@bytesbrains/cruise-worker 0.0.0-stage → 0.12.0

package/migrations/0122_batch_prices.sql

@@ -0,0 +1,188 @@
+-- A price for work that may wait: the provider batch tier (#716, epic #712).
+--
+-- A batch is billed at a different rate from the same model called
+-- synchronously — OpenAI's is half, on every model in this file. That is a
+-- third condition on a price, beside the two 0024 added (prompt-size tier and
+-- peak window), and it is unlike them in one way that decides the shape: **a
+-- synchronous request must never be priced at it.** A peak row is chosen by
+-- the clock and a tier row by the prompt, so the resolver picks between them
+-- per request. The batch row is chosen by *which surface the request came in
+-- on*, which the resolver has no way to see and must not be asked to guess.
+--
+-- ## So the condition is a column, and the existing view filters it out
+--
+-- `price_observations` gains `service_tier`: '' is the rate everything has
+-- been priced at until now, 'batch' the provider's batch rate. It joins the
+-- primary key, because the same model, date, tier and window are observed at
+-- both rates, often in the same pass.
+--
+-- `model_prices` keeps its name and its columns and now reads **only** the ''
+-- tier. Every reader of it today — the resolver, lane prices, the catalogue,
+-- doctor, reports, the demo parity check — goes on seeing exactly the rows it
+-- saw before this migration, and none of them can see a batch rate by
+-- accident. The alternative, a `service_tier` column on the view that every
+-- reader then filters, is one forgotten WHERE from billing a synchronous
+-- request at half price, and a lane's price would read the minimum through a
+-- maximum (CLAUDE.md: a lane's price is the maximum) the first time a reader
+-- forgot.
+--
+-- `model_batch_prices` is the same view over the 'batch' tier, and is read by
+-- the batch path alone (#716).
+--
+-- ## The rebuild
+--
+-- The primary key changes, so the table is rebuilt — the 0024 shape. Every
+-- observation is carried with `service_tier` '', which is what every one of
+-- them was. Nothing references `price_observations` but the view, which is
+-- dropped first and recreated after. The demo's seed zeroes every observation
+-- (`src/demo-catalogue.ts`), so the batch rows it inherits from a migration
+-- are zeroed with the rest.
+
+DROP VIEW model_prices;
+
+CREATE TABLE price_observations_new (
+  observed_at                  TEXT    NOT NULL DEFAULT (datetime('now')),
+  model                        TEXT    NOT NULL,
+  effective_from               TEXT    NOT NULL,
+  prompt_tokens_from           INTEGER NOT NULL DEFAULT 0,
+  peak_utc                     TEXT    NOT NULL DEFAULT '',
+  service_tier                 TEXT    NOT NULL DEFAULT '' CHECK (service_tier IN ('', 'batch')),
+  input_micros_per_mtok        INTEGER NOT NULL,
+  output_micros_per_mtok       INTEGER NOT NULL,
+  cached_input_micros_per_mtok INTEGER,
+  source                       TEXT    NOT NULL,
+  note                         TEXT,
+  retracted                    INTEGER NOT NULL DEFAULT 0 CHECK (retracted IN (0, 1)),
+  PRIMARY KEY (model, effective_from, prompt_tokens_from, peak_utc, service_tier, observed_at)
+);
+
+INSERT INTO price_observations_new
+  (observed_at, model, effective_from, prompt_tokens_from, peak_utc,
+   input_micros_per_mtok, output_micros_per_mtok, cached_input_micros_per_mtok, source, note, retracted)
+SELECT observed_at, model, effective_from, prompt_tokens_from, peak_utc,
+       input_micros_per_mtok, output_micros_per_mtok, cached_input_micros_per_mtok, source, note, retracted
+  FROM price_observations;
+
+DROP TABLE price_observations;
+ALTER TABLE price_observations_new RENAME TO price_observations;
+
+CREATE INDEX idx_price_observations_lookup ON price_observations (model, effective_from);
+
+-- Unchanged from 0048 but for the tier: the latest surviving observation per
+-- price key, standard rate only.
+CREATE VIEW model_prices AS
+SELECT o.model,
+       o.effective_from,
+       o.prompt_tokens_from,
+       o.peak_utc,
+       o.input_micros_per_mtok,
+       o.output_micros_per_mtok,
+       o.cached_input_micros_per_mtok,
+       o.source,
+       o.observed_at AS recorded_at
+  FROM price_observations o
+ WHERE o.retracted = 0
+   AND o.service_tier = ''
+   AND o.observed_at = (
+         SELECT max(x.observed_at)
+           FROM price_observations x
+          WHERE x.model              = o.model
+            AND x.effective_from     = o.effective_from
+            AND x.prompt_tokens_from = o.prompt_tokens_from
+            AND x.peak_utc           = o.peak_utc
+            AND x.service_tier       = ''
+       );
+
