Output quality: 60fps capture, no ghosting, natural camera and typing by Co-Messi · Pull Request #14 · Co-Messi/supercut

Co-Messi · 2026-09-30T12:34:26Z

Output-quality pass on capture → plan → compositor. Each item was measured first: a failing check in the new motion-quality suite or a failing e2e test, then the fix.

Items

#	Problem	Fix	Test	Commit
1	Nothing measured motion quality	`test/motion-quality.test.ts` + `test/helpers/motion-metrics.ts` score `buildRenderPlan` on synthetic takes (39fps and 60fps source, a click-triggered nav gap, typing, a scene reload, a punch at the end of the take) and on the real demo take from main. Checks that failed on main ran as `it.fails` and each fix commit flipped its own.	the suite itself	`a9fb580`, `968bff9`
2	Source ~39–53 fps	JPEG q92 screencast at DPR 2 (resolution unchanged). Also: the 1px repaint beacon stopped registering damage when a hovered, transformed row met a timer re-setting identical text, and the source fell to 20fps. It now covers the whole viewport at 1–2e-4 opacity, which moves no 8-bit channel. The renderer picks the frame MIME type by extension, so older PNG takes still render.	`frameMimeType` unit; e2e checks for `.jpg` frames; new e2e holds ≥45fps on a hovered dashboard (was 20.7)	`d98dbfd`
3	Double exposures from cross-blending	Every gap is floor-held and a page change is a hard cut. The plan has no pixels to prove two frames near-identical, and blending near-identical frames changes nothing visible. The `blend` array stays (all -1/0) so the host-page contract is unchanged.	3 old cross-blend plan tests rewritten to assert hold/cut (reason in the commit); blend-share checks	`52ba6b9`
4	Stale zoom across scenes, late punches, mid-move ending	Page boundaries come from scene markers plus any source gap ≥250ms. A scene opens at its first NEW frame. Shots end 800ms before a scene change, with no bridge and no glide across a boundary, and the spring snaps to wide at a real cut. The establishing shot is 800ms. Lead is 750ms, and a punch that can't reach 90% by its event is skipped. Tail is ≥1000ms plus ≥1700ms past the last punch. Picture fades in 0.6s and out 1.8s, the same constants as the music afades. Executor: 1s pre-roll at the take head and after each new page paints; the entry goto is skipped when `page.url()` already equals it. Also: fast local click-navigations leave no source gap because of paint holding, so the recorder now logs a strict `navigation` event (no URL) for navigations it didn't issue, and the plan cuts wide on it.	motion suite; new plan tests (no bridge across a nav gap, skip late punch, settled tail, fades, gapless logged nav); e2e: pre-roll, same-URL scene has no reload gap, one `navigation` logged for a clicked link and none for entry gotos	`e72c947`, `bb58ad4`, `a90d35f` (an unexplained gap must be ≥500ms to count as a cut, so a capture hiccup never snaps the camera on the same page)
5	Soft/aliased text, onion-ring blur, 8-bit dimming	`imageSmoothingQuality="high"`. Each source is pre-resampled once to 1.5× the content width. Blur pass count comes from the max displacement of all 4 corners, rounded up to a power of two and capped at 32. The accumulator is float16: in the bundled Chromium, 48 passes of white at 1/48 give 240/255 in 8-bit, while float16 with power-of-two passes stays within 1 level. The 8-bit fallback caps at 8 passes.	`blurPassCount` unit tests (the page embeds the same function via `toString`)	`abfa408`
6	Metronomic typing, instant scroll jumps, zero-length clicks	Seeded `typingPlan`: log-normal gaps around ~100ms, ×1.8 after spaces and punctuation, 45ms floor, mean never below 60ms (no paste), 250–400ms before the first key, 250–350ms before Enter. Eased 350–900ms page scroll that centres the target. 100ms settle before the press, 70ms hold.	`typingPlan` unit tests; new `/keys` fixture reports real pointer, key and scroll timings to an e2e	`55f2826`
7	Reasoning model returned empty `length` ×4 and the run failed	On `finish_reason: "length"` with empty content, `max_tokens` doubles (8k→16k→32k, ceiling 4×). A 400 on an escalated size falls back to the last accepted size. `BudgetedLlmClient` reserves `escalationCeiling(maxTokens)`.	escalation sequence, ceiling, no escalation on non-truncated empties, 400 fallback, budget reserve	`165b095`

