17 October 2026 · Agentic task loop · video edit
Custom Characters Promo · Sushi 1.2.4
The orchestrator reached COMPLETE after 3 h 55 min with no operator action and no failed session, and the final reviewer accepted all 27 requirements on its first cycle. The render passes 19 of 20 hidden acceptance cases: every caption in sync, both exclusions, the full call to action and a still final hold. It hides captions in the pauses between sentences. Visually it is clean but plainer than the Opus 5.5 render of the Spanish brief.
What was checked
- Bootstrap was accepted on its first session: 27 requirements and 13 task cards. T001 (the skeleton) was split twice; one of its parts (T001aa, writing the narration timings) blocked after three attempts and a replan dropped it, its work done by a sibling.
- Each task was accepted only after verify --all and verify --task <ID> passed and a fresh reviewer returned ACCEPT. Reviewers returned CHANGES on T001c, T006, T009 and T012. The final reviewer accepted on the first cycle.
- Independent check after the run: verify --all passed 28/28 in 24 s (limit 240 s). The committed render is 1080×1920, 30 fps, 20.000 s with audio, and was made after the last composition change.
- Hidden acceptance cases: 19 of 20. Failed: captions.always, captions hidden at 2.75–4.0 s and 8.25–8.75 s. The TensorFold attempt hid captions in the same pauses; the brief asks both for captions 'each while it is spoken' and 'always visible', so revision 2 should say which.
- Audio: the music bed is ducked to 12% under each sentence and rises to 35% in the pauses; the chimes sit at 9.8 s and 15.9 s.
- Visual judgement (operator, not a score): clean, readable and free of text collisions, but plainer than the Opus 5.5 render: one dark background for every scene, headlines only in the hook and the call to action, and few finishing details. Opus worked from a storyboard with the hyperframes-creative skill and seven revisions with a person; this run had neither, and the brief asked for restraint.
Limits of this evidence
Two changes from the earlier Sushi runs: Sushi 1.2.4 (faster Flash Next expert kernels, image-aware prefix cache) and the n-gram table. Following the model author's advice, the pack's 4-bit n-gram table (29.8 GiB, sha256 c8ab74bc…) was replaced with the bf16 table of the Sushi-4bpw pack, and config.json's ngram_table was set to 16 bits as the pack's README describes; the pack reports 3–5% lower error for it. Sushi cannot hold the 95.4 GB table in memory and reads rows from the SSD. Within this run the two changes cannot be told apart, and the workload differs from earlier Sushi runs. The rendered video is published with its narration made by Fish Audio Drama 3 (Built with Fish Audio). The footage, generated project, transcripts and engine logs stay on the workstation.
Configuration
The model, serving engine and agent used for this execution.
| Field | Value | Detail |
|---|---|---|
| Model · primary | Qwen 3.8 Flash Next · primary · reasoning medium | EXL3 2.625 bpw (K=2.625; MCG codebook, window 15); n-gram embedding table bf16 (95.4 GB, sha256 387f988e…, read from the SSD) instead of the pack's 4-bit table, as the Sushi startup log confirms |
| Checkpoint · primary | Qwen3.8-Flash-Next-Sushi-2.6bpw, with the bf16 n-gram table of the Sushi-4bpw pack | primary model artifact |
| Runtime | Sushi 1.2.4 · OpenAI-compatible server, greedy sampling (--temp 0), MTP drafts, 8-bit KV cache, SSD prefix cache (emptied before the run), vision on, /metrics enabled mode | inference engine and exact build |
| Harness | Pi coding agent, driven by the task-loop orchestrator, with four HyperFrames skills 1.1.0 | the agent or client that drove the run |
| Hardware | Apple MacBook Pro · 64 GB unified memory | hardware recorded for this execution |
| Workload | Custom Characters Promo · revision 1 | Second run of the Custom Characters Promo brief, with the same brief, asset pack, skills and orchestrator parameters as the TensorFold attempt. A clean start: an earlier attempt on this setup was stopped after its bootstraps kept writing verify helpers where the orchestrator reverts them, the orchestrator was changed to say so, and this run began from a fresh project with an empty Sushi prefix cache. No operator action during the run. |
| Context window | 131,072 tokens | maximum context configured for the serving runtime |
| Repetitions | 1 | executions recorded for this workload; one means an indicative result, not an estimate |
Inference measurements
Engine-side rates for this run. Each figure states its own statistic; figures from different runs are not interchangeable.
