@kenkaiiii/ggcoder 5.66.0 → 5.66.2

package/assets/motion/bin/motion-check.mjs

@@ -102,12 +102,18 @@ export function parseMotionOutput(text, slideshowRequested = false, holds = [])
       error:
         "Holds cannot exempt the entire video; explicitly requested slideshows use their own mode",
     };
-  if (
-    holds.some(
-      (hold) => !freezes.some((freeze) => freeze.end > hold.start && freeze.start < hold.end),
-    )
-  )
-    return { ok: false, error: "Stale hold declaration: no detected freeze overlaps its window" };
+  const staleHolds = holds.filter(
+    (hold) => !freezes.some((freeze) => freeze.end > hold.start && freeze.start < hold.end),
+  );
+  // Name the stale holds and where pixels actually freeze, so the plan is fixed in one
+  // step instead of guessed again against another full check.
+  if (staleHolds.length)
+    return {
+      ok: false,
+      error: "Stale hold declaration: no detected freeze overlaps its window",
+      staleHolds,
+      freezes,
+    };
   const unexpectedFreezes = freezes.filter(
     (freeze) =>
       !holds.some(

package/assets/motion/skills/motion/SKILL.md

@@ -81,7 +81,9 @@ images yourself against the `Concept` and `Language` in `frame.md`. That one too
 runs the technical checks; do not repeat them or ask another model to review
 them.
 
-Fix concrete defects, render and check the changed output. An unchanged export
-needs no repeated checking. Deliver the MP4, reveal it and state real limits.
+Fix concrete defects and render. For a targeted fix or small edit, spot-check
+the changed moments first (`spot: true`), then run one full check on the export
+you deliver. An unchanged export needs no repeated checking unless its hold plan
+changed. Deliver the MP4, reveal it and state real limits.
 Sampled frames are not full playback or audio listening; technical success is
 not proof of visual quality.

package/assets/motion/skills/video-qa/SKILL.md

@@ -17,6 +17,7 @@ Call `motion_check` with:
 - `holds`: only when the video has a deliberate static section (an end card, reading hold or pause you designed), the path to its render-bound hold plan (format below). Write it before this first call; an undeclared deliberate hold is flagged as frozen and costs a second full check.
 - `slideshowRequested`: true only when the user actually requested a slideshow; never use it to hide broken motion.
 - `range`: only for an export longer than 180 seconds or a targeted diagnostic, a start/end range of at most 180 seconds. Report the inspected range honestly; do not claim full-video visual coverage.
+- `spot`: true for a quick check of a targeted fix or small edit. Put short `windows` around the moments you changed, at most 240 rendered frames in total (4 s at 60 fps). The layout check then covers every rendered frame inside them, far faster than the full check, while the pixel, hold and audio checks still cover the whole export. A spot result is never delivery verification: once it is clean, run the full check (without `spot`) once on the export you deliver.
 
 The tool runs HyperFrames `check` (which already includes lint), decoded-pixel analysis with `motion-check.mjs` including canvas/WebGL, and audio analysis only when audio exists. It returns technical results and actual rendered overview, phone-size and consecutive-frame images **to you**, the working agent.
 
@@ -38,7 +39,7 @@ If the tool fails, its check details list each problem with where and when it oc
 
 ## 3. Fix only a concrete defect
 
-A failed check or visibly wrong output warrants a targeted fix, a new versioned render and a check of that changed output. An unchanged export does not need checking again. Do not iterate toward subjective perfection or route the result to another reviewer. Report an unresolved blocker as draft/unverified rather than looping indefinitely or declaring success.
+A failed check or visibly wrong output warrants a targeted fix, a new versioned render and a spot check of the fixed moments, then one full check once they are clean. An unchanged export does not need checking again unless you corrected its hold plan; that re-check reuses the passing source audit, so it is quick. Do not iterate toward subjective perfection or route the result to another reviewer. Report an unresolved blocker as draft/unverified rather than looping indefinitely or declaring success.
 
 The tool measures audio and rejects non-finite levels or clipping; it does not normalize the file. Follow the user's delivery loudness target when they give one. Do not add audio to silent work or force every supplied animation through a generic -14 LUFS mix.
 