Metrics

Regression suite: same take files, main's planner vs this branch

metric (target)	real demo take, main	real demo take, branch	synthetic 39fps, main	synthetic 39fps, branch
z at scene marker (≤1.05)	1.42	1.013	1.42	1.014
z on first new-page frame after nav (≤1.05)	1.10	1.00	1.42	1.00
min wide rest per scene (≥700ms)	0 ms	1050 ms	717 ms	867 ms
min punch arrival at click (≥0.90 or skipped)	0.98*	0.96	0.90	0.94
tail max \|Δz\|/frame (<1e-4)	~0	0	4.0e-3 (final z 1.009)	5.5e-5 (final z 1.0001)
blended frames (<5%)	13.4%	0%	52.1%	0%

* In the first commit the arrival window read the next beat's lead-in, so main's first click scored 0%. The window was tightened so it no longer does. As a result it can't see main's late punch on that click, which overlapped the next beat.

Real record + render, demo recipe (`record` then `render --music pulse`)

	main @ `013d6b8`	this branch
capture health (avg source fps)	52.7 (20fps for the last 0.7s)	57.9 (continuous 60 except the one reload gap)
ffprobe	h264 1920×1080 60/1, 10.05s	h264 1920×1080 60/1, 11.83s
motion: z at marker / after nav / wide rest / blended	1.42 / 1.10 / 0ms / 3.1%†	1.014 / 1.00 / 1017ms / 0%
`freezedetect n=-60dB d=0.5`	2.23 (0.55s), 2.78 (0.53s), 4.57 (1.18s), 7.13 (0.58s), 7.72→end (~2.3s)	3.50 (0.75s), 7.13 (1.32s), 8.45 (1.55s)

† main's planner scored on the branch's JPEG take.

The branch's freezedetect stretches are static shots, not capture stalls. The source is continuous 60fps there. They cover the typing hold and the dashboard's wide establishing shot plus the hover hold on a full-width row, which gets no punch. The dashboard's small counter changes fall under the −60dB threshold.

Real `generate` on the demo app (DeepSeek v4-pro, text-only)

The run includes a click on a nav link, a same-URL next scene (no reload now), and a typed form. Results: capture 58.6fps; z at scene markers 1.014; z after the logged click-navigation 1.42 → 1.00 (main planner vs branch on the same take); wide rest ≥917ms per scene; min arrival 0.96; final z 1.0001 at rest; 0% blended. ffprobe: h264 1920×1080 60/1, 17.6s. The LLM usage for this run (~23.5k tokens) did not trigger max_tokens escalation.

Behaviour changes to be aware of

Token budget: each call now reserves 4× its maxTokens, the escalation ceiling. Analyze and script each reserve about 32k plus the prompt, so --max-tokens values below roughly 40–50k now refuse the first call up front.
Take length: the head pre-roll adds 1s. Each reloaded scene adds up to about 0.8s, and typing at a human pace can overrun short slots. The demo went from 10.05s to 11.83s, so a recipe budgeted at exactly 60s can render a little past 60s.
Schema: Event-Log v0 gains an optional navigation event. Old parsers drop it with their forward-compat warning. TypeScript consumers that switch exhaustively on KnownEvent will need a new case.

Not verified

DeepSeek's actual output cap at 16k/32k max_tokens. The 400 fallback is covered by unit tests only, and no real run truncated.
The sharpness gain at z=1 is not obvious in still crops. What is measured: pass spacing ≤1px at every corner, and float16 accumulation confirmed in the render log (blur accumulator: float16).
The pulse dashboard from the original report was not re-run (its app isn't in this repo).

