "Synthesize a take per line; defaults to every draft or stale line.", obj({
+            script_id: S,
+            targets: ANY_ARRAY,
+            provider: S,
+            model: S,
+            voice: S,
+            transcribe: B
+        }, ["script_id"]), (params) => {
+            const id = str(params["script_id"]);
+            const doc = world.scripts.get(id);
+            if (!doc)
+                throw new Error(`no script with id "${id}"`);
+            const requested = strList(params["targets"]);
+            const targets = requested.length > 0
+                ? requested.map((target) => {
+                    const line = doc.lines.find((l) => l.id === target);
+                    if (!line)
+                        throw new Error(`no line with id "${target}"`);
+                    return line;
+                })
+                : doc.lines.filter((l) => l.status !== "voiced");
+            const voiced = [];
+            const skipped = [];
+            for (const line of targets) {
+                if (line.status === "voiced" && requested.length === 0) {
+                    skipped.push(line.id);
+                    continue;
+                }
+                line.status = "voiced";
+                line.take = `asset://take_${line.id}`;
+                line.duration_seconds = Math.max(1, Math.round(line.text.length / 10));
+                voiced.push(line.id);
+            }
+            return { script_id: id, voiced, skipped };
+        }),
+        tool("assemble_script_timeline", "Lay the voiced takes end to end into a saved timeline sequence.", obj({ script_id: S, name: S }, ["script_id"]), (params) => {
+            const id = str(params["script_id"]);
+            const doc = world.scripts.get(id);
+            if (!doc)
+                throw new Error(`no script with id "${id}"`);
+            const voiced = doc.lines.filter((l) => l.status === "voiced");
+            if (voiced.length === 0) {
+                throw new Error(`script "${id}" has no voiced takes to assemble`);
+            }
+            const timelineId = `tl_${id}`;
+            let startMs = 0;
+            const clips = voiced.map((line) => {
+                const clip = {
+                    id: `clip_${line.id}`,
+                    trackId: "track_voiceover",
+                    name: line.text.slice(0, 24),
+                    startMs,
+                    durationMs: line.duration_seconds * 1000,
+                    animations: []
+                };
+                startMs += clip.durationMs;
+                return clip;
+            });
+            world.timelines.set(timelineId, {
+                id: timelineId,
+                name: `${doc.name} voiceover`,
+                tracks: [{ id: "track_voiceover", type: "audio", name: "Voiceover" }],
+                clips
+            });
+            return {
+                timeline_id: timelineId,
+                clips: clips.length,
+                skipped: doc.lines.length - voiced.length,
+                duration_seconds: startMs / 1000
+            };
+        }),
+        tool("edit_script", "Apply cast/line edits to a saved script.", obj({ script_id: S, ops: ANY_ARRAY }, ["script_id", "ops"]), (params) => {
+            const id = str(params["script_id"]);
+            const doc = world.scripts.get(id);
+            if (!doc)
+                throw new Error(`no script with id "${id}"`);
+            const ops = recList(params["ops"]);
+            if (ops.length === 0)
+                throw new Error("edit_script needs at least one op");
+            ops.forEach((op, index) => {
+                try {
+                    applyScriptOp(doc, op);
+                }
+                catch (e) {
+                    throw new Error(`edit_script: op ${index + 1} (${str(op["op"], "?")}): ` +
+                        (e instanceof Error ? e.message : String(e)));
+                }
+            });
+            return { ok: true, applied: ops.length, document: clone(doc) };
+        }),
+        // -- storyboards ------------------------------------------------------
+        tool("list_storyboards", "List saved storyboards.", obj({ limit: N }), () => ({
+            storyboards: [...world.storyboards.values()].map((doc) => ({
+                storyboard_id: doc.storyboard_id,
+                name: doc.name,
+                shots: doc.shots.length,
+                stills: doc.shots.filter((s) => s.still !== null).length,
