@bytesbrains/cruise-worker 0.0.0-stage → 0.12.1

package/migrations/0023_model_limits.sql

@@ -0,0 +1,48 @@
+-- Give `checkLimits` something to check.
+--
+-- `max_context` and `max_output` have been columns since 0007 and NULL on every
+-- row since. `checkLimits` reads `max_output` and returns ok when it is NULL, so
+-- the output cap has been live code enforcing nothing — a sixth instance of the
+-- pattern this project keeps recording, and a new variant: not a dead column,
+-- but a live gate with no data. It even has a passing test, because the test
+-- seeds the value the production rows lack.
+--
+-- **Every number here was fetched from the host's own page, not from the model's.**
+-- That distinction is the whole of this migration.
+--
+-- naderu.com's model board tracks 44 models with verified prices and context
+-- windows, refreshed weekly, and is a better-disciplined source than anything we
+-- would write. But it is keyed on the **model**, and Cruise is keyed on the
+-- **model at a given upstream**. For closed-weights models those are the same
+-- thing. For open weights they are not, and the gap is not small:
+--
+--   deepseek-v4-flash   board 1M / 384K      DeepSeek 1M / 384K      identical
+--   llama-3.3-70b       board 131,072        Workers AI 24,000       5.5x apart
+--
+-- Workers AI serves Llama with a 24,000-token window, roughly a fifth of the
+-- model's native 131k. Importing the board's figure would have had Cruise accept
+-- requests up to 131k against a deployment that takes 24k — and the gate that
+-- exists to catch exactly that would have passed them through. The same shape as
+-- 0020's price error, in the same direction: model-level data applied to a
+-- host-level deployment.
+
+-- Workers AI, https://developers.cloudflare.com/workers-ai/models/llama-3.3-70b-instruct-fp8-fast/
+-- The page states a 24,000-token context window. It states a *default* max_tokens
+-- of 256 and no upper bound, so `max_output` stays NULL rather than being
+-- invented — the rule from 0011 and 0020 holds: never write a number you did not
+-- fetch. `checkLimits` now derives a bound from the window instead.
+UPDATE models
+   SET max_context = 24000
+ WHERE id = 'workers-ai/@cf/meta/llama-3.3-70b-instruct-fp8-fast';
+
+-- DeepSeek, https://api-docs.deepseek.com/quick_start/pricing
+-- 1M context, 384K maximum output, both stated.
+UPDATE models
+   SET max_context = 1000000, max_output = 384000
+ WHERE id = 'deepseek/deepseek-v4-flash';
+
+-- Mistral, https://docs.mistral.ai/models/model-cards/mistral-small-4-0-26-03
+-- 256k context stated; maximum output not stated on the card, so left NULL.
+UPDATE models
+   SET max_context = 262144
+ WHERE id = 'mistral/mistral-small-latest';

package/migrations/0024_conditional_prices.sql

@@ -0,0 +1,84 @@
+-- A price that depends on when you ask, and on how much you send.
+--
+-- `model_prices` holds one rate per model per effective date. Two providers on
+-- the board now price by conditions that a flat row cannot express, and one of
+-- them is a provider we route:
+--
+--   DeepSeek, since 16:00 UTC on 2026-08-16: every rate doubles during peak
+--   hours, 01:00-04:00 and 06:00-10:00 UTC. Off-peak is the rate "most of the
+--   day"; peak is exactly 2x.            https://api-docs.deepseek.com/news/news260813
+--
+--   Gemini 3.1 Pro and Grok 4.5/4.6: at or above 200,000 input tokens the rate
+--   doubles, for the whole request. Qwen3.7-Plus steps at 256K, MiniMax M3 at
+--   512K. None is on a keyed upstream yet; all arrive with the first OpenRouter
+--   key, and the shape is the same.                   https://naderu.com/models-board
+--
+-- So a price row gains two conditions, each with a value meaning "no condition":
+--
+--   prompt_tokens_from  the row applies when the request's prompt is at least
+--                       this many tokens. 0 is the base tier. Tiers apply to the
+--                       whole request, which is how Google and xAI bill it; if a
+--                       provider turns out to bill marginally, correcting down is
+--                       the cheap direction.