🤖 Generated with Claude Code

Scores buildRenderPlan on a synthetic multi-scene take (click-triggered navigation gap, typing, scene reload, end-of-take punch) at ~39fps and ~60fps source cadence, plus the real demo-recipe take recorded on main @ 013d6b8 (events + frame index only). Asserts: z <= 1.05 at scene markers and on the first new-page frame after a nav gap; >= 700ms wide rest per scene; punches reach >= 90% by the click (or are skipped); |dz| < 1e-4 per frame over the final 300ms; < 5% of frames blend different sources. Checks that fail on main run as it.fails (PENDING table); each fix commit removes its entries. Baseline on the real take: z at scene marker 1.42, z after nav gap 1.10, 0ms wide rest on scene 2, first click arrival 0%, 13.4% blended frames (52% on the 39fps synthetic). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a9fb580e2b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-09-30T12:39:25Z

+  const sceneStarts = markers.map((m, i) => {
+    if (i === 0) return 0;
+    const g = gaps.find((g) => g.tA >= m - MARKER_WINDOW_BEFORE_MS && g.tA <= m + MARKER_WINDOW_AFTER_MS);
+    return g ? g.tB : m;


Include unmarked navigation gaps in wide-rest checks

When navigation is triggered by a click without a scene marker—as the synthetic fixture deliberately does between 5200 ms and 5900 ms—this list contains no start for the newly loaded page. Consequently, wideRestMs only measures the take head and marked reload, so the wide-rest test can pass even if the camera is wide for only one frame after the unmarked navigation and immediately zooms again; include navigation gaps not already attributed to a marker as scene starts.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-09-30T12:39:26Z

+/** the real demo-recipe take recorded on main @ 013d6b8 (events + frame
+ *  index only — the frames themselves are not needed to score the plan) */
+function demoTake(): { log: EventLog; index: FrameIndexEntry[] } {
+  const dir = join(import.meta.dirname, "fixtures", "takes", "demo-main");


Avoid using import.meta.dirname at the Node 20 floor

On Node 20.0–20.10, which is allowed by the package's node >=20 engine constraint, import.meta.dirname is unavailable, so join() receives undefined and this suite fails while constructing the real-demo scenario. Resolve the directory from import.meta.url with fileURLToPath instead, or raise the declared minimum Node version to 20.11.

Useful? React with 👍 / 👎.

