agent-nuvira 3.5.6 → 3.5.9

package/dist/tools/registry.js

@@ -261,6 +261,45 @@ const imageGenSchema = z.object({
         .optional()
         .describe('Optional backend override: gemini (Imagen/Nano Banana), openai (DALL-E/gpt-image), stability, comfyui (local), or pollinations (free). Defaults to the first backend whose API key is configured.'),
 });
+/**
+ * Motion tool args — code-authored animation rendered to a video file.
+ *
+ * Kept deliberately narrow: the SOURCE (html / html_path / preset+data) and the
+ * handful of settings that change the output, because a long schema is prompt
+ * budget spent on every turn whether or not motion is being rendered.
+ */
+const motionRenderSchema = z.object({
+    action: z
+        .enum(['render', 'revise', 'list', 'presets'])
+        .optional()
+        .describe("Default 'render'. 'revise' re-renders a previous render (source_id) with edits — deterministic because the stored source is re-rendered, not re-invented; 'list' shows this workspace's renders; 'presets' lists the ready-made scenes."),
+    html: z.string().optional().describe('A self-contained HTML animation (no network: inline CSS/JS/SVG). Must be seekable — see the motion skill.'),
+    html_path: z.string().optional().describe('Path to an HTML animation file instead of inline html'),
+    preset: z
+        .enum(['title-card', 'bar-race', 'line-reveal', 'code-walkthrough'])
+        .optional()
+        .describe('Use a ready-made scene instead of hand-written html: title-card, bar-race, line-reveal (charts from your data), code-walkthrough'),
+    params: z.record(z.string(), z.unknown()).optional().describe('Preset parameters: title, subtitle, kicker, accent, background, foreground, unit, decimals'),
+    data: z.unknown().optional().describe('Data to animate: JSON (array of numbers, [{label,value}], {series:[…]}) or CSV text'),
+    data_path: z.string().optional().describe('Path to a CSV/JSON file to animate (bar-race, line-reveal, code-walkthrough)'),
+    source_id: z.string().optional().describe("Render id to revise (from a previous motion_render or action:'list')"),
+    title: z.string().optional().describe('Convenience title (merged into preset params)'),
+    fps: z.number().int().min(1).max(60).optional().describe('Frames per second (default 30)'),
+    duration_ms: z.number().int().min(200).max(600000).optional().describe('Length in ms (default 8000)'),
+    width: z.number().int().min(64).max(3840).optional().describe('Frame width (default 1280)'),
+    height: z.number().int().min(64).max(2160).optional().describe('Frame height (default 720)'),
+    aspect_ratio: z.string().optional().describe("Shape instead of a size, e.g. '16:9', '9:16', '1:1' (720p/1080p by default)"),
+    format: z.enum(['mp4', 'gif', 'webm']).optional().describe('Container (default mp4; gif carries no audio)'),
+    crf: z.number().int().min(0).max(51).optional().describe('Quality: lower is better (default 20)'),
+    narration: z.string().optional().describe('Narration text, synthesized locally (edge-tts/Piper) and muxed onto the video'),
+    voice: z.string().optional().describe('TTS voice id for narration'),
+    captions: z
+        .array(z.object({ from_ms: z.number(), to_ms: z.number(), text: z.string() }))
+        .optional()
+        .describe('Caption cues, written as a .srt next to the video'),
+    burn_captions: z.boolean().optional().describe('Burn the captions into the frames (needs an ffmpeg with libass)'),
+    out_name: z.string().optional().describe('Filename prefix for the artifact (defaults to the preset name)'),
+});
 /** I4 — voice tools: TTS + transcription (optional installs). */
 const speakSchema = z.object({
     text: z.string().min(1).describe('The text to speak'),
@@ -880,7 +919,10 @@ registerTool({
     endsAgentStep: false,
     run: async (args, ctx) => {
         const { action, goal, steps, id, status, note } = planTodoSchema.parse(args);
-        const { defaultPlanStore } = await import('./plan-store.js');
+        const { defaultPlanStore, planProvenanceFromActions } = await import('./plan-store.js');
+        // R1 — the tool result line that tells the model its claim was recorded as
+        // UNVERIFIED rather than as a done step. Empty for every other update.
