Same proxy as netlify/functions/ai.mjs, rebuilt as a standalone Bun service on
the CVM host so AI calls stop spending Netlify compute/bandwidth credits and
the sats-paid Routstr key lives on our box.
- ai-proxy/server.ts: Bun.serve on 127.0.0.1:8787 — NIP-98 + Pro gate, model
allowlist, per-IP throttle, /api/ai + /api/ai/status. Own tsconfig
(typecheck:ai in npm test); eslint-ignored like cvm-server.
- ai-proxy/{package.json,.env.example,skatespots-ai.service,Caddyfile,DEPLOY.md}.
- scripts/deploy-ai-proxy.sh (npm run deploy:ai): scp + bun install + unit +
restart, mirroring deploy-cvm.sh.
- Client: one place decides the AI base, src/lib/sasquatchAi.ts (also the NIP-98
u tag). Defaults unchanged (/api/ai on Netlify); set VITE_SASQUATCH_AI_BASE to
move it. Fully reversible.
Deployed and verified on the host: /api/ai/status returns the live Routstr
balance. Public TLS (DNS + Caddy) is the remaining step before switching the app
over, so no client behaviour changes yet.
- sasquatchBrain: PROXY_MODELS now match the Routstr node's slugs
(gemini-3-6-flash, z-ai-glm-5-3, gemini-3-7-flash).
- ai.mjs: new owner-only POST /api/ai/status (NIP-98 + owner pubkey) returning the
upstream balance + a 'low' flag (<500 sats); config.path now serves both routes.
- useSasquatchAiStatus + SasquatchAiStatusCard (owner-only, in Settings): shows the
balance and warns when low — the top-up nudge, no webhook/notification key needed.
- netlify/functions/ai.mjs: optional Pro enforcement (SASQUATCH_REQUIRE_PRO=1).
Requires a NIP-98 kind-27235 auth header (x-sasquatch-auth) and, unless the
caller is a platform owner, an active Pro pass (x-sasquatch-pro = a
provider-authored zap receipt) verified OFFLINE — signature, user binding,
platform tag, and paid-amount → tier. Off by default.
- src/lib/proAuth.ts: signs the NIP-98 once per session (no Amber popup per
message) and attaches the Pro receipt.
- sasquatchBrain: threads auth headers through the proxy (streaming + non).
- useSasquatchPro: exposes the winning receipt event.
- Verified: 200 with the gate off, 401 with it on (env reset afterwards).
- netlify/functions/ai.mjs: POST /api/ai runs against Netlify's AI Gateway —
Netlify injects OPENROUTER_API_KEY + OPENROUTER_BASE_URL into Functions, so no
key is managed or shipped. Curated model allowlist, best-effort per-IP
throttle, CORS for web + Capacitor, and SSE streamed straight through.
- sasquatchBrain: call the proxy instead of OpenRouter directly; drop
VITE_OPENROUTER_API_KEY entirely; PROXY_MODELS set to the slugs that actually
work via the gateway (its ZDR policy blocks gemini/anthropic/gpt-4o-mini).
- netlify.toml: [functions] directory.
- Verified in production: POST https://skate-spots.org/api/ai -> 200.
- sasquatchBrain: new streamSasquatchAI — OpenAI-style SSE via OpenRouter (if
the key is set) then Pollinations; falls back to the non-streaming caller only
when nothing at all was delivered, so blocked-stream environments still answer.
- SasquatchGuide: stream the first pass into a live bubble (raw JSON tool calls
are suppressed so partial JSON never shows; prose-then-tool output is dropped),
finalise the bubble in place, show a blinking cursor, and hide the generic
'thinking...' bubble while a reply streams.
- Drop Pollinations 'mistral' (now a 404) leaving the live 'openai' slug.
- sasquatchBrain: abort each provider request after 15s and give up the whole
multi-provider attempt after 35s, so a hung OpenRouter/Pollinations call can no
longer wedge the chat (it now falls through to the static/local fallbacks).
- SasquatchGuide: send the system prompt as a real system message (both the
first pass and the tool-synthesis pass) instead of smuggling it into a user
turn, and mark tool output as untrusted — blunts prompt injection via event
content.
- useCVMClient: signature-verify (verifyEvent) every relay result, matching the
standalone server and the release-notes claim.
- useCVMServerHealth: stale threshold 1h -> 45m to match the server heartbeat.
Three low-risk speedups focused on the slowest path — questions that hit live
relay data:
1. Relay timeout 10s → 4.5s (6s for the fan-out community stats). The useful
relays answer in ~2s; the old 10s just let one slow/dead relay stall the
whole reply.
2. Results-first: for finder tools (spots/events/coaches/gear/crews/shops/
tours/stats/weather) format the CVM result locally and SKIP the second
"prettify" AI call. Removes a full model round-trip (~2–4s) — users want the
list, not extra prose. New toolCallToIntent() maps a tool call to the existing
formatLocalResults() path. Personal/Pro tools (get_my_*) still use AI synthesis.