| Metric | Value | Definition |
|---|---|---|
| TTFT · median | 1.813 s | Median engine-measured TTFT from Sushi /metrics deltas around each response (p90 5.11 s). · n=798 |
| Decode · output-weighted | 47.22 tok/s | Output tokens / engine decode seconds over requests with engine timings (output-weighted). · n=798 |
| Draft acceptance | 70.19 % | Accepted MTP draft tokens / drafted tokens from Sushi spec-stats over matched requests. · n=798 · 234,678 accepted of 334,368 drafted. |
| Swapped in during the run | 0.76 GiB | Pages swapped in during the run, summed over sessions (vm_stat Swapins deltas), over 520 samples. · n=520 · Memory pressure reached warning at least once and free memory fell to 0.03 GiB; no request failed. Sushi's process reached 51 GiB, against about 46 GiB with the 4-bit table in earlier runs. |
Agent session counters
Agent-wide usage totals. These include reasoning, tool calls and waiting, and are not model decode speed.
| Metric | Value | Definition |
|---|---|---|
| Run elapsed | 234.7 min | From the orchestrator's preflight (2026-10-17 14:23:59 UTC) to COMPLETE (18:18:41 UTC) in one segment, with no block or pause. · n=1 |
| Agent sessions | 55 sessions | Fresh Pi sessions started by the orchestrator: 1 bootstrap, 28 builder, 19 reviewer, 2 split, 4 replan, 1 final reviewer. None stopped or failed. · n=1 |
| Model responses | 798 responses | Assistant responses with usage records across all sessions: 742 tool calls and 56 stops; no engine errors. · n=1 |
| Matched model requests | 798 requests | Pi responses matched in order to the Sushi 1.2.4 process log (PID 31139) by prompt total, output tokens and SSD-restored tokens. · n=1 · Every response matched and no key repeats. Of 830 request summaries in the log, the 5 before the run are the orchestrator's preflight and the 27 after it are later requests from other clients; they stay unassigned. |
| Tool calls | 936 calls | Pi tool calls across all sessions. · n=1 |
| Context compactions | 0 compactions | Pi context compactions across all sessions. · n=1 |
Context and token counters
Token totals as reported by the harness or engine. Cached and cumulative figures are not additive with fresh input.
| Metric | Value | Definition |
|---|---|---|
| Maximum prompt context | 87,063 tokens | Largest engine-reported prompt among matched requests. · n=798 |
| Fresh input | 1,377,700 tokens | Sum of Pi uncached input counters across all sessions. · n=1 |
| Cached input | 19,118,297 tokens | Sushi SSD-restored tokens over matched requests (equal to Pi's cache-read counters). · n=798 |
| Prompt-cache token share | 93.28 % | Cached prompt tokens / prompt tokens over matched requests (19,118,297 / 20,495,997). · n=798 |
| Output tokens | 443,643 tokens | Pi output usage across all sessions; includes reasoning tokens. · n=1 |
| of which reasoning output | 254,710 tokens | Pi-reported reasoning tokens included in output tokens (57.4%). · n=1 |
Recorded check counters
Counts from the checks recorded with this run.
| Metric | Value | Definition |
|---|---|---|
| Tasks completed | 15 tasks | Task cards with status done at the end. Two more were split and one (T001aa) was dropped by a replan after it blocked. · n=1 |
| Reviews returning changes | 4 reviews | Task reviewer sessions that returned CHANGES (T001c, T006, T009, T012). · n=1 |
| Project checks passing | 28 checks | Checks passed by the project's own ./verify --all at the end (24 s in the independent run). · n=1 |
| Hidden acceptance cases passed | 19 checks | Hidden acceptance cases passed, of 20: written from the brief, never shown to the model, and read from the rendered video. · n=1 · No reference render that passes all 20 exists yet. |
Request records & provenance
Per-request numbers are not published. All 798 Pi responses were matched in order to the Sushi 1.2.4 process log; TTFT and decode come from Sushi /metrics read before and after each response.
Run ID: 2026-10-17-qwen38-flash-next-sushi-v124-pi-task-loop-hyperframes-promo
One execution, with checks transcribed from the retained record. A completed session is not a quality score. Measurement method · All runs