+                clips: doc.shots.filter((s) => s.clip !== null).length
+            }))
+        })),
+        tool("create_storyboard", "Create a blank storyboard.", obj({ name: S, brief: S, style: S, aspect_ratio: S, id: S }, ["name"]), (params) => {
+            const id = str(params["id"], `sb_${world.storyboards.size + 1}`);
+            if (world.storyboards.has(id)) {
+                const existing = world.storyboards.get(id);
+                return {
+                    ok: true,
+                    storyboard_id: existing.storyboard_id,
+                    name: existing.name
+                };
+            }
+            const name = str(params["name"], "Untitled");
+            const doc = {
+                storyboard_id: id,
+                name,
+                image_model: { provider: "fal_ai", model: "fal-ai/flux/schnell" },
+                video_model: { provider: "fal_ai", model: "fal-ai/ltx" },
+                shots: []
+            };
+            world.storyboards.set(id, doc);
+            return { ok: true, storyboard_id: id, name, shots: 0 };
+        }),
+        tool("get_storyboard", "One board: its shots, their status, and its models.", obj({ storyboard_id: S }, ["storyboard_id"]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            return {
+                storyboard_id: doc.storyboard_id,
+                name: doc.name,
+                image_model: doc.image_model,
+                video_model: doc.video_model,
+                shots: doc.shots.map((shot) => ({
+                    id: shot.id,
+                    slug: shot.slug,
+                    action: shot.action,
+                    duration_seconds: shot.duration_seconds,
+                    duration_source: shot.duration_source,
+                    has_still: shot.still !== null,
+                    has_clip: shot.clip !== null,
+                    revision: shot.revision
+                }))
+            };
+        }),
+        tool("render_storyboard_stills", "Render a keyframe per shot; defaults to every shot without one.", obj({ storyboard_id: S, targets: ANY_ARRAY, provider: S, model: S }, [
+            "storyboard_id"
+        ]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            const requested = resolveShots(doc, params["targets"]);
+            const targets = requested.length > 0
+                ? requested
+                : doc.shots.filter((s) => s.still === null);
+            const rendered = [];
+            for (const shot of targets) {
+                shot.still = `asset://still_${shot.id}`;
+                rendered.push(shot.id);
+            }
+            return {
+                storyboard_id: id,
+                rendered,
+                model: str(params["model"], doc.image_model.model)
+            };
+        }),
+        tool("render_storyboard_clips", "Animate each shot's keyframe into a clip; a shot without a still is skipped.", obj({ storyboard_id: S, targets: ANY_ARRAY, provider: S, model: S }, [
+            "storyboard_id"
+        ]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            const requested = resolveShots(doc, params["targets"]);
+            const targets = requested.length > 0
+                ? requested
+                : doc.shots.filter((s) => s.clip === null);
+            const rendered = [];
+            const skipped = [];
+            for (const shot of targets) {
+                if (shot.still === null) {
+                    skipped.push({
+                        id: shot.id,
+                        reason: "no keyframe — render its still first"
+                    });
+                    continue;
+                }
+                shot.clip = `asset://clip_${shot.id}`;
+                rendered.push(shot.id);
+            }
+            return {
+                storyboard_id: id,
+                rendered,
+                skipped,
+                model: str(params["model"], doc.video_model.model)
+            };
+        }),
+        tool("revise_storyboard_clip", "Revise one shot's clip.", obj({ storyboard_id: S, target: S, instruction: S }, [
+            "storyboard_id",
+            "target",
+            "instruction"
+        ]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            const [shot] = resolveShots(doc, [str(params["target"])]);
+            if (shot.clip === null)
+                throw new Error(`shot "${shot.id}" has no clip to revise`);
+            shot.revision++;
