streetlight：面向西雅图住房困难人群的免费匿名 AI 支持工具

Free, no-account AI support for people navigating housing instability, paperwork, benefits, and hard decisions in Seattle / King County.

Anonymous by default. No accounts, no PII retention, no ads, no upsell, no data harvesting, no premium tier. Open source from day one.

A software engineer in SOMA uses Claude every day. He pays $20/month and it's a rounding error in his budget. A guy sleeping in a shelter ten blocks away has the exact same brain. Often harder problems. Zero access. That gap is not a moral failure on his part. It's an access gap. This project closes it, in one city, for one population, as a public utility.

Status

Pre-launch. V1 in development. Soft-launch through named frontline partners when the build is ready.

Documents

Agent instructions — shared coding-agent entry point for Claude Code, Codex, and other agents.
Thesis — why this tool exists and the reasoning behind it.
V1 Specification — what the tool is, concretely.
Data and Privacy Architecture — how it's built, what it logs (and doesn't), and the threat model.
What this tool will not do — the "no" list.
Translation handoff — how locale files are produced, reviewed, and audited for completeness.
Translation worklist — generated per-language missing-key checklist for translation maintenance.
Resource maintenance — how the hand-curated human-help lists are verified and kept fresh.
Operational runbook — dashboard-first incident and pause procedures.
Partner bug report template — structured no-content issue intake for frontline partners.
Partner launch packet — ready-to-send framing, limits, and bug-report links for soft launch.
Partner materials index — quick map of the soft-launch docs set.
Security policy — how to report vulnerabilities safely.
Contributing guide — how to open issues and send pull requests.
Code of conduct — participation and moderation standards.

Project structure

This is a one-person, one-year commitment with a graceful sunset path if no institutional home is found by year-end. The about page is honest about this. Code stays public on GitHub regardless.

Useful commands

npm run lint
npm run dev:mock — runs the app with zero-cost local mock chat replies
npm run dev:live — runs the app against the configured live model path
npm run build
npm run check:audit — runs npm audit for dependency advisories
npm run check:locales
npm run check:locales:summary — quieter locale-gap summary for routine verification loops
npm run check:launch — static soft-launch readiness check for docs, placeholders, translations, and env contract
npm run check:env — validates local/env-example boolean and numeric value shapes without printing secrets
npm run check:forbidden-integrations — fails if forbidden analytics, error-reporting, or log-forwarding packages are added
npm run check:no-tracked-secrets — scans tracked files for common secret-shaped values before push
npm run check:ops — dry-run validation for the incident and ops helper scripts
npm run check:response-style — runs live brief-first response-shape cases against the configured /api/chat path
npm run check:translation-worklist — verifies that the generated translation worklist is current
npm run cost:status — shows the local model pair and whether cheap local mode is active
npm run diagnostics:local — prints the current local server mode and the report-page metadata snapshot
npm run ops:status — prints the local health, deploy metadata, model mode, and any launch-doc placeholders still hanging around
npm run build:translation-worklist
npm run incident:new -- --slug short-name --severity Sev-2 — scaffolds a public Sev-1/Sev-2 incident write-up from incidents/TEMPLATE.md
npm run incident:sev3 -- --what "..." --action "..." --outcome "..." — appends a structured Sev-3 note to incidents/log.md
add --dry-run true to either incident command if you want a preview without writing files
npm run validate:data — validates referral/crisis schema plus source and last-verified metadata
npm run data:status — shows the oldest referral/crisis entries and whether any have drifted past the freshness threshold
npm run smoke — checks key pages plus a small synthetic end-to-end chat flow against ACCESS_TOOL_BASE_URL or http://localhost:3000
npm run smoke:quick — cheaper local smoke pass with a smaller synthetic sample
npm run regression:prompts — runs the synthetic prompt suite in tests/prompts/ against the configured live model path
npm run regression:mock — runs the full prompt set in mock-local plumbing mode without live model calls; this is what PR verification uses
npm run regression:quick — cheaper prompt plumbing check; in mock mode it verifies streaming/classifier wiring without pretending to validate live model behavior
npm run verify:quick — lower-cost local check for active setup work, now including the diagnostics snapshot and quick regression plumbing pass
npm run verify:mock — full local no-spend verification with all smoke cases and all prompt fixtures in mock plumbing mode
npm run verify — runs the local pre-launch verification stack

npm run dev:mock is for zero-cost local UI work. npm run dev:live is the easy way back to real model calls. Full prompt regression still expects live model replies, so npm run regression:prompts will stop with a clear message while mock mode is active. Use npm run regression:mock for a full no-spend plumbing pass. The GitHub Actions Verify workflow runs the full mock suite on PRs and pushes to main; run npm run regression:prompts with the Haiku testing model path before model, prompt, or deploy changes where live behavior matters. Run npm run check:response-style after prompt changes that affect answer length or structure.

Utility routes

/healthz — simple no-store JSON health check for local smoke checks and future monitoring
/report-problem — structured no-content bug report draft builder that stays client-side and opens a mailto: draft

License

Licensed under the Apache License, Version 2.0. See LICENSE.

Contact

Bug reports currently go to jesse@balancedlivingtherapy.com.