dewpt

An ambient ideation canvas where words condense out of latent space. See SPEC.md for the concept and vocabulary; dewpt-demo.html is the static reference implementation whose look and interaction behavior this build preserves.

This is M1 (live field): seed input, Worker + per-session Durable Object, pool generation via Workers AI, drip client, evaporated sidebar.

Run

npm install
npm run dev        # wrangler dev on http://localhost:8787

The AI binding runs remotely ("remote": true), so wrangler login must have happened once on this machine. Probe it with:

curl http://localhost:8787/api/debug/ai   # {"ok":true,...} means generation will work

WARP users: workerd (the wrangler dev runtime) cannot open outbound connections while Cloudflare WARP intercepts its sockets — the AI binding fails with InferenceUpstreamError: Network connection lost and the probe above hangs. Pause WARP (or exclude workerd in split-tunnel settings) while developing. As a last resort, echo 'DEV_FAKE_AI=1' > .dev.vars swaps in a canned offline generator (src/dev-fake-ai.ts) so the field machinery can be exercised without egress. Production generation is Workers AI only.

Local inference

Generation and embedding can run against any OpenAI-compatible server — Ollama, llama.cpp's server, LM Studio, vLLM — by pointing .dev.vars at it:

LOCAL_AI_BASE_URL=http://localhost:11434/v1
GEN_MODEL=qwen2.5-coder:7b
EMBED_MODEL=bge-m3

/api/debug/ai reports which backend answered ("mode":"local"), so the probe stays honest once you switch. LOCAL_AI_API_KEY is there for servers that want one; Ollama ignores it. This is a dev convenience — production is Workers AI, and these vars live only in .dev.vars, never in wrangler.jsonc.

Two things that bite:

• Thinking models return nothing. They spend the whole max_tokens budget on message.reasoning and hand back an empty content, which parses to zero candidates and a field that never fills. Add LOCAL_AI_CHAT_OPTIONS={"reasoning_effort":"none"} — measured on qwen3.5:4b, that turns 0 candidates into 8. Ollama silently ignores both think: false and chat_template_kwargs.enable_thinking; only reasoning_effort works there.
• Embeddings are persisted per session. Vectors live in the DO's SQLite, so changing EMBED_MODEL mid-session mixes incompatible dimensions into the cosine scoring. Start a new session after a model change.

WARP blocks workerd's egress to localhost too, so this is not a way around that trap — it is a way to stop needing a Cloudflare account and an Access token to develop.

Tests & checks

npm test           # vitest: pool logic + generation prompt/parsing (AI mocked)
npm run typecheck  # tsc over src/test and scripts

Prompt calibration

Iterate on generation quality without touching the app — edit the prompt and few-shots in generation.ts, then eyeball three labeled batches (strangeness 0.2 / 0.5 / 0.85, altitude 0.3):

CLOUDFLARE_ACCOUNT_ID=… CLOUDFLARE_API_TOKEN=… \
  npm run calibrate -- "security awareness people actually enjoy"
# options: --model=@cf/… --count=24

The token needs the Workers AI — Read permission. Calibrate goes through the Cloudflare REST API, so it works even where workerd egress doesn't (WARP).

Architecture

• Worker (src/index.ts) — /api routes; static client served from public/ via Workers assets (run_worker_first).
• SessionDO (src/session-do.ts) — one per session (the URL hash is the session). SQLite persistence, alarm-driven generation pump. Serving never waits on generation.
• PoolCore (src/pool-core.ts) — pure pool logic: 6 buckets (3 dewpoint tiers × 2 altitudes, mirroring the demo's pools), fresh-first draws, lazy invalidation on pins/param changes, embedding cosine dedupe (> 0.92), exclude LRU (~300), evaporated ring buffer (20).
• Generation (src/generation.ts) — prompt building (the model hears "strangeness", never the weather vocabulary), few-shot examples per strangeness band seeded from the demo pools, robust JSON parsing, embeddings via @cf/baai/bge-m3.
• Client (public/field.js) — the demo's field machinery verbatim (spawn/decay timing, depth, tier colors, pulse-on-prospect, pin-to-condense, reduced-motion), fed by per-bucket local buffers (public/pool-client.js) that refill in the background. An empty buffer skips a spawn tick; it never blocks.
• Teaching surface (docs/ui-teaching-research.md) — slider endpoint labels + screen-reader glosses, a progressive hint line that advances as the user acts (pure state machine in src/hint-machine.ts, mirrored for the browser in public/hint-machine.js; a localStorage taught flag quiets it for returning users), and a non-modal ? legend for recall.
• Explainer surface (docs/explainer-research.md) — before a seed exists the field explains itself by demonstration: a centered manifesto plus meta-words about dewpt condensing and evaporating through the real spawn/decay motion (public/preseed.js, static pool in public/preseed-pool.js; zero network, torn down the moment a session starts). A ? what is this? toggle in the header opens a non-modal about panel — concept, weather glossary, and the Kevin Kelly credit — that never pauses the field.