+            shot.clip = `asset://clip_${shot.id}_r${shot.revision}`;
+            return { storyboard_id: id, revised: shot.id, revision: shot.revision };
+        }),
+        tool("assemble_storyboard_timeline", "Lay the rendered clips into a saved timeline sequence.", obj({ storyboard_id: S, name: S }, ["storyboard_id"]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            const shots = doc.shots.filter((s) => s.clip !== null);
+            if (shots.length === 0) {
+                throw new Error(`board "${id}" has no rendered clips to assemble`);
+            }
+            const timelineId = `tl_${id}`;
+            let startMs = 0;
+            const clips = shots.map((shot) => {
+                const clip = {
+                    id: `clip_${shot.id}`,
+                    trackId: "track_video",
+                    name: shot.slug,
+                    startMs,
+                    durationMs: shot.duration_seconds * 1000,
+                    animations: []
+                };
+                startMs += clip.durationMs;
+                return clip;
+            });
+            world.timelines.set(timelineId, {
+                id: timelineId,
+                name: `${doc.name} cut`,
+                tracks: [{ id: "track_video", type: "video", name: "Video" }],
+                clips
+            });
+            return {
+                timeline_id: timelineId,
+                clips: clips.length,
+                skipped: doc.shots.length - shots.length,
+                duration_seconds: startMs / 1000
+            };
+        }),
+        tool("edit_storyboard", "Apply shot-list edits to a saved board.", obj({ storyboard_id: S, ops: ANY_ARRAY }, ["storyboard_id", "ops"]), (params) => {
+            const id = str(params["storyboard_id"]);
+            const doc = world.storyboards.get(id);
+            if (!doc)
+                throw new Error(`no storyboard with id "${id}"`);
+            const ops = recList(params["ops"]);
+            if (ops.length === 0)
+                throw new Error("edit_storyboard needs at least one op");
+            ops.forEach((op, index) => {
+                try {
+                    applyStoryboardOp(doc, op);
+                }
+                catch (e) {
+                    throw new Error(`edit_storyboard: op ${index + 1} (${str(op["op"], "?")}): ` +
+                        (e instanceof Error ? e.message : String(e)));
+                }
+            });
+            return { ok: true, applied: ops.length, document: clone(doc) };
+        })
+    ];
+}
+// ---------------------------------------------------------------------------
+// cases
+// ---------------------------------------------------------------------------
+const field = (result, key) => rec(result)[key];
+const asNumber = (value) => isNumber(value) ? value : Number.NaN;
+const asString = (value) => isString(value) ? value : "";
+export const CODEACT_API_SURFACE_CASES = [
+    {
+        id: "game-whole-document-save",
+        description: "Write and read a complete game level in one atomic edit",
+        namespaces: ["games"],
+        objective: 'Create a native game named "Long Level". Replace its entire draft ' +
+            "in one set_document operation with schema version 2 and a scene named level containing " +
+            "117 entities: one camera, one solid terrain tilemap using the built-in " +
+            "wall asset, and 115 entities named item-0 through item-114. The terrain " +
+            "must have 3,446 tiles. Use the created draft timestamp as the edit guard. " +
+            "Read the saved document back. Finish with {gameId, entityCount, tileCount, " +
+            "entrySceneId, preservedIdentity}, where preservedIdentity confirms the " +
+            "saved document kept the created game id and revision.",
+        outputSchema: obj({ gameId: S, entityCount: N, tileCount: N, entrySceneId: S, preservedIdentity: B }, [
+            "gameId", "entityCount", "tileCount", "entrySceneId", "preservedIdentity"
+        ]),
+        expect: {
+            requiredTools: ["create_native_game", "edit_native_game", "get_native_game"],
+            maxActions: 3,
+            minToolCalls: 3,
+            maxToolCalls: 3,
+            resultCheck: (r) => asString(field(r, "gameId")) === "game_1" &&
+                asNumber(field(r, "entityCount")) === 117 &&
+                asNumber(field(r, "tileCount")) === 3446 &&
+                asString(field(r, "entrySceneId")) === "level" &&
