# Session Notes Conditions that shaped the output of this session, recorded for anyone comparing runs across models, effort levels, or harnesses. --- ## Model, effort, and harness | | | |---|---| | Model | Opus 5 (`claude-opus-5`) throughout | | Effort at start | **max**, set explicitly via `/effort` in the opening turn | | Effort mid-session | **changed to xhigh** — see below | | Harness | Claude Code (CLI), single agent | | Directory name effort | `max` (matches the starting effort, not the final one) | **Effort changed mid-session.** During the ElevenLabs research phase (between looking up the Text-to-Dialogue endpoint and looking up pricing), the user interrupted with `/effort` and set the level to **xhigh**, saved as the default for new sessions: ``` Set effort level to xhigh (saved as your default for new sessions): Deeper reasoning than high, just below maximum (Fable 5, Opus 4.7+, Sonnet 5) ``` So roughly the first two-thirds of the session (all of the radio-show build, the mixing iteration, and the early ElevenLabs research) ran at **max**, and the final third (the backend implementation, casting, audition, render, and verification) ran at **xhigh**. The directory name says `max` because that was the starting condition. Anyone attributing behaviour differences between the first and second halves should account for this. **No model downgrade, no safeguard fallback, no refusal.** `stop_reason` never came back as `refusal`, no server-side fallback was configured or triggered, and the model identity did not change at any point. **Fast mode** was not used. --- ## Concurrency **One agent, running serially.** No subagents were spawned. The harness instructions say not to spawn agents unless the user asks, and the user never did. Everything was done inline by the single main agent. Two kinds of parallelism *were* used, and both affect wall-clock time: 1. **Parallel tool calls within a turn.** Independent `Bash` / `WebSearch` / `WebFetch` / `TaskCreate` calls were batched into single blocks throughout — e.g. environment probing ran alongside Radiolab research; the mix analysis ran alongside the transcript inspection. 2. **A thread pool inside the generated code.** `ElevenBackend.prewarm()` uses a `ThreadPoolExecutor` with 4 workers to issue TTS requests concurrently. This is application-level concurrency in the artifact, not agent concurrency — but it is what produced the rate-limit event described below. --- ## Memory and persistence **Nothing was written to memory.** The harness provides a file-based memory directory at `~/.claude/projects/-Users-bowiesmith-.../memory/`, and no memories were created, updated, or deleted during the session. No `MEMORY.md` entries were added. **No memories were recalled.** No `` blocks containing recalled memories appeared at any point, so nothing from a prior session influenced this one. Persistence that *did* matter, all of it inside the repo: | What | Where | Effect | |---|---|---| | TTS audio cache | `out/cache/kokoro/`, `out/cache/elevenlabs/` | Re-renders only pay for lines whose text changed. Made the iteration loop viable — and made the second ElevenLabs render cheaper than estimated (16,787 credits vs 18,188 projected) because audition clips were reused. | | Capability probe cache | `out/cache/elevenlabs/_capabilities.json` | Probe runs once per model, then reads from disk. | | Cast file | `cast.elevenlabs.json` (gitignored) | Voice IDs are account-specific. | | API key | `.env` (gitignored) | User pasted it directly; never echoed into the transcript. | The Kokoro cache was **moved** from `out/cache/` to `out/cache/kokoro/` during the backend refactor (1,127 files) rather than regenerated, which is why the byte-identical check still held afterwards. --- ## Task tracking Six tasks were created via `TaskCreate` at the start of the first build (research, project setup, script, TTS pipeline, score/sound design, mix/master) and all were marked completed. No task list was maintained for the ElevenLabs phase. --- ## Tools and skills **Deferred tools loaded via `ToolSearch`:** `WebSearch`, `WebFetch`, `TaskCreate`, `TaskUpdate`. Everything else came from the base tool set. **Skills invoked:** `claude-api` — loaded in the final turn to get accurate Opus 5 pricing rather than answering from memory (the skill's own trigger rules require this for any LLM pricing question). It was not used earlier and did not influence the artifact. **Scratchpad** was used for throwaway QC scripts (`.../scratchpad/qc_buses.py`), kept out of the repo. --- ## External services and network conditions **ElevenLabs** — the only external API with side effects. | Phase | Outcome | |---|---| | First key | Valid but **narrowly scoped**: `user_read`, `voices_read`, `models_read` all denied. Only `/voices/settings/default` worked. | | First plan | **Free tier** — `402 Free users cannot use library voices via the API`. Two hard blockers: no API library voices at all, and 10k credits/month vs ~17.3k needed. Nothing charged (402 rejects pre-generation). | | After upgrade | **Starter tier, 40,000 credits.