…60fps source - Screencast JPEG q92 at DPR 2 (resolution unchanged) instead of PNG; frames are written as frames/NNNNNN.jpg. - The renderer serves frames by extension (frameMimeType), so takes recorded as PNG before this change still render. - Repaint beacon now covers the viewport at 1-2e-4 opacity (no 8-bit channel moves). The 1px corner beacon stopped registering damage on the demo dashboard once a hovered row met a timer re-setting identical text: the source fell to that timer's 20Hz for the rest of the scene. Demo recipe source fps (capture health): 52.7 -> 57.5 avg including the scene-change reload gap; no sub-60 stretch left outside the nav gap. Tests: frameMimeType unit test; record e2e asserts .jpg frames and a new e2e holds >= 45fps over a hovered-dashboard hold (was 20.7fps). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Linear cross-blending across any 25-500ms source gap (and a 350ms late crossfade across navigation/long gaps) mixed two DIFFERENT source frames: a visible double exposure on 13% of the real demo take's frames and 52% of a 39fps take's. The plan has no pixels to prove two frames near-identical, and blending near-identical frames is a visual no-op, so every gap is now floor-held and a page change is a clean cut. The blend array stays in the plan (all -1/0) so the host-page contract is unchanged. Behaviour change: the three "source cross-blend" plan tests encoded the old dissolve; they are rewritten to assert the hold/cut (short residual gap, nav gap, irregular sub-60fps source) and the dense-60fps case is kept. Motion suite: blended share 13.4% / 52.1% -> 0% (blend checks flipped). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…, settled ending Render plan: - Page boundaries = scene markers + any source gap >= 250ms (catches click-triggered navigations the event log never marks). A marker's scene opens at the first NEW frame after its reload gap, not at the marker (emitted before the goto, while frames are dropped). - Shots never outlive a boundary: the zoom-out completes 800ms before a scene change (or as the old page freezes for a click-nav); no bridging and no GLIDE_Z park on a stale focus across a boundary; the spring state snaps to wide at a real cut, hidden by the picture change. - Establishing shot (800ms at z=1) anchored to the new page's first frame. - ZOOM_LEAD 600 -> 750ms; a punch that cannot reach 90% by its event (ARRIVE 600ms) is skipped instead of landing after the click, as is one the next navigation would cut to < 300ms after the event. - Tail >= 1000ms and the take runs >= 1700ms past the last punch so the camera is at rest at the end; plan.fade drives a picture fade in/out from black with the same lengths as the music afades (shared constants). Executor: - 1000ms pre-roll at the take head and after each new page has painted, so the establishing shot plays and the first punch arrives before the first click. - The next scene's entry goto is skipped when page.url() already equals it (and the previous scene completed): no redundant reload/freeze. Behaviour change: the establishing-shot plan test asserted the camera "glides" at z > 1.05 between scenes (6000ms); it now asserts z < 1.02 there and at the next scene marker. The motion metric's click window no longer reaches into the next beat's lead-in. Motion suite (real demo take): z at scene marker 1.42 -> 1.01, z after nav gap 1.10 -> 1.00, scene-2 wide rest 0 -> 4300ms, all PENDING flipped. Real record+render: marker 1.01, arrival min 0.96, final z 1.000 at rest. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…at16 accumulation - imageSmoothingQuality "high" on both canvases (the default low/bilinear filter aliased text when downscaling the 3840px source). - Each source frame is pre-resampled once (resizeQuality high) to 1.5x the content width, so every blur pass draws a <= 1.5x downscale. - Blur pass count from the MAX displacement of all four content corners (was top-left only: up to ~7px ghost spacing on the far side during zooms), rounded up to a power of two and capped at 32. - Accumulate in a float16 canvas when available: in 8-bit, 48 passes of white at 1/48 summed to 240/255 (probed in the bundled Chromium); float16 with power-of-two passes stays within 1 level. 