+-- The same, for the batch rate. Same columns, so the resolver reads either.
+CREATE VIEW model_batch_prices AS
+SELECT o.model,
+       o.effective_from,
+       o.prompt_tokens_from,
+       o.peak_utc,
+       o.input_micros_per_mtok,
+       o.output_micros_per_mtok,
+       o.cached_input_micros_per_mtok,
+       o.source,
+       o.observed_at AS recorded_at
+  FROM price_observations o
+ WHERE o.retracted = 0
+   AND o.service_tier = 'batch'
+   AND o.observed_at = (
+         SELECT max(x.observed_at)
+           FROM price_observations x
+          WHERE x.model              = o.model
+            AND x.effective_from     = o.effective_from
+            AND x.prompt_tokens_from = o.prompt_tokens_from
+            AND x.peak_utc           = o.peak_utc
+            AND x.service_tier       = 'batch'
+       );
+
+-- ## OpenAI's batch rates
+--
+-- Fetched 2026-10-29 UTC from the host's own pricing page, Batch tab,
+-- https://developers.openai.com/api/docs/pricing — not the board. The Batch
+-- guide (https://developers.openai.com/api/docs/guides/batch) states the
+-- discount as "50% cost discount compared to synchronous APIs" and a 24h
+-- completion window; every figure below is the listed batch rate, and each is
+-- half the standard row this catalogue already holds, which was checked
+-- against the same page's Standard tab on the same date. No standard row
+-- needed correcting.
+--
+-- The long-context tier applies to batch too: >272K input tokens, the same
+-- `prompt_tokens_from` = 272001 boundary 0049 reads.
+--
+-- `effective_from` is each model's own first standard date, because the
+-- batch rate was not introduced after the model — the page lists both, and a
+-- model's batch rate starting later than its standard one would be a claim
+-- nobody observed. It is also what lets a batch resolve on any day the model
+-- has a standard price.
+--
+-- gpt-5.6-sol's standard $4 / $20 is promotional, "available at least through
+-- November 21, 2026" (same page); its batch rate is half the promotional one.
+-- When the promotion ends both tiers move, and both are observed then.
+--
+-- **Not written: embeddings.** text-embedding-3-small and -large are listed
+-- as batch-capable on their model pages, and the pricing page's Batch table
+-- has no row for either. A batch rate that is not stated is not one to infer,
+-- so embeddings are not batchable on Cruise until one is: a model with no
+-- batch row is refused at submit (#716).
+INSERT INTO price_observations
+  (observed_at, model, effective_from, prompt_tokens_from, peak_utc, service_tier,
+   input_micros_per_mtok, output_micros_per_mtok, cached_input_micros_per_mtok, source, note)
+VALUES
+  ('2026-10-29 00:00:00', 'openai/gpt-6-astra',    '2026-09-13',      0, '', 'batch',
+    5000000, 25000000,  500000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122).'),
+  ('2026-10-29 00:00:00', 'openai/gpt-6-astra',    '2026-09-13', 272001, '', 'batch',
+   10000000, 37500000, 1000000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-6.1-sol',    '2026-10-02',      0, '', 'batch',
+    1000000,  5000000,   50000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122).'),
+  ('2026-10-29 00:00:00', 'openai/gpt-6.1-sol',    '2026-10-02', 272001, '', 'batch',
+    2000000,  7500000,  100000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-6-luna',     '2026-10-02',      0, '', 'batch',
+      50000,   250000,    5000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122).'),
+  ('2026-10-29 00:00:00', 'openai/gpt-6-luna',     '2026-10-02', 272001, '', 'batch',
+     100000,   375000,   10000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-sol',    '2026-09-02',      0, '', 'batch',
+    2000000, 10000000,  200000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122). Half the promotional standard rate, promotion at least through 2026-11-21.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-sol',    '2026-09-02', 272001, '', 'batch',
+    4000000, 15000000,  400000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-luna',   '2026-09-02',      0, '', 'batch',
+     100000,   600000,   10000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122).'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-luna',   '2026-09-02', 272001, '', 'batch',
+     200000,   900000,   20000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-terra',  '2026-09-02',      0, '', 'batch',
+    1000000,  6000000,  100000, 'https://developers.openai.com/api/docs/pricing',
+   'First batch observation (#716, 0122).'),
+  ('2026-10-29 00:00:00', 'openai/gpt-5.6-terra',  '2026-09-02', 272001, '', 'batch',
+    2000000,  9000000,  200000, 'https://developers.openai.com/api/docs/pricing',
+   'Batch long-context tier: >272K input tokens.');

