Duskline

Duskline publishes a daily, provider-authoritative register of model deprecations and deletion dates. The MVP intentionally focuses on explicit lifecycle notices rather than mirroring every active model.

The repository is a Bun/Turborepo workspace built with Effect v4, TanStack Start, TanStack Table, Alchemy v2, and Cloudflare Workers.

Applications

• apps/collector fetches and validates five independent official sources.
• apps/web renders the deprecation register. Its lifecycle column displays Deprecated: … · Deletion: … on one line, ordered by recent deletions newest-first, followed by upcoming deletion dates nearest-first. Model, provider, lifecycle, and replacement are intentionally fixed-order columns. Notices without a deletion date are omitted. The same Cloudflare deployment serves the committed contract as static assets at /current.json, /changes.json, /source-status.json, and /check-status.json.

The complete source snapshot stays behind a TanStack Start server function. The browser receives only display-ready deprecation notices with known deletion dates, plus the small source-verification summary. Models deleted within the previous 30 days appear before upcoming notices; older records, unknown-date notices, and collector-only metadata are not included in the table payload. The /sources page documents each source's authority, coverage, latest check, and retained snapshot. Every retained lifecycle identity also has a prerendered /models/:identity evidence page with dates, official replacements, source links, observation metadata, and Duskline publication history.

Provider and platform identity are deliberately separate. An upstream OpenAI retirement and OpenRouter availability for the same family remain distinct records.

Data contract

data/current.json is the primary frontend artifact. It contains:

• schema_version: currently 1.
• last_published_at: the time of the last semantic publication, not the last successful check.
• records: normalized lifecycle notices with official source URLs, observation metadata, nullable dates, replacement models, and stable hashes.
• source_status: collection health, separate from model lifecycle state.

data/changes.json contains compact field-level additions, removals, status changes, date changes, and replacement changes. It retains the latest 500 events. data/source-status.json is the last published snapshot-health view. data/check-status.json records the outcome and check time of the latest completed daily verification, including runs where no lifecycle fact changed. Model evidence URLs use a deterministic SHA-256 key for the complete provider/platform/model/region identity; display names are never used as joins.

Lifecycle states (legacy, deprecated, retired, removed) describe a model. Snapshot states (healthy, stale, error, invalid, not_collected) describe whether published records can be trusted. Daily check states (healthy, degraded, not_checked) describe the latest verification run. If a fetch, parser, or schema check fails, that source's prior records remain published and its health becomes stale. A failed source never produces an empty replacement.

Development

bun install
bun run test
bun run typecheck
bun run infra:check
bun run collect
bun run dev

Copy .env.example to .env and set CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN. Alchemy reads .env automatically. The local file is ignored by Git. Lifecycle collection uses public official sources and requires no provider credentials.

The live collector uses official sources:

• OpenAI API deprecations
• Anthropic model deprecations
• Amazon Bedrock model lifecycle
• Fireworks changelog Markdown feed
• OpenRouter models API with output_modalities=all

Provider fixture tests characterize each parser and fail closed when expected structure disappears.

Publication and deployment

The daily workflow runs at 00:17 UTC and can also be started manually. Collection and publication are serialized. Tests and schema validation run before a short-lived artifact is handed to a minimally privileged publication job. Only the four allowlisted data files can be committed. Semantic lifecycle files remain byte-identical when facts are unchanged, while check-status.json gets a small daily verification commit. Every successful publication deploys the prerendered site to Cloudflare, so the homepage can distinguish “last checked” from “lifecycle facts last changed.”

bun run deploy first performs an authenticated collection and validates the artifacts, then deploys the TanStack Start site and its static JSON as one Cloudflare Worker:

bun run deploy -- --stage prod

The homepage, source coverage, and model evidence pages are prerendered. Browser caches must revalidate documents, while Cloudflare may keep them at the edge for five minutes with stale-while-revalidate and stale-if-error protection. Fingerprint assets are immutable for one year.

Configure those values as a GitHub Actions variable and secret. On the first deployment, Cloudflare.state() provisions Alchemy's state-store authentication and encryption secrets automatically. The repository should remain private when reliable GitHub schedules are required; if made public, add an external freshness monitor because GitHub can disable schedules in inactive public repositories.

Move collection into Cloudflare Workflows/D1/R2 only when the product adds sub-daily checks, user subscriptions, account-specific Bedrock regions, or raw response retention. Those additions do not require changing the versioned frontend contract.