8-bit fallback caps at 8. blurPassCount is exported and embedded verbatim into the host page. Demo render time 2.7s -> 3.7s for 711 frames (float16 path confirmed in the render log). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…licks - typingPlan (seeded): log-normal inter-key gaps around ~100ms, x1.8 after spaces/punctuation, 45ms floor, mean compressed no lower than 60ms on a short slot (overruns shift the schedule instead of pasting), a 250-400ms beat before the first key and 250-350ms before Enter. Replaces the uniform min(90, remaining/len) metronome. - Off-screen targets are reached by an eased (cubic in-out, 350-900ms) page scroll that centres them, filmed as motion; scrollIntoViewIfNeeded stays as a backstop for nested scroll containers. Post-scroll settle 350->150ms. - The pointer rests 100ms on the target before pressing and holds the press 70ms (was press/release back to back). Tests: typingPlan unit tests; a new /keys fixture route reports pointer, press, keystroke, Enter and scroll timings to the test server, and a record e2e asserts the eased scroll (>= 5 intermediate positions), >= 80ms settle, >= 55ms hold, a beat before the first key and Enter, varied >= 35ms gaps. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…an honest budget reserve A reasoning model (DeepSeek v4-pro on the pulse dashboard) spent all 8000 max_tokens thinking and returned empty content with finish_reason "length" on every one of the 4 attempts, failing the run. Retrying at the same size mostly truncates again. - OpenAICompatibleClient doubles max_tokens after an empty answer with finish_reason "length" (8k -> 16k -> 32k), capped at escalationCeiling() = 4x the requested size. Empty answers that were not truncated retry at the same size. - If the provider rejects an escalated size with a 400, the client falls back to the largest size it accepted and stops escalating (instead of failing the run on a provider output cap). - BudgetedLlmClient reserves escalationCeiling(maxTokens) for the call's completion, so the pre-send refusal covers the escalated attempt. Tests: escalation sequence, ceiling, no escalation on non-truncated empties, 400 fallback, budget reserve of the ceiling. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Found verifying on a real generate run: a click on a nav link navigated with NO source gap (Chromium paint holding keeps a fast local navigation streaming at 60fps), so the gap-based boundary never fired and the camera carried its punch into the old page's link straight onto the new page. - Event-Log Schema v0 gains an optional, strict `navigation` event ({t, observed_t}, no URL). Additive: older parsers drop unknown types with their existing forward-compat warning. - The recorder logs it when a main-frame navigation request the recipe did not issue (entry/goto navigations are wrapped and excluded) commits. - The plan treats it as a snap boundary at the commit time (unless its reload also left a source gap, which already covers it): the navigating click's punch is skipped (it would last < 300ms), and the new page opens wide with its establishing read. - resolveFocus finds the action's own event even if a navigation was logged after it. Tests: schema accepts the event (and stays strict), plan cuts wide at a gapless logged navigation, record e2e logs exactly one navigation for a clicked link and none for the next scene's entry goto. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The "wide on the first new-page frame" metric only looked at source gaps; it now also checks the frame after each logged action-triggered navigation (gapless fast navigations). On the real generate take whose first scene clicks a nav link: main's planner 1.42 -> 1.00 on this branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

