# Session notes Conditions that shaped the output of this session, recorded for anyone comparing runs across models, effort levels, or harnesses. Every number here was read from the session transcript (`~/.claude/projects/…/548ff887-….jsonl`) and the subagent journals, not estimated. --- ## Model and effort | | | | --- | --- | | Model | `claude-opus-5` for **every** request — main loop and all subagents | | Effort | `xhigh` on all 339 main-loop assistant messages | | Service tier | `standard` on all requests | | Speed | `standard` — fast mode was **not** used | | `inference_geo` | `not_available` (no US-pinned routing, no 1.1× multiplier) | | Harness | Claude Code `2.1.243`, permission mode `auto`, mode `normal` | **No downgrade occurred and no restoration was needed.** Every one of the 339 real assistant messages carries `model: claude-opus-5`. Five additional `` assistant records appear in the transcript — these are harness-generated notices (usage-limit messages and one request timeout), not model output, and they carry no usage. Effort was set by the user's opening `/effort` command: ``` Set effort level to ultracode (this session only): xhigh + dynamic workflow orchestration ``` Ultracode is the material variable here. It carries a standing instruction to author and run a `Workflow` for every substantive task and to treat token cost as a non-constraint. Three workflows were run (46 distinct agent invocations) where a non-ultracode session would very likely have used one or two subagents, or none. The verification pass in §"Adversarial verification" below is directly attributable to it. --- ## Interruptions that affected the run ### Two usage-limit lockouts The session was cut off mid-workflow twice. | # | Hit at | Reset | Idle | What was lost | | --- | --- | --- | --- | --- | | 1 | ~00:44 EDT | 05:10 EDT | 4 h 27 m | 16 of 23 agents in workflow #1: all 9 judges, both page drafts, both syntheses, all 3 verifiers | | 2 | ~05:36 EDT | 10:10 EDT | 4 h 35 m | 4 of 23 agents in the resumed run: `synthesize:page` and all 3 verifiers | **This is the single biggest structural influence on the artifacts**, and it changed the shape of the work in two ways: 1. **Recovery used workflow resume rather than re-running.** After the first lockout the journal was inspected to confirm the seven completed agents had cached, non-empty results, then the run was relaunched with `{scriptPath, resumeFromRunId}`. The unchanged prefix replayed from cache instantly and only the failed agents re-ran. Without this the four grounding briefs (~194 KB of research) would have been regenerated at full cost, twice. 2. **`synthesize:page` never ran, and the essay prose is therefore mine.** The page-content synthesis agent died in both lockouts. Rather than spend a third workflow on it, the essay was written directly from the same four research briefs the agent would have used. Only the *narration script* is a workflow product (drafted 3×, judged 9×, synthesized at `xhigh`). A run without the lockouts would have had two synthesized artifacts instead of one — worth knowing when comparing prose quality across sessions. 3. **The verification workflow was re-scoped as a result.** Because the original `verify:*` agents were written to check *drafts* and kept dying anyway, a fresh workflow was authored to check the **actual shipped files on disk**. That is strictly better — it caught 15 real defects in the files as written, including several introduced during hand-editing that no draft-checker could have seen. ### One request timeout The user first sent this meta-task (script.md / notes.md / metadata.toml) on **2026-08-28 14:18 EDT**. It failed with a synthetic `Request timed out` at 14:27 and produced no output. The prompt was re-sent on **2026-08-29 16:48 EDT** and completed. Both copies of the prompt are present in the conversation, which is why the instruction text appears twice in the final turn. The 26.4-hour gap between them is idle wall-clock and dominates the raw session span. --- ## Concurrency Work was **not** serial. Three levels of parallelism ran, often at the same time: | Level | Detail | | --- | --- | | **Workflows in background** | All three `Workflow` runs were backgrounded. Scaffolding, the design system, the SHA-256 implementation and its fuzzing, and the scene renderer were all written while agents ran. | | **Fan-out inside workflows** | Grounding = 4 parallel; drafting = 3 parallel; judging = 3 drafts × 3 lenses concurrently; verification = 4 checks pipelined into per-finding adjudication with no barrier between stages. | | **Background Bash** | The two full video renders (≈12 min each) ran via `run_in_background`, with README writing, mobile screenshots and linting proceeding meanwhile. | Agent counts: 23 + 23 + 27 = **73 agent slots**, of which 46 executed live (the rest were cache replays or lockout failures). Subagents produced 796 assistant messages against the main loop's 339 — roughly 70 % of all model messages in this session were subagent messages. **Notable:** the resumed workflow's cached prefix means the two `wf_302f6e13-1e2` rows are the same run ID, not two runs. Comparing "agents spawned" across sessions should count 46 live agents, not 73 slots. --- ## Caching and persistence - **Every cache write used the 1-hour TTL.** `cache_creation.ephemeral_1h_input_tokens` = 8,913,854; `ephemeral_5m_input_tokens` = 0 across both the main loop and all subagents. At 2× base input, 1h writes cost $89.14 of the $176.31 total — the largest single line item. The trade paid off: the cache-read share was **92.8 %** of all input tokens, and the same work without caching would have cost **$649.71** instead of **$176.31** (a 3.7× saving). - **Session-level context compaction was not triggered.