+                field(r, "preservedIdentity") === true,
+            resultCheckLabel: "one whole-document save preserves identity and 117 entities / 3446 tiles"
+        }
+    },
+    {
+        id: "rag-index-and-answer",
+        description: "Index a collection's documents, then answer from vector search",
+        namespaces: ["collections"],
+        objective: "The `handbook` collection holds our support policy, one document per " +
+            "topic. Put those documents into the vector index — each under its own " +
+            "document id as the source id — then use vector search to find how many " +
+            "days the refund window is. Finish with {sourceId, days}: the source id " +
+            "of the chunk that answered it, and the number of days as a number.",
+        outputSchema: obj({ sourceId: S, days: N }, ["sourceId", "days"]),
+        expect: {
+            requiredTools: ["query_collection", "vector_text_search"],
+            maxActions: 5,
+            resultCheck: (r) => asString(field(r, "sourceId")) === "policy-refunds" &&
+                asNumber(field(r, "days")) === 30,
+            resultCheckLabel: "sourceId=policy-refunds days=30"
+        }
+    },
+    {
+        id: "memory-lifecycle",
+        description: "Save a memory, correct it in place, and read it back",
+        namespaces: ["memory"],
+        objective: 'Remember for this conversation, under the title "preferred image ' +
+            'model", that the house image model is fal_ai/flux/schnell. Then the ' +
+            "user corrects you: it is really fal_ai/flux/dev. Update the note you " +
+            "just saved in place — do not delete it or add a second one. Finish " +
+            "with {memoryId, content, " +
+            "total} — the id of that note, its content as the memory list now " +
+            "reports it, and the FULL count of memories the list holds, unrelated " +
+            "notes included.",
+        outputSchema: obj({ memoryId: S, content: S, total: N }, [
+            "memoryId",
+            "content",
+            "total"
+        ]),
+        expect: {
+            requiredTools: [
+                "memory_save",
+                "memory_update",
+                "memory_list"
+            ],
+            maxActions: 4,
+            resultCheck: (r) => {
+                const content = asString(field(r, "content"));
+                return (asString(field(r, "memoryId")).startsWith("mem_") &&
+                    content.includes("flux/dev") &&
+                    !content.includes("schnell") &&
+                    asNumber(field(r, "total")) === 2);
+            },
+            resultCheckLabel: "corrected note, 2 memories total"
+        }
+    },
+    {
+        id: "threads-recall",
+        description: "Find a past conversation and read what it actually did",
+        namespaces: ["threads"],
+        objective: "In an earlier conversation we made the launch banner. Find that " +
+            "thread among the past ones, read it, and report which image model the " +
+            "banner was really generated with — the model id is in the arguments " +
+            "of the tool call that made it, not in the text, so read that message " +
+            "in full rather than guessing from the transcript. Finish with " +
+            "{threadId, model}.",
+        outputSchema: obj({ threadId: S, model: S }, ["threadId", "model"]),
+        expect: {
+            requiredTools: ["list_threads", "get_thread", "get_message"],
+            maxActions: 4,
+            resultCheck: (r) => asString(field(r, "threadId")) === "th_banner" &&
+                asString(field(r, "model")) === "fal_ai/flux/dev",
+            resultCheckLabel: "th_banner generated with fal_ai/flux/dev"
+        }
+    },
+    {
+        id: "shared-handoff",
+        description: "Read what upstream steps shared, then publish a derived value",
+        namespaces: ["shared"],
+        objective: "Earlier steps of this run left their results in shared memory. Find " +
+            "the pricing result among them and read it — the numbers are only " +
+            "there, so do not invent them. Work out the annual price at twelve " +
+            "times the monthly one, and publish it into shared memory under the " +
+            'key "annual_price" so later steps can use it. Finish with ' +
+            "{monthlyUsd, annualUsd, publishedKey} — the monthly price you read, " +