package/migrations/0123_batch_files.sql

@@ -0,0 +1,50 @@
+-- Batch files, and what a batch job says about itself (#716, epic #712).
+--
+-- OpenAI's batch shape, which `/v1/files` and `/v1/batches` follow so that a
+-- stock SDK works against Cruise unchanged: a project uploads a JSONL file,
+-- then creates a batch naming it. The bytes live in R2 (`BATCH_FILES`); this
+-- table is what Cruise knows about them — whose, how big, and when they go.
+--
+-- **A file belongs to a project**, through the one path to that fact every
+-- other row uses (0004): a key's project, never a column that could disagree
+-- with it. A key from another project asking for a file id is answered 404,
+-- the same as an id that does not exist.
+--
+-- Deleted, not disabled: unlike a lane or a key, a file's bytes are the
+-- payload Cruise promised not to keep (#712, 24h after its job is over), and
+-- the ledger never references a file. `deleted_at` records that the bytes are
+-- gone; the row stays, so a job's `input_file_id` still names something.
+CREATE TABLE batch_files (
+  id          TEXT PRIMARY KEY,
+  project_id  TEXT NOT NULL REFERENCES projects(id),
+  purpose     TEXT NOT NULL CHECK (purpose IN ('batch', 'batch_output')),
+  filename    TEXT NOT NULL,
+  bytes       INTEGER NOT NULL CHECK (bytes >= 0),
+  -- The R2 object. Its own column rather than derived from the id, so a
+  -- change to the key layout is a migration and not a silent miss.
+  object_key  TEXT NOT NULL,
+  created_at  TEXT NOT NULL DEFAULT (datetime('now')),
+  deleted_at  TEXT
+);
+
+CREATE INDEX idx_batch_files_project ON batch_files (project_id, created_at);
+
+-- ## What a batch job carries
+--
+-- 0121 made the job and its money; these are the fields the Batch object
+-- reports back. `model` is the one model every line names — OpenAI requires a
+-- batch's lines to share one, and admission checks it once rather than per
+-- line. `line_count` is fixed at admission. `completed` and `failed` move as
+-- results are settled (#716's Workflow), and `provider_batch_id` is the
+-- provider's own id once the batch has been sent.
+ALTER TABLE jobs ADD COLUMN endpoint TEXT;
+ALTER TABLE jobs ADD COLUMN model TEXT;
+ALTER TABLE jobs ADD COLUMN input_file_id TEXT;
+ALTER TABLE jobs ADD COLUMN output_file_id TEXT;
+ALTER TABLE jobs ADD COLUMN error_file_id TEXT;
+ALTER TABLE jobs ADD COLUMN line_count INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE jobs ADD COLUMN completed_count INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE jobs ADD COLUMN failed_count INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE jobs ADD COLUMN provider_batch_id TEXT;
+-- The caller's `metadata`, as JSON: up to 16 string pairs, as OpenAI caps it.
+ALTER TABLE jobs ADD COLUMN metadata TEXT;

