Journal generation

When the 15th layer lands, the bridge calls Kimi K2 once, asynchronously with the 8-second master fade, to produce the journal. Code lives in proxy/src/kimi.ts.

Model choice

moonshot-v1-128k — explicitly the non-reasoning chat variant. Two reasons:

1. Reasoning models burn most of the token budget on hidden chain-of-thought before emitting any content. With a tight prose target (≤480 chars) we hit empty-content responses repeatedly on reasoning variants.
2. The reasoning scratchpad occasionally leaked into the visible journal, breaking the second-person geological voice.

The 128k context window is overkill for a 15-line transcript but means we never have to truncate or summarise the placement log going in.

Input format

The full transcript is formatted as one numbered placement per line, with the comment in quotes when present:

01. Hermes placed drone — "first tone, the cave breathes"
02. Player placed bell
03. Hermes placed texture — "dust between the marks"
04. Player placed pulse
...
15. Hermes placed swell — "lift, then dissolve"

That transcript plus a short scale header (key + mode feel) is the entire prompt context.

Format rules

The system prompt enforces:

• Exactly 3 paragraphs, 2–3 sentences each
• Under 480 characters total (we cap at 720 in clamp logic because moonshot tends to overshoot the advisory)
• Tone: introspective, slightly mineral — “a geologist taking notes inside a cave”
• Reference 1–2 striking moves from the log
• No title, no markdown, no preamble — output only the prose

Clamp logic

If the response overshoots the cap, clampJournal() truncates in preference order:

1. Cut at the last paragraph break that fits under the cap.
2. If a single paragraph is too long, cut at the last sentence end (., !, ?) that fits.
3. Last resort, cut at the last whitespace that fits.

It never chops mid-sentence. If the model returns something completely out-of-format (no paragraph breaks, no punctuation), the descent fallback kicks in (see below).

Fallback

If KIMI_API_KEY is unset or the call fails, the bridge produces a deterministic stub journal — a one-line acknowledgement that the descent completed, with the layer count and key. The descent always ends gracefully; mints never block on Kimi availability.

Why one call, not three

The journal is single-shot because variance there is small. We’ve never observed a “bad” journal from moonshot in the way we observe bad glyphs — the model is reliably introspective in the requested voice. Glyphs need best-of-3 with structural scoring; journals don’t.