+--   peak_utc            comma-separated HH:MM-HH:MM windows, UTC, start inclusive
+--                       and end exclusive, during which the row applies. '' is
+--                       any hour.
+--
+-- The resolver takes the rows in force, keeps the highest tier at or below the
+-- prompt, and within it prefers a window containing the request's time over the
+-- '' row. Both are '' / 0 for every existing row, so every existing lookup
+-- resolves exactly as before.
+--
+-- **Not** added: a region. No keyed provider prices by geography today, and a
+-- column with nothing to read it has been wrong six times out of six in this
+-- repository. When the first regionally priced upstream is routed, it is one more
+-- condition column on this row and the same resolver.
+--
+-- Primary key changes, so the table is rebuilt. The '' and 0 defaults are what
+-- let the key stay NOT NULL: SQLite admits NULL in a primary key and then cannot
+-- enforce uniqueness across it.
+
+CREATE TABLE model_prices_new (
+  model                        TEXT    NOT NULL,
+  effective_from               TEXT    NOT NULL,
+  prompt_tokens_from           INTEGER NOT NULL DEFAULT 0,
+  peak_utc                     TEXT    NOT NULL DEFAULT '',
+  input_micros_per_mtok        INTEGER NOT NULL,
+  output_micros_per_mtok       INTEGER NOT NULL,
+  cached_input_micros_per_mtok INTEGER,
+  source                       TEXT    NOT NULL,
+  recorded_at                  TEXT    NOT NULL DEFAULT (datetime('now')),
+  PRIMARY KEY (model, effective_from, prompt_tokens_from, peak_utc)
+);
+
+INSERT INTO model_prices_new
+  (model, effective_from, input_micros_per_mtok, output_micros_per_mtok,
+   cached_input_micros_per_mtok, source, recorded_at)
+SELECT model, effective_from, input_micros_per_mtok, output_micros_per_mtok,
+       cached_input_micros_per_mtok, source, recorded_at
+  FROM model_prices;
+
+DROP TABLE model_prices;
+ALTER TABLE model_prices_new RENAME TO model_prices;
+
+-- DeepSeek V4-Flash at the price it has actually charged since 2026-08-16.
+--
+-- The row on record is $0.14 / $0.28 as of 2026-08-14, which was right for two
+-- days. Since 16:00 UTC on the 16th the ledger has under-counted every DeepSeek
+-- request by 1.6x off-peak and up to 4.7x at peak — the failure the budget exists
+-- to prevent, live for six days. The effective date is the calendar day, which
+-- over-counts the sixteen hours before the change took effect; over-counting is
+-- the safe direction and the alternative is a timestamped effective_from for one
+-- row.
+--
+-- Figures from https://api-docs.deepseek.com/quick_start/pricing, fetched
+-- 2026-08-22: off-peak $0.22 in / $0.007 cache hit / $0.66 out; peak $0.44 /
+-- $0.014 / $1.32. Model version DeepSeek-V4-Flash-0731.
+INSERT INTO model_prices
+  (model, effective_from, prompt_tokens_from, peak_utc,
+   input_micros_per_mtok, output_micros_per_mtok, cached_input_micros_per_mtok, source)
+VALUES
+  ('deepseek/deepseek-v4-flash', '2026-08-16', 0, '',
+   220000, 660000, 7000, 'https://api-docs.deepseek.com/quick_start/pricing'),
+  ('deepseek/deepseek-v4-flash', '2026-08-16', 0, '01:00-04:00,06:00-10:00',
+   440000, 1320000, 14000, 'https://api-docs.deepseek.com/quick_start/pricing');

package/migrations/0025_catalogue_from_keyed_hosts.sql

@@ -0,0 +1,172 @@
+-- Every chat model the three keyed hosts serve, at the price and window each
+-- host states.
+--
+-- Three models were routable. A lane (auto-mode.md) is a choice among
+-- candidates, and with three candidates across eight jobs most lanes would have
+-- had one or none. This is the whole current catalogue of the upstreams we hold
+-- keys for — DeepSeek, Mistral, Workers AI — minus what is retiring, unpriced,
+-- or not a chat model.
+--
+-- Rules, the same as 0011, 0020 and 0023:
+--
+--   Every number is quoted from the host's own page, fetched 2026-08-21 UTC.