+            "the annual price you worked out, and the FULL key the publish call " +
+            "reported back.",
+        outputSchema: obj({ monthlyUsd: N, annualUsd: N, publishedKey: S }, [
+            "monthlyUsd",
+            "annualUsd",
+            "publishedKey"
+        ]),
+        expect: {
+            requiredTools: ["list_shared", "read_shared", "share_result"],
+            maxActions: 4,
+            resultCheck: (r) => asNumber(field(r, "monthlyUsd")) === 49 &&
+                asNumber(field(r, "annualUsd")) === 588 &&
+                asString(field(r, "publishedKey")) === "shared:annual_price",
+            resultCheckLabel: "49/588 published as shared:annual_price"
+        }
+    },
+    {
+        id: "web-research-brief",
+        description: "Search the web, open the page, and report facts only the page carries",
+        namespaces: ["web"],
+        objective: "Find our NodeTool 4.2 release-notes page on the web and read it. " +
+            "Finish with {url, version, sandbox}: the page's URL, the release " +
+            "version, and the name of the JavaScript engine the release runs agent " +
+            "actions in. The engine is not in the search snippet — you have to open " +
+            "the page.",
+        outputSchema: obj({ url: S, version: S, sandbox: S }, [
+            "url",
+            "version",
+            "sandbox"
+        ]),
+        expect: {
+            requiredTools: ["web_search"],
+            maxActions: 4,
+            resultCheck: (r) => asString(field(r, "url")) ===
+                "https://example.dev/nodetool/release-notes" &&
+                String(field(r, "version")) === "4.2" &&
+                asString(field(r, "sandbox")).toLowerCase().includes("quickjs"),
+            resultCheckLabel: "release-notes url, 4.2, QuickJS"
+        }
+    },
+    {
+        id: "document-extraction-pipeline",
+        description: "PDF → text, tables, and markdown in one pass",
+        namespaces: ["documents"],
+        objective: "The workspace holds reports/q3.pdf. Pull its tables out to " +
+            "reports/q3-tables.json, convert the PDF itself to markdown at " +
+            "reports/q3.md, and read the report's text. Finish with {accounts, " +
+            "tableRows, markdownFile}: how many accounts the quarter's revenue " +
+            "spanned (a number, no separators), how many table rows were " +
+            "extracted, and the markdown file you wrote.",
+        outputSchema: obj({ accounts: N, tableRows: N, markdownFile: S }, [
+            "accounts",
+            "tableRows",
+            "markdownFile"
+        ]),
+        expect: {
+            requiredTools: [
+                "extract_pdf_text",
+                "extract_pdf_tables",
+                "convert_pdf_to_markdown"
+            ],
+            maxActions: 4,
+            resultCheck: (r) => asNumber(field(r, "accounts")) === 1280 &&
+                asNumber(field(r, "tableRows")) === 3 &&
+                asString(field(r, "markdownFile")) === "reports/q3.md",
+            resultCheckLabel: "accounts=1280 rows=3 reports/q3.md"
+        }
+    },
+    {
+        id: "timeline-fix-and-validate",
+        description: "Snapshot, repair a failing timeline, and re-validate clean",
+        namespaces: ["timelines"],
+        objective: "The timeline sequence tl_promo does not pass validation. Snapshot it " +
+            "first so the repair is undoable, then fix what validation complains " +
+            "about with a document edit, and check it again. Finish with " +
+            "{issuesBefore, issuesAfter, trackCount, snapshotVersion}: the issue " +
+            "counts before and after your edit, how many tracks the document ends " +
+            "up with, and the version number your snapshot got.",
+        outputSchema: obj({ issuesBefore: N, issuesAfter: N, trackCount: N, snapshotVersion: N }, ["issuesBefore", "issuesAfter", "trackCount", "snapshotVersion"]),
+        expect: {
+            requiredTools: [
+                "validate_timeline",
+                "create_timeline_version",
+                "edit_timeline"
+            ],
+            maxActions: 5,
+            resultCheck: (r) => asNumber(field(r, "issuesBefore")) === 1 &&
+                asNumber(field(r, "issuesAfter")) === 0 &&
+                asNumber(field(r, "trackCount")) === 2 &&
+                asNumber(field(r, "snapshotVersion")) === 1,
+            resultCheckLabel: "1 issue → 0, 2 tracks, snapshot v1"