3. Faster AI fallback: cap OpenRouter attempts to the top 3 free models before
the keyless backup, and shorten 429 backoff (≤800ms vs ≤2500ms). Avoids the
worst-case 15–20s grind through 8 models on a busy free tier.
Conversational/knowledge answers (single AI call) keep Sasquatch's personality
and are untouched.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
Address the two genuinely exploitable findings from the OpenCode audit:
C1 (critical) — sasquatchBrain.ts shipped a hardcoded OpenRouter API key in
the client bundle, extractable and abusable by anyone. Removed it; OpenRouter
is now opt-in via VITE_OPENROUTER_API_KEY and skipped when unset. Sasquatch
still works for everyone via the existing keyless Pollinations fallback.
(The leaked key must also be rotated/revoked on openrouter.ai.)
H1 (high) — SpotMap popup interpolated external weather-API data directly into
innerHTML. Added an escapeHtml() helper and applied it to every weather/AQI/
forecast value so a compromised weather API can't inject HTML/JS.
Triage of the rest: C3 (crew agent derived key) is a documented, deliberate
tradeoff — agent holds no funds and only speaks for itself. H2 (chart.tsx
dangerouslySetInnerHTML) is the stock shadcn chart with app-controlled config,
not user input. H3/H4 (GitHub Actions) don't apply — no .github/ in this repo.
C2 (NWC in localStorage) is standard for web Nostr clients. npm audit fix +
key rotation are operator steps done outside the app.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
1. Pro purchase confirmation: on a paid pass, send the buyer an encrypted
NIP-17 receipt to their own inbox (tier, expiry, unlocked perks). Best-effort
so it never blocks the purchase. Success toast now points to the DM.
2. Frictionless Pro answers: add local static-FAQ entries for "sasquatch pro",
"pro perks/version", pricing, "go pro", "upgrade" — so those questions are
answered instantly and never dead-end when the free AI model is rate-limited
(the exact failure seen when asking "what perks has the pro version?").
3. FAQ matching now prefers the longest matching key so specific phrases win
over broad ones ("sasquatch pro" > "pro").
4. Refresh the free-model fallback list (add deepseek-v3, gemini-2.0-flash-exp,
qwen-2.5-72b) for better free-tier AI availability.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
Fixes the "brain's catching traffic" dead-end for trick/terminology
questions and reduces reliance on a single flaky free API.
- Local glossary answerer (answerFromGlossary): trick/how-to/terminology
questions ("what is a feeble grind", "how do I do a kickflip", "grosso
grind") are now answered instantly from the 150+ entry glossary with
ZERO API calls — so they never hit a rate limit.
- Wired into SasquatchGuide as (a) a fast path before the AI and (b) a
last-resort fallback when the AI is unreachable.
- sasquatchBrain now has multi-provider fallback: OpenRouter free models
(with 429 backoff) → keyless Pollinations backup. Only throws if both
fail, letting the caller fall back to local knowledge.
- Replaced the unhelpful fallback message with useful local links +
a nudge to ask about a specific trick.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
Glossary:
- New /glossary page: 80+ tricks across 11 categories (fundamentals,
flips, shoves/spins, grinds, slides, grabs, airs, lip, stalls,
handplants, boneless, freestyle) with alternate names, difficulty
badges, live search, and category filters.
- Added to desktop "More" dropdown and mobile burger menu.
Sasquatch AI:
- Fixed/expanded free-model fallback list with valid OpenRouter slugs
(dropped the non-existent gemma-4 slug).
- Added exponential backoff + one retry on 429 rate limits, respecting
Retry-After headers.
- SasquatchGuide now reuses the shared sasquatchBrain caller instead of
a duplicated implementation.
- Friendlier fallback message when the AI is saturated.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
Phase 0 — identity (lib/crewAgent.ts):
- Deterministic per-crew agent keypair derived from the crew coordinate
(sha256 of a public salt + coord). Documented security tradeoff: the key
is reproducible by design and holds no funds — used only for the agent to
speak for itself. Includes kind-0 profile builder (bot:true, NIP-24) and
a signAsCrewAgent helper for Phase 3.
Phase 1 — opt-in toggle:
- SkateCrew.agentEnabled parsed from a queryable ["t","crew-agent"] marker.
- "Crew Agent" toggle (Beta) in CrewForm; publishes the marker when on.
- NIP.md documents the crew-agent tag and the agent identity derivation.
Phase 2 — crew-scoped Sasquatch with receipts:
- lib/sasquatchBrain.ts: shared AI caller (model fallback) + tool-call parser.
- useCrewAgentChat: reuses the CVM data tools, seeded with crew context
(name/location) so "near us" resolves to the crew's turf. Ephemeral/local,
no keys, no publishing yet.
- CrewAgentPanel: chat dialog rendering agent answers PLUS structured receipt
cards (spots / events / Lab pros / gear / crews) linking into the app.
- CrewDetailPage: "Ask Sasquatch" card + panel, shown only when the crew
enabled the agent.
Also: Lab category tabs → fixed 5-across grid (no horizontal scroll). Bump 1.0.68.
Co-authored-by: shakespeare.diy <assistant@shakespeare.diy>