+--   A figure the host does not state stays NULL. `max_output` is NULL for all
+--   but four models because no host page states a maximum; `checkLimits`
+--   derives a bound from the window (0023).
+--
+--   Where two host sources disagree, the gate takes the smaller. Mistral's card
+--   says 256k for ministral-3b and the API says 131072; the card says 128k for
+--   codestral and the API says 256000. Refusing a request the host would have
+--   taken is visible and correctable; passing one it would not is neither.
+--
+--   Nothing here is routable yet. A row is admitted by conformance measuring it
+--   (#10), which is the only writer of `measured_at`. Inserting is not
+--   admitting.
+--
+--   `effective_from` is today in UTC, 2026-08-21. The wall clock that wrote this
+--   read 03:00 on the 22nd. See 0020 and 0021.
+--
+-- Excluded, and why:
+--   Workers AI: 17 models marked "Deprecated 5/30/2026" on their pages, a date
+--   already past; 9 beta LoRA-era models with no price stated;
+--   llama-3.1-8b-instruct-fast, which has a page and no price; llama-guard, a
+--   classifier. Mistral: mistral-medium-2505 and -2508 (retire 2026-08-31),
+--   devstral-2512 (retired 2026-07-31), magistral (deprecated), voxtral (audio).
+--   Every `-latest` alias: an alias moving is a rate change with no signal
+--   (0020), so each row pins the dated id the host lists.
+--
+-- Three of the Workers AI rows are the same weights DeepSeek serves under its
+-- own key, at DeepSeek's peak rate all day and without a peak window. Whether
+-- that is cheaper depends on the hour, which is precisely the decision a lane
+-- exists to make and a human never would.
+
+-- ── DeepSeek ──────────────────────────────────────────────────────────────────
+-- https://api-docs.deepseek.com/quick_start/pricing — 1M context, 384K maximum
+-- output, both stated for all three; tool calls and JSON output stated for all
+-- three; image input only on the vision model (guides/vision). Peak 01:00-04:00
+-- and 06:00-10:00 UTC at 2x, in force since 16:00 UTC 2026-08-16 (news260813).
+
+INSERT INTO models (id, upstream_id, upstream_model, max_context, max_output) VALUES
+  ('deepseek/deepseek-v4-pro',              'deepseek', 'deepseek-v4-pro',              1000000, 384000),
+  -- Experimental on the host's own label; images are billed as input tokens at
+  -- V4-Flash rates, up to 384 tokens each (news260821).
+  ('deepseek/deepseek-v4-flash-vision-exp', 'deepseek', 'deepseek-v4-flash-vision-exp', 1000000, 384000);
+
+INSERT INTO model_prices
+  (model, effective_from, prompt_tokens_from, peak_utc,
+   input_micros_per_mtok, output_micros_per_mtok, cached_input_micros_per_mtok, source)
+VALUES
+  ('deepseek/deepseek-v4-pro', '2026-08-21', 0, '',
+   660000, 1980000, 22000, 'https://api-docs.deepseek.com/quick_start/pricing'),
+  ('deepseek/deepseek-v4-pro', '2026-08-21', 0, '01:00-04:00,06:00-10:00',
+   1320000, 3960000, 44000, 'https://api-docs.deepseek.com/quick_start/pricing'),
+  ('deepseek/deepseek-v4-flash-vision-exp', '2026-08-21', 0, '',
+   220000, 660000, 7000, 'https://api-docs.deepseek.com/quick_start/pricing'),
+  ('deepseek/deepseek-v4-flash-vision-exp', '2026-08-21', 0, '01:00-04:00,06:00-10:00',
+   440000, 1320000, 14000, 'https://api-docs.deepseek.com/quick_start/pricing');
+
+-- ── Mistral ───────────────────────────────────────────────────────────────────
+-- Prices: https://docs.mistral.ai/inference/pricing (cached input is 10% of
+-- input for every listed model). Windows: each model's card, which states a
+-- rounded figure ("256k"), read against the API's max_context_length, which
+-- states an exact one; the smaller wins where they differ. Max output is stated
+-- on two cards only.
+--
+-- `mistral-medium-2604` is the API's id; the card is titled Mistral Medium 3.5
+-- and never prints "2604". The API's own alias list maps the two
+-- (mistral-medium-2604 ↔ mistral-medium-3-5), which is the host saying so.