With action-triggered navigations now logged, a bare source gap no longer has to stand in for them. Gaps near a scene marker or a logged navigation still cut at >= 250ms; an unattributed gap must be >= 500ms (above the ~400ms stall ceiling the capture e2e documents). A 300ms hiccup mid-shot on a slow machine previously truncated the punch and snapped z ~1.18 -> 1.0 with the same page still showing. Tests: new plan test (300ms unlogged stall: no discontinuity, punch kept); the gap-bridging test now logs its navigation, as the recorder does; the motion metric applies the same attribution rule. CI robustness: the same-URL e2e now counts GET /dash hits on the fixture server (1, not 2) instead of asserting a < 250ms max frame gap, and the hovered-dashboard fps floor is 35 (CI's software compositor sustains ~47; the regression was 20). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Messi · 2026-10-03T00:29:44Z

Independent review — progress log (updated as findings land)

[started] reviewing at a90d35f
Item 5 (pulse): real generate on the pulse dashboard at a90d35f succeeded: 3 scenes, 940 frames, avg 60.1 fps source, QC clean, ~12.2k LLM tokens (no max_tokens failure).
Item 1 (framing) confirmed, worse than reported: on the gen2 take the plan showed up to 307px of wallpaper on the right while the left overflowed (1.5–2.1s, z=1.42), and 131px on top at 7s.
Item 1 fixed in 176aca8: a shared cameraTransform in plan.ts, embedded in the host page via toString for the blur passes and the cursor. It centres the focus and clamps per axis: the window stays on canvas below coverage and covers the canvas above it, and the range ramps up from 0 at z=1. Failing-first framing check (it.fails on 3 scenarios plus a new edge/corner take) is now at 0px one-sided wallpaper, 0px uncovered, centring error 0. Re-rendered gen2: zoomed shots fill the frame. Gates green.
Snap smear fixed in 92b4876: at gen2's logged click-navigation (~2.15s), the camera snap z=1.42→1 fell inside one frame's shutter, so that frame was a motion-blur smear. This is in the PR's own render too. Snaps now land on frame boundaries (failing-first maxIntraFrameDz on a gen2-shaped gapless-nav scenario). Residual, not fixed: a gapless logged navigation is stamped at commit, and paint holding keeps the old page on screen ~2 more frames (commit 2139.8ms, first new-page frame 2171.9ms). The old page shows wide for ~33ms before the cut.
Item 5 beacon: Playwright probe on pulse (Next.js) and the demo app, with and without the beacon. Header screenshots byte-identical. elementFromPoint identical at 4 points. Scroll dims unchanged. No console errors or hydration warnings. Minor, not fixed: on plain pages the beacon becomes body's last element child. pulse freezedetect: longest static stretch 2.6s at 8.95s (wide dashboard; every scene-2/3 click carries focus_bbox = the full-height main panel, so z fits to ≤1 and no punch is planned).
Item 6 mutation check: 13 mutants on plan.ts. All killed except one equivalent mutant (removing the establishing segment; the start clamp already enforces it). The narrowed arrival window hid main's late first-click punch; added the maxLateRise check in 546da4d (0.016 on the branch vs 0.060 under main's start rule).
Item 7 fixed in e8fde15: one call can run 4 attempts (8k+16k+32k+32k ≈ 88k + 4 prompts) against a 4×maxTokens reserve, so it could overshoot the "hard" ceiling by ~56k. The 4× reserve also refused the first call for any --max-tokens < ~40k. The wrapper now passes the budget left as spendLimit, and the client checks it before every attempt. The 400 fallback is correct.
Item 8 fixed in 11b0131: the recipe ceiling ignored the take's fixed overhead. A 58s 4-scene recipe validated but would render ~65s. parseRecipe now enforces estimated video length ≤ 60s. Also moved 10 stray (gitignored) .bak files out of the checkout.
Ramp test cfb4568: mutant gx=gy=1 survived all 407 tests; two cameraTransform unit tests now kill it.
(log rebuilt once: an API EOF during an update had truncated it; the helper now refuses to patch on a failed read)