+        let assertedNote = '';
         const store = ctx.planStore ?? defaultPlanStore();
         // C4 — a re-declaration of the SAME goal carries the progress of the steps it
         // re-declares (see PlanStore.create). Disclosed rather than silent: a model
@@ -926,7 +968,26 @@ registerTool({
                         `Create them in ${ctx.cwd}, then mark the step done — a step is not done while its artifact is missing.`);
                 }
             }
-            store.update(resolved, status, note);
+            // ── R1 — WHAT STOOD BEHIND THIS CLAIM ───────────────────────────────────
+            // Marking a step done is a claim about the past, and the defect this closes
+            // is measured: a plan table reported `14 failed / 422 passed ✅ done` before
+            // any command in the transcript had run. When the loop supplied its action
+            // ledger, the claim is recorded AGAINST it — the backing call becomes the
+            // step's evidence, and an empty ledger marks the step `asserted`, which
+            // renders ⚠️ rather than ✅ from then on.
+            //
+            // The update is never REFUSED. A step whose proof is a decision rather than
+            // an execution is legitimate, and a tool that argues with the model instead
+            // of informing the reader would be a worse bug than the one being fixed;
+            // what changes is that the claim can no longer LOOK measured. The model is
+            // told which way it was recorded, so it can go and produce the evidence.
+            const provenance = planProvenanceFromActions(ctx.executedActions);
+            store.update(resolved, status, note, provenance);
+            if (status === 'done' && provenance.hasLedger && !provenance.evidence) {
+                assertedNote =
+                    `\n⚠️ Step "${resolved}" is recorded as ASSERTED, not verified: no tool call has run in this turn, so nothing backs the claim yet. ` +
+                        `Run the check that produces the result, then update the step again — or leave it asserted and say so in your answer.`;
+            }
         }
         else {
             return 'Error: plan_todo update needs id + status (pending|running|done|blocked).';
@@ -939,7 +1000,7 @@ registerTool({
                 ctx.emit?.('plan:changed', snapshot);
             }
         }
-        return (store.toText ? store.toText() : `Plan updated (${action}).`) + carriedNote;
+        return (store.toText ? store.toText() : `Plan updated (${action}).`) + carriedNote + assertedNote;
     },
 });
 // ─── WS1 (#23) finding tool — a claim, and the verdict the gate computed ────
@@ -2795,6 +2856,42 @@ registerTool({
         }
     }),
 });
+// Motion — animation authored as code, rendered to a real video file.
+registerTool({
+    name: 'motion_render',
+    capability: { effectClass: 'local-write', reversible: true },
+    description: 'Render an ANIMATED video from code: a self-contained HTML/CSS/SVG animation (html or html_path) or a ready-made scene (preset + data). Use for animated explainers, animated charts of the user\'s own data, kinetic titles and code walkthroughs — the output is deterministic (frames are captured from the animation timeline, not recorded), so a revise re-renders the stored source. Needs chromium (playwright) and ffmpeg; a missing one is named with its install command. Optional narration is synthesized locally and muxed in; captions are written as .srt. The video is saved to the sandbox motion/ dir and, in a gateway context, sent back to the originating channel.',
+    category: 'workflow',
+    inputSchema: motionRenderSchema,
+    endsAgentStep: false,
+    run: async (args, ctx) => {
+        const parsed = motionRenderSchema.parse(args);
+        const { runMotionTool } = await import('./modality/motion-tool.js');
+        const result = await runMotionTool(parsed, ctx);
+        if (result.media) {
+            // Same artifact path as generate_image/speak: a render is a deliverable, so
+            // the dashboard's Artifacts tab can show (and re-open) it.
+            recordArtifact(ctx?.artifacts, {
+                kind: 'media',
+                title: result.media.title,
+                path: result.media.path,
+                mime: result.media.mime,
+            });
+            if (result.sidecar) {
+                recordArtifact(ctx?.artifacts, {
+                    kind: 'file',
+                    title: result.sidecar.title,
+                    path: result.sidecar.path,
+                    mime: result.sidecar.mime,
+                });
+            }
+            const caption = parsed.title ?? parsed.narration ?? undefined;
+            const delivered = await tryAutoDeliver(result.media.path, caption, ctx, 'video');
+            return `${result.text}${delivered}`;
+        }
+        return result.text;
+    },
+});
 // Voice mode tool
 registerTool({
     name: 'voice_mode',