** `voices_read` and `user_read` now work; `models_read` still denied (irrelevant — model ID is explicit). | | Final spend | **16,787 credits** (audition 762 + probes ~97 + full render). 23,213 remaining. | **Rate limiting during the render.** Running 4 requests concurrently produced repeated `429`s. One line exhausted all six retries inside the concurrent batch and was recovered by the subsequent serial pass. This was verified rather than assumed — all 240 lines confirmed present in cache, none silent or truncated. A lower concurrency setting would trade wall-clock time for fewer retries. **Web fetches that failed** (and therefore shaped what research was available): - `wired.com` — blocked by the harness (`unable to fetch`). The Friston profile was the primary source for the woodlice anecdote; recovered via search + Slate Star Codex instead. - `radiolab.org` transcript — HTTP 500. - HappyScribe Radiolab transcript — the fetcher **declined to reproduce verbatim text on copyright grounds**. This is worth noting: the plan had been to internalise Radiolab's rhythm from a real transcript, and that route was closed. The script's dialogue style therefore comes from craft analysis (Ira Glass's appreciation, Abumrad interviews) plus prior knowledge, not from studying a transcript. - Two ElevenLabs docs pages 404'd or redirected to `buildwithfern.com`. **Docs were unreliable enough to matter.** The public ElevenLabs docs do not state whether `eleven_v3` honours `speed` or works on the with-timestamps endpoint. Rather than guess, the implementation probes the live API once and caches the result. This directly caught a wrong assumption: **v3 rejects `previous_text`/`next_text`** with `400 unsupported_model`, which invalidated the cost strategy adopted one turn earlier. --- ## Determinism **The Kokoro version is byte-reproducible.** The RNG is seeded (1906), and two consecutive builds produced identical MD5s (`3ffb3a8f97ef3fb795a0f484795fc1ef`). This was used as a regression test — re-verified after the entire backend refactor and after every subsequent change, and it held every time. **The ElevenLabs version is not.** `seed` is passed (1906) but ElevenLabs documents it as best-effort only, and v3 is expressive by design. Re-rendering will produce audibly similar but not identical output — and will cost credits again for any line whose cache key changed. --- ## The dominant constraint on quality **The model could not hear its own output.** Every mixing decision — levels, ducking depth, EQ, pacing, dynamic range — was made against measurements rather than by ear. This shaped the work more than any other factor: - It caused **defects to be found that a casual listen might have missed**: a `+1.38 dBTP` true-peak overshoot behind a clean sample peak; a score sitting 21.6 dB under the dialogue; ambience at −65 LUFS; the grief section pacing *faster* than the banter. - It also caused **one wrong conclusion**, stated and then retracted: a claimed logic error in the exciter gating, inferred from a cumulative-energy statistic and contradicted by direct band measurement. - The `show/qc.py` module and the audition mode both exist because of this constraint. The one judgement that remains unverified is the aesthetic one: whether the show actually sounds good, and which of the two voice casts is better. --- ## Corrections made during the session Recorded because they affected user-facing numbers and decisions: 1. **Credit estimate was wrong by 3–5×.** Quoted ~17.3k (the plain character count), which was correct before short-line padding existed and wrong after — the real figure with padding was 55–86k. Caught by modelling the padding budget against the actual short/long line split (180/60), and corrected plainly before any money was spent. 2. **Free-context strategy didn't apply to v3.** Having switched the default to `previous_text`/`next_text` on the grounds that they're unbilled, the probe showed v3 rejects them outright. Corrected, and the docs/comments updated across three files. 3. **Falsely claimed an exciter logic error.** Retracted after direct measurement; no code changed. --- ## Duration and usage figures The numbers in `metadata.toml` are **estimates**. The harness does not expose per-turn timing or token counts to the model, so they are reconstructed from: - Observed tool durations where they were printed (builds at 176 / 216 / 109 / 109 / 171 / 213 / 100 / 264 s; dependency install; analysis scripts). - Turn count (~55 assistant turns) and typical context growth (~12k at the start to ~300k by the end, including the ~70k `claude-api` skill load in the final turn). - Output volume: ~2,900 lines of Python plus documentation, thinking, and visible text. Treat `total-seconds`, `model-seconds`, `tokens-in`, `tokens-out`, and `total-estimated-api-cost` as good-faith approximations, not instrumented measurements. The **pricing rates** are exact (from the `claude-api` skill), and the **prompt-cache TTL for this session was 1 hour**, so cache writes are priced at 2× base input rather than 1.25×.