** The session ran to ~15 M tokens of budget with no compaction event in the transcript. - **Memory directory: created, never written.** The harness provisioned `~/.claude/projects/…/memory/` at 00:21. Nothing was saved to it and no `MEMORY.md` exists. The one genuinely non-obvious discovery of the session (the espeak path-length bug) was documented in the code that works around it — `video/src/bigworld_video/espeak_shim.py` — which the memory guidance prefers over a memory file, since the repo already records it. - **Scratchpad used for all intermediates.** Research briefs (~194 KB), contact sheets and probe screenshots went to the session scratchpad, not the repo. The repo contains only shippable files. - **Workflow journals are the durable artifact.** `journal.jsonl` per run is what made resume-from- cache possible after both lockouts. --- ## Adversarial verification — and what it caught The final workflow was structured so that every finding faced a referee whose default was that the finding is wrong ("plausible-sounding findings that do not survive checking are worse than no findings at all"). Result: **16 raised, 18 adjudications, 15 confirmed, 3 rejected.** The rejections mattered as much as the confirmations. One flagged a "parse hazard" in the Dyson-shell line; the referee showed the grammatical rationale was backwards and the line was left alone. Another claimed the closing line landed weakly; the referee showed the alleged mishearing required two simultaneous errors. A one-sided review would have degraded the script on both counts. The confirmations were substantive and would have shipped otherwise — a magnitude wrong by 2× inside a sentence the page's own instrument contradicts; AES aged at 40 years instead of 25; two post-quantum breaks conflated; "four words" attached to a five-word phrase. **Three defects were found by the model's own checking rather than by any agent**, and are worth noting because they are the kind a review pass usually misses: 1. **A real SHA-256 bug.** Fuzzing the hand-written implementation against Node's `crypto` revealed a padding error at every length ≡ 55 (mod 64). Found only because 1,100 vectors were run instead of the usual three. 2. **The espeak 160-byte path buffer.** Five hypotheses were tested before measuring the path length. A machine with a shallower checkout would never have hit it. 3. **The atom-computer cliché is false.** The `crypto-numbers` brief flagged it, and the futility instrument computes it live — so the page's own tool refutes the line the page quotes. This became a design feature rather than something to hide. --- ## Environment specifics that could change the output elsewhere - **macOS `say` was rejected** in favour of Kokoro-82M. On a machine without network access at first run, Kokoro cannot fetch its checkpoint and the video pipeline would fall back or fail. This is the one dependency `uv.lock` cannot pin; it is called out in `README.md`. - **The espeak workaround is path-length dependent.** On a shallower checkout the shim is a no-op (`len(parent) <= 110` short-circuits it). The bug is invisible on most machines. - **Frame render throughput varied 5×** — ~16 fps on simple scenes, ~3.5 fps once the `lattice` and `onevsall` beats (O(n²) over 209 nodes) came in. Full render: ~12 minutes for 1,508 frames. - **Playwright pinned Chromium** (`chromium-1187`, headless shell 140.0.7339.16). Font resolution in the video comes from macOS system fonts; a Linux render would substitute and shift metrics. - **Two live pricing fetches** were made while writing `metadata.toml`, via the `claude-api` skill and `WebFetch`, rather than pricing from memory. One URL 404'd; the second and third resolved and agreed with the skill's bundled table. --- ## Duration accounting The raw span is misleading, so here is the decomposition. | Measure | Seconds | Note | | --- | --- | --- | | **Full span** (first → last message) | 145,770 | 40.5 h, 2026-08-28 00:20:47 → 2026-08-29 16:50:16 EDT | | Usage-limit lockout #1 | 16,011 | 4.45 h, nothing running | | Usage-limit lockout #2 | 16,523 | 4.59 h, nothing running | | Gap before the re-sent meta-task | 94,885 | 26.4 h, user away; includes the timed-out attempt | | Other idle > 240 s | 14,025 | mostly background renders and workflow waits | | **Active session time** | **4,326** | 1.20 h of turns actually in flight | | **Model processing time** | **4,043** | 1.12 h; sum of main-loop response latencies | Model time is 93 % of active time — the session was model-bound, not tool-bound, which is what `xhigh` on a 1 M-token context buys. Subagent model time is *not* included in `model-seconds`: it ran concurrently inside the three workflows (56 m 53 s of combined workflow wall time) and would double-count against the main loop's own clock. --- ## What a comparison run should hold constant If this artifact is being compared against another model, effort level, or harness: 1. **The lockouts are not reproducible.** They cost one synthesized artifact (`synthesize:page`) and forced the essay prose to be written directly. A clean run gets more agent-authored prose. 2. **The research briefs are cached in the journals**, not in the repo. A fresh run regenerates them and will not get byte-identical grounding. 3. **The narration script is a judged-panel product**; the essay prose is not. Do not read the two as products of the same process. 4. **All quantitative claims were verified by computation in-session** (see `script.md` Appendix C). A run that skips that step will produce prose that reads the same and is wrong in at least a dozen places — this run's verification pass found exactly that many.