package/migrations/0124_batch_run.sql

@@ -0,0 +1,21 @@
+-- Running a batch (#716 part 2b, epic #712): what the Workflow needs to keep
+-- on the job between submit and settle.
+--
+-- `markup_basis_points` — the markup the hold was taken at. A line settles at
+-- its batch price plus this rate, never the rate in force when the provider
+-- finishes up to a day later: the hold was sized at the submit-time rate, and
+-- a markup raised overnight would settle a line above the bound that was
+-- reserved for it. The same rule `accounting.ts` keeps for a request, which
+-- carries the rate on its hold for the same reason (#99). 0 for a tenant that
+-- is not billed, and for every job written before this column.
+--
+-- `provider_status` — OpenAI's own word for the batch (`validating`,
+-- `in_progress`, `finalizing`, `completed`, `expired`, `cancelling`,
+-- `cancelled`, `failed`), as last polled. `jobs.status` keeps Cruise's five;
+-- this keeps the two distinctions they fold away and a caller still needs:
+-- an `expired` batch is a `failed` job whose finished lines were billed and
+-- returned, and its event is `batch.expired`, not `batch.failed`; and a
+-- `running` job the caller asked to cancel reads `cancelling` until the
+-- provider confirms it. Null until the batch has been sent.
+ALTER TABLE jobs ADD COLUMN markup_basis_points INTEGER NOT NULL DEFAULT 0 CHECK (markup_basis_points >= 0);
+ALTER TABLE jobs ADD COLUMN provider_status TEXT;

package/migrations/0125_anthropic_path_prefix.sql

@@ -0,0 +1,22 @@
+-- Anthropic's version segment moves from the call into its upstream row.
+--
+-- `prepareAnthropicStructuredAttempt` asked `providerPath` for `/v1/messages`.
+-- Every other provider-path caller passes a path without a version
+-- (`/embeddings`, `/responses`, `/images/generations`) and lets the upstream
+-- row's `provider_path_prefix` (0063) supply what that provider needs — Google's
+-- `/v1beta/openai`, xAI's `/v1`. Anthropic was the one caller carrying its own.
+--
+-- Through AI Gateway the two spellings are the same URL: the gateway forwards
+-- whatever follows the provider segment, so `…/anthropic` + `/v1/messages` and
+-- `…/anthropic` + prefix `/v1` + `/messages` both reach `api.anthropic.com/v1/messages`.
+-- The difference is a direct upstream, which joins `base_url` and the path with
+-- nothing between them (`providerPath`, `src/route-run.ts`). A self-hosted
+-- deployment with no gateway points Anthropic at `https://api.anthropic.com/v1`,
+-- because that is where its OpenAI-compatible `/chat/completions` lives — and
+-- with the version inside the call, structured output went to `/v1/v1/messages`.
+--
+-- Written only where nothing is set, so an operator who has already chosen a
+-- prefix for this row keeps it.
+
+UPDATE upstreams SET provider_path_prefix = '/v1'
+ WHERE id = 'anthropic' AND provider_path_prefix IS NULL;

package/package.json

@@ -1,6 +1,39 @@
 {
   "name": "@bytesbrains/cruise-worker",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "0.12.0",
+  "description": "The Cruise Worker, built and ready to deploy: one OpenAI-compatible endpoint in front of every model provider, with budgets that stop.",
+  "license": "Apache-2.0",
+  "type": "module",
+  "exports": {
+    ".": {
+      "types": "./index.d.ts",
+      "default": "./index.mjs"
+    },
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "index.mjs",
+    "index.d.ts",
+    "worker",
+    "migrations",
+    "wrangler.template.jsonc",
+    "README.md",
+    "LICENSE"
+  ],
+  "engines": {
+    "node": ">=22"
+  },
+  "repository": {
+    "type": "git",
+    "url": "https://gitlab.com/bytesbrains-oss/cruise.git",
+    "directory": "packages/worker"
+  },
+  "homepage": "https://gitlab.com/bytesbrains-oss/cruise",
+  "keywords": [
+    "llm",
+    "openai",
+    "gateway",
+    "cloudflare-workers",
+    "budget"
+  ]
+}

package/worker/VERSION

@@ -0,0 +1 @@
+0.12.0+c7b94bc