+
+INSERT INTO models (id, upstream_id, upstream_model, max_context, max_output) VALUES
+  -- https://docs.mistral.ai/models/mistral-large-3-25-12
+  ('mistral/mistral-large-2512',  'mistral', 'mistral-large-2512',  262144, NULL),
+  -- https://docs.mistral.ai/models/mistral-medium-3-5-26-04
+  ('mistral/mistral-medium-2604', 'mistral', 'mistral-medium-2604', 262144, NULL),
+  -- https://docs.mistral.ai/models/ministral-3-3b-25-12 — card 256k, API 131072
+  ('mistral/ministral-3b-2512',   'mistral', 'ministral-3b-2512',   131072, NULL),
+  -- https://docs.mistral.ai/models/ministral-3-8b-25-12
+  ('mistral/ministral-8b-2512',   'mistral', 'ministral-8b-2512',   262144, NULL),
+  -- https://docs.mistral.ai/models/ministral-3-14b-25-12
+  ('mistral/ministral-14b-2512',  'mistral', 'ministral-14b-2512',  262144, NULL),
+  -- https://docs.mistral.ai/models/codestral-25-08 — card 128k, API 256000
+  ('mistral/codestral-2508',      'mistral', 'codestral-2508',      131072, NULL),
+  -- https://docs.mistral.ai/models/zai-glm-5-2 — 1M context, 128k max output.
+  -- Third-party weights, public preview, on the host's label.
+  ('mistral/zai-glm-5-2',         'mistral', 'zai-glm-5-2',         1048576, 128000),
+  -- https://docs.mistral.ai/models/leanstral-1-5 — 256k, 128k max output.
+  -- Public preview and free, which a price-first lane will prefer on sight; the
+  -- quality data that should temper that is exactly what conformance and the
+  -- ledger exist to collect.
+  ('mistral/labs-leanstral-1-5',  'mistral', 'labs-leanstral-1-5',  262144, 128000);
+
+INSERT INTO model_prices
+  (model, effective_from, input_micros_per_mtok, output_micros_per_mtok,
+   cached_input_micros_per_mtok, source)
+VALUES
+  ('mistral/mistral-large-2512',  '2026-08-21',  500000, 1500000,  50000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/mistral-medium-2604', '2026-08-21', 1500000, 7500000, 150000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/ministral-3b-2512',   '2026-08-21',  100000,  100000,  10000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/ministral-8b-2512',   '2026-08-21',  150000,  150000,  15000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/ministral-14b-2512',  '2026-08-21',  200000,  200000,  20000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/codestral-2508',      '2026-08-21',  300000,  900000,  30000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/zai-glm-5-2',         '2026-08-21', 1400000, 4400000, 140000, 'https://docs.mistral.ai/inference/pricing'),
+  ('mistral/labs-leanstral-1-5',  '2026-08-21',       0,       0,      0, 'https://docs.mistral.ai/inference/pricing');
+
+-- ── Workers AI ────────────────────────────────────────────────────────────────
+-- Prices: https://developers.cloudflare.com/workers-ai/platform/pricing/, which
+-- prints three decimals where the model pages round to two (0011 took the
+-- pricing page for the same reason). Windows: each model's page at
+-- https://developers.cloudflare.com/workers-ai/models/<slug>/. No page states a
+-- maximum output. Five are "paid billing only" on the pricing page.