+        }
+    },
+    {
+        id: "sketch-layer-repair",
+        description: "Fix a dangling active layer and add a styled layer",
+        namespaces: ["sketches"],
+        objective: "The image document img_cover does not pass validation. Fix it, and " +
+            'while you are in there add a layer named "Glow" at 0.4 opacity in the ' +
+            "multiply blend mode and leave it as the document's active layer. " +
+            "Finish with {issuesBefore, issuesAfter, layerCount, glowOpacity, " +
+            "activeLayerId} read back from the document.",
+        outputSchema: obj({
+            issuesBefore: N,
+            issuesAfter: N,
+            layerCount: N,
+            glowOpacity: N,
+            activeLayerId: S
+        }, [
+            "issuesBefore",
+            "issuesAfter",
+            "layerCount",
+            "glowOpacity",
+            "activeLayerId"
+        ]),
+        expect: {
+            requiredTools: ["validate_sketch", "edit_sketch"],
+            maxActions: 5,
+            resultCheck: (r) => asNumber(field(r, "issuesBefore")) === 1 &&
+                asNumber(field(r, "issuesAfter")) === 0 &&
+                asNumber(field(r, "layerCount")) === 3 &&
+                asNumber(field(r, "glowOpacity")) === 0.4 &&
+                asString(field(r, "activeLayerId")).length > 0 &&
+                asString(field(r, "activeLayerId")) !== "layer_ghost",
+            resultCheckLabel: "1 issue → 0, 3 layers, glow active at 0.4"
+        }
+    },
+    {
+        id: "script-voice-and-assemble",
+        description: "Voice only the lines that need it, then assemble the voiceover",
+        namespaces: ["scripts", "timelines"],
+        objective: "Script scr_intro has lines that still need recording. Record exactly " +
+            "those — leave the ones already voiced alone — then assemble the " +
+            "voiceover timeline. Finish with {voiced, clips, timelineId}: the ids " +
+            "you recorded (sorted), how many clips the assembled sequence holds, " +
+            "and its timeline id.",
+        outputSchema: obj({ voiced: ANY_ARRAY, clips: N, timelineId: S }, [
+            "voiced",
+            "clips",
+            "timelineId"
+        ]),
+        expect: {
+            requiredTools: ["voice_script_lines", "assemble_script_timeline"],
+            maxActions: 4,
+            resultCheck: (r) => {
+                const voiced = strList(field(r, "voiced")).slice().sort().join(",");
+                return (voiced === "line_2,line_3" &&
+                    asNumber(field(r, "clips")) === 3 &&
+                    asString(field(r, "timelineId")) === "tl_scr_intro");
+            },
+            resultCheckLabel: "voiced line_2,line_3; 3 clips in tl_scr_intro"
+        }
+    },
+    {
+        id: "storyboard-direct-shots",
+        description: "Direct a board's shot list headlessly, before anything renders",
+        namespaces: ["storyboards"],
+        objective: "Board sb_lighthouse needs one more beat. Add a closing shot — the " +
+            "lamp goes dark — as the last shot, 2 seconds long, timed off its " +
+            "audio rather than a fixed length, and drop the shot showing the " +
+            "keeper closing the door. Change nothing else and render nothing. " +
+            "Finish with {shots, lastAction, lastDurationSource} read back from " +
+            "the board.",
+        outputSchema: obj({ shots: N, lastAction: S, lastDurationSource: S }, ["shots", "lastAction", "lastDurationSource"]),
+        expect: {
+            requiredTools: ["edit_storyboard", "get_storyboard"],
+            forbiddenTools: ["render_storyboard_stills", "render_storyboard_clips"],
+            maxActions: 4,
+            resultCheck: (r) => asNumber(field(r, "shots")) === 3 &&
+                /lamp/i.test(asString(field(r, "lastAction"))) &&
+                asString(field(r, "lastDurationSource")) === "audio",
+            resultCheckLabel: "3 shots, closing lamp shot timed off audio"
+        }
+    },
+    {
+        id: "storyboard-render-and-assemble",
+        description: "Stills, then clips, then a cut — in the order the board allows",
+        namespaces: ["storyboards", "timelines"],
+        objective: "Board sb_lighthouse is directed but unfinished. Every shot needs a " +
+            "still and then a clip, and a clip can only be animated from a shot's " +
+            "still. Get all three shots to a clip and assemble the cut. Finish " +