… wallpaper slab The compositor offset fx(1−z) + (cx−fx)(1−1/z) moved the focus only ~30% toward centre at z=1.42 and had no bounds: on the real gen2 take the plan shoved the window off the left edge while 307px of wallpaper showed on the right (and 131px on top at another punch). cameraTransform (plan.ts) now aims the focus at the canvas centre and clamps the offset per axis: content stays fully on canvas while z·content < canvas and covers it once z·content ≥ canvas. The legal range ramps open from the plain scale-about-centre offset (g = 0 at z=1 → 1 at coverage), so wide shots stay exactly centred and a z≈1.1 glide only nudges the window. The host page embeds the same function via toString for both the blur passes and the cursor, replacing two inline copies of the old formula. Failing-first: new framing check in the motion-quality suite (one-sided wallpaper, uncovered-while-coverable, focus-centring error vs the closest feasible position) ran as it.fails on all three scenarios plus a new edge/corner-targets take; all four now pass at 0px. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

A logged click-navigation with no source gap snaps the camera from a held punch straight to wide at the commit time. That time fell inside an output frame's recorded shutter half, so the compositor's motion blur averaged the whole window from z=1.42 to z=1 into one smeared flash frame — visible at ~2.15s of the PR's own gen2 render. Snaps now apply at the start of the first output frame at/after the cut, so every frame is wholly before or after the jump. For gap boundaries this is the same frame that first shows the new page. Failing-first: new maxIntraFrameDz metric (z range across one frame's shutter samples; a moving spring spans < 0.01) on a scenario shaped like gen2's first beat (hover punch, click, logged nav 80ms later, no gap) measured a straddled snap; it now stays < 0.02 on every scenario. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The arrival metric's window was cut off at the next beat's lead-in, so on main's demo take the first click's late punch (starting ~260ms AFTER the click) read as "never punched" and scored as skipped. New maxLateRise metric: zoom gained from the click to 400ms later, stopped where the next beat's 750ms lead-in could begin, so only this click's own punch counts. Branch: 0.016 on the real demo take. Mutation with main's start rule (lead 600, late start allowed, no skip): 0.060 on that same take, and the new check fails on all five scenarios (arrival alone failed on three). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…n budget BudgetedLlmClient checked the budget once per call and reserved the 4x escalation ceiling (32k at maxTokens 8000), but one call runs up to four attempts — 8k → 16k → 32k → 32k bills ~88k plus four prompts — so a single truncating call could overshoot the advertised "hard" ceiling by ~56k. The up-front 4x reserve also refused the first call outright for any --max-tokens below ~40k, even when no escalation was ever needed. The wrapper now passes the budget left as ChatOptions.spendLimit and the OpenAI-compatible client checks it before every attempt: an escalated max_tokens shrinks to the room left, and when even the requested size no longer fits it stops with TokenBudgetExceededError. The wrapper's pre-check reserves only the first attempt (prompt + maxTokens). Failing-first: a provider that truncates every attempt and bills it in full metered ~88k against a 40k budget (now ≤ 40k, TokenBudgetExceededError); a 20k budget refused the call up front (now attempts 8k, then escalates only into the remaining room). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The recipe ceiling summed only action durations and holds, but a take adds a 1s head pre-roll, ~1.5s per later scene (entry reload + 400ms settle + 1s pre-roll on the new page, at least the 1s entry allowance) and a ~1.7s settled tail. A 4-scene recipe at 58s therefore validated yet rendered ~65s, breaking the README's ≤60s promise (real runs: pulse 11.9s of recipe → 16.65s video, gen2 13.1s → 17.6s). parseRecipe now enforces estimatedTakeMs (budgets + head + per-scene change + tail) ≤ 60s. The script stage's own target (≤ 50s over ≤ 4 scenes, ≈ 57.2s estimated) still fits; an action that overruns its slot can still add a little. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The framing checks only scored coverage and one-sided exposure, so setting the ramp to full strength at every z (gx = gy = 1) survived the whole suite: wide shots would slide sideways while the focus spring returns, and a z≈1.1 glide would jam the window against an edge. Two unit tests on cameraTransform: at z=1 the offset is exactly zero for focus points on every edge and corner; at z=1.1 the smaller margin stays ≥ half the centred margin while the window still moves toward an off-centre focus. The gx = gy = 1 mutant now fails both. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Messi · 2026-10-03T02:10:51Z

Independent review: summary (head `cfb4568`)

I checked the PR's claims against real renders (contact sheets and full-res stills of the branch vs main), real generate runs, and mutation checks. The motion-quality direction is right and clearly better than main. Main's render has a crossfade ghost frame, a zoomed first frame on the new page, and an abrupt ending; the branch has none of these. I found and fixed six more issues, listed below.

Fixed on this branch

commit	issue	failing-first evidence
`176aca8`	Zoomed framing pushed the window off one edge while wallpaper showed on the other. On the gen2 take this reached 307px of wallpaper (z=1.42). The fix adds a shared `cameraTransform` that centres the focus and clamps per axis: the window stays on canvas until the content can cover it, then covers it. The ramp keeps wide shots exactly centred. The host page embeds the same function (toString) for the blur passes and the cursor.	New framing check: it.fails on all scenarios plus an edge/corner take, now 0px
`92b4876`	Smeared flash frame at a gapless click-navigation. The camera snap z=1.42→1 landed mid-shutter, so motion blur averaged the whole window across the jump. It is visible in the PR's own gen2 render at ~2.15s. Snaps now land on frame boundaries.	maxIntraFrameDz on a gen2-shaped scenario
`546da4d`	The narrowed arrival window hid late punches. Main's late first-click punch read as "skipped". New `maxLateRise` check: 0.016 on the branch vs 0.060 under main's start rule.	mutation
`e8fde15`	The "hard" token budget could be overshot by ~56k inside one call. A call makes up to 4 attempts (8k+16k+32k+32k). The 4× up-front reserve also refused the first call for any `--max-tokens` below ~40k. The wrapper now passes the budget left (`spendLimit`), and the client keeps every attempt and escalation inside it. This corrects the PR description: a call no longer reserves 4× maxTokens, and small `--max-tokens` values are no longer refused up front; escalation shrinks to the room left.	88k metered vs 40k budget → now ≤ budget
`11b0131`	Videos could exceed 60s. The recipe ceiling ignored the take's fixed overhead (1s pre-roll, ~1.5s per scene change, ~1.7s tail), so a 58s 4-scene recipe validated and rendered at ~65s. `parseRecipe` now enforces the estimated video length. Checked against real runs, the estimate is about 1s high, which is the safe direction: pulse 17.6s estimated vs 16.65s actual, gen2 18.8s vs 17.6s. The LLM's ≤50s target still fits.	schema test
`cfb4568`	The framing ramp had no test. A mutant applying full centring at every z survived all 407 tests; two unit tests now kill it.	mutation

