Bot replies were running long (4 paragraphs of action+dialogue beats
per turn) because we never set max_tokens on the narrative call. Three
tunable knobs now in Settings (set in data/config.toml to override):
- narrative_max_tokens: int = 400
Hard cap on each generated response. ~400 tokens ≈ 1–2 short
paragraphs. Drop to 200 for terse banter, bump to 800+ for longer
scenes.
- narrative_temperature: float = 0.85
Sampling temperature. 0.7 = grounded/consistent (slightly stiff),
0.85 = creative-but-in-character (default), 1.0 = wide variety,
>1.0 = often off-the-rails.
- prompt closing instruction now nudges: "Keep your response to a
single beat — one or two short paragraphs at most. Don't monologue;
leave room for the other person to react."
Both turns.py (post_turn) and regenerate.py forward the params to
client.stream(). FeatherlessClient already passes **params through to
the OpenAI-compat endpoint.
Note: temperature doesn't control length — that was a common
misconception. max_tokens is the actual length cap. Lower temperature
makes word choice more predictable (slightly stiffer voice), not
shorter. Both knobs are useful for different goals.
The kickoff parse-and-confirm route was 500-ing intermittently because
Hermes-3 + Featherless's response_format={"type":"json_object"} only
guarantees JSON output, NOT a particular schema. The model was inventing
its own field names (sceneTime, entities, settingDetails) instead of
the KickoffParse fields, causing Pydantic validation to fail on both
classify() retries.
Three changes:
1. Include the Pydantic JSON schema in the system prompt so the model
knows exactly which keys to produce. Affects every classify() call
(kickoff parse, turn parse, scene-close detect, significance,
state-update, scene summarize). Strip ```json fences if the model
wraps its output. Bump retries 2 → 3 (model is stochastic; one extra
attempt closes most of the remaining gap).
2. parse_kickoff() now passes a default empty KickoffParse so the
route degrades to a fillable form instead of 500 when the classifier
ultimately fails. The confirm form is the human-in-the-loop; an
empty form is strictly better UX than a stack trace.
3. Tests updated: bumped canned-failure arrays from 2 → 3 entries to
match the new attempt count; renamed kickoff test from
"raises_when_classifier_fails_twice" to
"falls_back_to_empty_when_classifier_fails" reflecting the new
degraded-but-usable behavior.
Verified live with all 3 sample bots (maya/eli/sam) — kickoff route
returns 200 across multiple attempts. Full suite: 168 passed.
Joseph Doherty