+            "with {shotsWithClips, timelineId, durationSeconds} read back from the " +
+            "board and the assembled sequence.",
+        outputSchema: obj({ shotsWithClips: N, timelineId: S, durationSeconds: N }, ["shotsWithClips", "timelineId", "durationSeconds"]),
+        expect: {
+            requiredTools: [
+                "render_storyboard_stills",
+                "render_storyboard_clips",
+                "assemble_storyboard_timeline"
+            ],
+            maxActions: 5,
+            resultCheck: (r) => asNumber(field(r, "shotsWithClips")) === 3 &&
+                asString(field(r, "timelineId")) === "tl_sb_lighthouse" &&
+                asNumber(field(r, "durationSeconds")) === 12,
+            resultCheckLabel: "3 clips, tl_sb_lighthouse, 12s"
+        }
+    },
+    {
+        id: "settings-and-credentials",
+        description: "Change a setting, and ask the user for the credential you cannot set",
+        namespaces: ["settings"],
+        objective: "Prepare this install for billing work. Turn autosave off, and make " +
+            "sure the Stripe credential is available — you cannot write a " +
+            "credential yourself, so ask the user for it and say why. Finish with " +
+            "{autosave, missingBefore, stripeReady}: the value autosave now holds, " +
+            "the names of the credentials that were NOT configured before you " +
+            "started, and whether Stripe is configured afterwards.",
+        outputSchema: obj({ autosave: S, missingBefore: ANY_ARRAY, stripeReady: B }, [
+            "autosave",
+            "missingBefore",
+            "stripeReady"
+        ]),
+        expect: {
+            requiredTools: ["list_secrets", "request_secret", "set_setting"],
+            maxActions: 4,
+            resultCheck: (r) => asString(field(r, "autosave")) === "false" &&
+                strList(field(r, "missingBefore")).join(",") === "STRIPE_API_KEY" &&
+                field(r, "stripeReady") === true,
+            resultCheckLabel: "autosave off, Stripe asked for and now configured"
+        }
+    },
+    {
+        id: "email-triage",
+        description: "Search, label, and archive the matching messages",
+        namespaces: ["email"],
+        objective: "Triage the mailbox: every message from the last 48 hours whose " +
+            'subject mentions an invoice gets the label "billing" and is then ' +
+            "archived. Finish with {labeled, archived, remaining}: the ids you " +
+            "labeled (sorted), how many messages you archived, and how many " +
+            "matching messages are still sitting in the inbox afterwards.",
+        outputSchema: obj({ labeled: ANY_ARRAY, archived: N, remaining: N }, [
+            "labeled",
+            "archived",
+            "remaining"
+        ]),
+        expect: {
+            requiredTools: ["search_email", "add_label_to_email", "archive_email"],
+            maxActions: 4,
+            resultCheck: (r) => {
+                const labeled = strList(field(r, "labeled")).slice().sort().join(",");
+                return (labeled === "msg_1,msg_2" &&
+                    asNumber(field(r, "archived")) === 2 &&
+                    asNumber(field(r, "remaining")) === 0);
+            },
+            resultCheckLabel: "msg_1+msg_2 labeled and archived, inbox clear"
+        }
+    },
+    {
+        id: "app-wiring-check",
+        description: "Check a saved app's wiring for free",
+        namespaces: ["apps"],
+        objective: 'The app "app_notes" drafts a short note from a prompt. Run the FREE ' +
+            "wiring check on that app — no run, it costs nothing and spends " +
+            "nothing. Finish with {verdictOk, ran, widgets} — the check's verdict, " +
+            "whether it executed the app, and how many widgets it reported.",
+        outputSchema: obj({ verdictOk: B, ran: B, widgets: N }, [
+            "verdictOk",
+            "ran",
+            "widgets"
+        ]),
+        expect: {
+            requiredTools: ["debug_app"],
+            maxActions: 5,
+            resultCheck: (r) => field(r, "verdictOk") === true &&
+                field(r, "ran") === false &&
+                asNumber(field(r, "widgets")) === 3,
+            resultCheckLabel: "free check ran, 3 widgets"
+        }
+    }
+].map((entry) => ({ ...entry, createTools: createSurfaceApiTools }));
+//# sourceMappingURL=codeact-api-surfaces.js.map