Verdicts on the PR's claims

60fps capture: pulse generate reached avg 60.1fps source and the demo 57.9. The full-viewport beacon is pixel-neutral (byte-identical screenshots), does not intercept hit-testing (elementFromPoint unchanged), and leaves layout and scroll unchanged. It caused no console or hydration errors on Next.js.
Pulse generate (previously failed on max_tokens): succeeded with 3 scenes, QC clean, ~12.2k tokens and no escalation.
ffprobe: every branch render is h264 1920×1080 60/1.
Hard cut at page changes (item 3): keep it. The cut is clean and wide-to-wide, while main's crossfade produced a double-exposure frame. I did not add a crossfade.
0.75–2.6s still stretches (item 4): acceptable. On pulse, about 8s pass without camera motion because the director gave every click focus_bbox = the full-height main panel, so the fit gives z ≤ 1. That is a follow-up for the director, not a planner bug.
Float16 blur: confirmed in the render log. I saw no dimming dip during zooms.
Mutation check: 13 real mutants, all killed. M4 (removing the establishing segment) is an equivalent mutant: the start clamp already enforces the same thing. One mis-built mutant (+0.3-frame snap threshold) was not a real revert and was discarded; its true revert is killed.

Residuals (not fixed)

A gapless logged navigation is stamped at commit, and paint holding keeps the old page on screen ~2 frames longer. The old page therefore shows wide for ~33ms before the cut.
On plain pages the beacon becomes body's last element child, which could affect a :last-child-style selector.
Typing or slot overruns, and slow remote reloads (nav time + 1.4s per scene), can still push a video a little past 60s.

Gates at cfb4568: typecheck ✓ · vitest 409/409 ✓ · build ✓ · npm audit 0 vulns ✓ · check:pack ✓ · CI all green, including browser-e2e.

Recommendation: ready to merge.

🤖 Generated with Claude Code

Replaces the June 760x428 13fps GIF with an unedited `generate` take of the Pulse dashboard, rendered at 1080p60 on this branch and encoded as a 1200px 30fps animated WebP (6.6MB, about the same weight as the old GIF). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

chatgpt-codex-connector Bot reviewed Sep 30, 2026

View reviewed changes

Co-Messi and others added 9 commits September 30, 2026 20:47

Co-Messi and others added 6 commits October 3, 2026 08:41

Navigation Menu

Output quality: 60fps capture, no ghosting, natural camera and typing - #14

Output quality: 60fps capture, no ghosting, natural camera and typing#14
Co-Messi wants to merge 17 commits into
mainfrom
fix-motion-quality

Co-Messi commented Sep 30, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Sep 30, 2026

Uh oh!

chatgpt-codex-connector Bot Sep 30, 2026

Uh oh!

Co-Messi commented Oct 3, 2026 •

edited

Loading

Uh oh!

Co-Messi commented Oct 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Co-Messi commented Sep 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Items

Metrics

Regression suite: same take files, main's planner vs this branch

Real record + render, demo recipe (record then render --music pulse)

Real generate on the demo app (DeepSeek v4-pro, text-only)

Behaviour changes to be aware of

Not verified

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

Co-Messi commented Oct 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Independent review — progress log (updated as findings land)

Uh oh!

Co-Messi commented Oct 3, 2026

Independent review: summary (head cfb4568)

Fixed on this branch

Verdicts on the PR's claims

Residuals (not fixed)

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Co-Messi commented Sep 30, 2026 •

edited

Loading

Real record + render, demo recipe (`record` then `render --music pulse`)

Real `generate` on the demo app (DeepSeek v4-pro, text-only)

Co-Messi commented Oct 3, 2026 •

edited

Loading

Independent review: summary (head `cfb4568`)