@@ -54,7 +55,7 @@ Declare only deliberate static intervals: an end card, a reading hold or another
 }
 ```
 
-Do not mark the whole video as a hold to suppress broken animation. Retain any existing meaningful motion assertions; do not invent a new assertion sidecar for every video.
+When a hold is stale, the check lists where the pixels actually freeze; correct the plan from those times instead of guessing. Do not mark the whole video as a hold to suppress broken animation. Retain any existing meaningful motion assertions; do not invent a new assertion sidecar for every video.
 
 ## 4. Deliver
 

package/dist/core/agent-session.d.ts

@@ -21,6 +21,13 @@ export interface SessionAttachment {
     name: string;
     path?: string;
 }
+/** Per-prompt run controls. */
+export interface PromptRunOptions {
+    /** Offer the model no tools for this prompt. */
+    disableTools?: boolean;
+    /** Hold reasoning effort at the plan-mode ceiling for this prompt. */
+    capThinking?: boolean;
+}
 export interface AgentSessionOptions {
     provider: Provider;
     model: string;
@@ -477,10 +484,10 @@ export declare class AgentSession {
     willExpandPromptTemplate(content: string): Promise<boolean>;
     /**
      * Process user input. Handles slash commands or runs agent loop.
+     * `capThinking` holds reasoning effort at the plan-mode ceiling for this
+     * prompt — for turns that must answer quickly from what is already known.
      */
-    prompt(content: string, provenance?: MessageProvenance, options?: {
-        disableTools?: boolean;
-    }): Promise<void>;
+    prompt(content: string, provenance?: MessageProvenance, options?: PromptRunOptions): Promise<void>;
     /**
      * Prompt with multimodal attachments (images / videos) alongside optional
      * text. Images and videos become native content blocks the model can see;

package/dist/core/agent-session.js

@@ -48,7 +48,7 @@ import { evaluateIdealReview, buildIdealReviewMessage, buildReviewCoverageEscala
 import { evaluateLoopBreak, buildLoopBreakMessage, CycleDetector, ToolCallProgressTracker, detectTextRepetition, } from "./loop-breaker.js";
 import { buildRegroundingMessage, requestTextForRegrounding } from "./regrounding.js";
 import { buildSemanticLoopJudgePrompt, buildSemanticLoopMessage, MAX_SEMANTIC_LOOP_CALLS, parseSemanticLoopVerdict, shouldRunSemanticLoopCheck, SEMANTIC_LOOP_JUDGE_TIMEOUT_MS, withJudgeTimeout, } from "./semantic-loop-check.js";
-import { buildIndependentReviewMessage, buildReviewerTask, INDEPENDENT_REVIEW_SCORE_THRESHOLD, parseReviewerFindings, REVIEWER_TOOLS, REVIEWER_WAIT_MS, } from "./ideal-review-subagent.js";
+import { buildIndependentReviewMessage, buildReviewerTask, INDEPENDENT_REVIEW_SCORE_THRESHOLD, parseReviewerFindings, REVIEWER_TOOLS, REVIEWER_TURN_TIMEOUT_MS, REVIEWER_WAIT_MS, } from "./ideal-review-subagent.js";
 import { buildEnvDeltaMessage } from "./env-delta.js";
 import { wrapSteeringText, buildNotificationSteeringText, STEERING_PREFIX } from "./steering.js";
 import { AgentNotificationQueue } from "./agent-notifications.js";
@@ -911,6 +911,8 @@ export class AgentSession {
     }
     /**
      * Process user input. Handles slash commands or runs agent loop.
+     * `capThinking` holds reasoning effort at the plan-mode ceiling for this
+     * prompt — for turns that must answer quickly from what is already known.
      */
     async prompt(content, provenance = {
         source: "human",
@@ -1645,9 +1647,12 @@ export class AgentSession {
                 triggerReasons: decision.reasons,
             });
             // Active model forced at spawn time — never routed to a fast/review model.
+            // The reviewer's own time limit ends it with a verdict on what it read;
+            // the wait below is only a backstop against a hung child.
             const snapshot = await this.subAgentManager.spawn(taskName, task, undefined, {
                 model: this.model,
                 tools: REVIEWER_TOOLS,
+                turnTimeoutMs: REVIEWER_TURN_TIMEOUT_MS,
             });
             agentId = snapshot.agent_id;
             const waited = await this.subAgentManager.wait([agentId], "all", REVIEWER_WAIT_MS);
@@ -1663,7 +1668,11 @@ export class AgentSession {
             }
             const findings = parseReviewerFindings(agent.output ?? "");
             if (!findings) {
-                log("WARN", "ideal", "Independent reviewer output unparseable; falling back", { agentId });
+                log("WARN", "ideal", "Independent reviewer output unparseable; falling back", {
+                    agentId,
+                    state: agent.state,
+                    ...(agent.error ? { error: agent.error } : {}),
+                });
                 return [];
             }
             if (findings.clean) {
@@ -2132,8 +2141,9 @@ export class AgentSession {
                 // Plan mode caps effort at medium (Codex `plan_mode_reasoning_effort`
                 // preset): read-only exploration doesn't need xhigh/max reasoning, and
                 // deep-reasoning models left at the ceiling burn enormous thinking
-                // budgets re-deriving context they cannot act on.
-                thinking: this.planModeRef.current
+                // budgets re-deriving context they cannot act on. A capped prompt
+                // (a sub-agent's timed answer) gets the same ceiling for the same reason.
+                thinking: this.planModeRef.current || options.capThinking
                     ? clampThinkingForPlanMode(this.thinkingLevel)
                     : this.thinkingLevel,
                 apiKey,