+
+INSERT INTO models (id, upstream_id, upstream_model, max_context) VALUES
+  ('workers-ai/@cf/deepseek-ai/deepseek-r1-distill-qwen-32b',  'workers-ai', '@cf/deepseek-ai/deepseek-r1-distill-qwen-32b',    80000),
+  ('workers-ai/@cf/deepseek-ai/deepseek-v4-flash-0731',        'workers-ai', '@cf/deepseek-ai/deepseek-v4-flash-0731',        1048576),
+  ('workers-ai/@cf/deepseek-ai/deepseek-v4-pro-0813',          'workers-ai', '@cf/deepseek-ai/deepseek-v4-pro-0813',          1048576),
+  ('workers-ai/@cf/google/gemma-4-26b-a4b-it',                 'workers-ai', '@cf/google/gemma-4-26b-a4b-it',                  256000),
+  ('workers-ai/@cf/aisingapore/gemma-sea-lion-v4-27b-it',      'workers-ai', '@cf/aisingapore/gemma-sea-lion-v4-27b-it',       128000),
+  ('workers-ai/@cf/zai-org/glm-4.7-flash',                     'workers-ai', '@cf/zai-org/glm-4.7-flash',                      131072),
+  ('workers-ai/@cf/zai-org/glm-5.2',                           'workers-ai', '@cf/zai-org/glm-5.2',                            262144),
+  ('workers-ai/@cf/openai/gpt-oss-120b',                       'workers-ai', '@cf/openai/gpt-oss-120b',                        128000),
+  ('workers-ai/@cf/openai/gpt-oss-20b',                        'workers-ai', '@cf/openai/gpt-oss-20b',                         128000),
+  ('workers-ai/@cf/ibm-granite/granite-4.0-h-micro',           'workers-ai', '@cf/ibm-granite/granite-4.0-h-micro',            131000),
+  ('workers-ai/@cf/moonshotai/kimi-k2.6',                      'workers-ai', '@cf/moonshotai/kimi-k2.6',                       262144),
+  ('workers-ai/@cf/moonshotai/kimi-k2.7-code',                 'workers-ai', '@cf/moonshotai/kimi-k2.7-code',                  262144),
+  ('workers-ai/@cf/meta/llama-3.1-8b-instruct-fp8',            'workers-ai', '@cf/meta/llama-3.1-8b-instruct-fp8',              32000),
+  ('workers-ai/@cf/meta/llama-3.2-11b-vision-instruct',        'workers-ai', '@cf/meta/llama-3.2-11b-vision-instruct',         128000),
+  ('workers-ai/@cf/meta/llama-3.2-1b-instruct',                'workers-ai', '@cf/meta/llama-3.2-1b-instruct',                  60000),
+  ('workers-ai/@cf/meta/llama-3.2-3b-instruct',                'workers-ai', '@cf/meta/llama-3.2-3b-instruct',                  80000),
+  ('workers-ai/@cf/meta/llama-4-scout-17b-16e-instruct',       'workers-ai', '@cf/meta/llama-4-scout-17b-16e-instruct',        131000),
+  ('workers-ai/@cf/mistralai/mistral-small-3.1-24b-instruct',  'workers-ai', '@cf/mistralai/mistral-small-3.1-24b-instruct',   128000),
+  ('workers-ai/@cf/nvidia/nemotron-3-120b-a12b',               'workers-ai', '@cf/nvidia/nemotron-3-120b-a12b',                256000),
+  ('workers-ai/@cf/qwen/qwen2.5-coder-32b-instruct',           'workers-ai', '@cf/qwen/qwen2.5-coder-32b-instruct',             32768),
+  ('workers-ai/@cf/qwen/qwen3-30b-a3b-fp8',                    'workers-ai', '@cf/qwen/qwen3-30b-a3b-fp8',                      32768),
+  ('workers-ai/@cf/qwen/qwen3.8-27b',                          'workers-ai', '@cf/qwen/qwen3.8-27b',                           262144),
+  ('workers-ai/@cf/qwen/qwq-32b',                              'workers-ai', '@cf/qwen/qwq-32b',                                24000);
+
+INSERT INTO model_prices
+  (model, effective_from, input_micros_per_mtok, output_micros_per_mtok,
+   cached_input_micros_per_mtok, source)
+VALUES
+  ('workers-ai/@cf/deepseek-ai/deepseek-r1-distill-qwen-32b', '2026-08-21',  497000, 4881000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/deepseek-ai/deepseek-v4-flash-0731',       '2026-08-21',  440000, 1320000,  14000, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/deepseek-ai/deepseek-v4-pro-0813',         '2026-08-21', 1320000, 3960000,  44000, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/google/gemma-4-26b-a4b-it',                '2026-08-21',  100000,  300000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/aisingapore/gemma-sea-lion-v4-27b-it',     '2026-08-21',  351000,  555000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/zai-org/glm-4.7-flash',                    '2026-08-21',   60000,  400000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/zai-org/glm-5.2',                          '2026-08-21', 1400000, 4400000, 260000, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/openai/gpt-oss-120b',                      '2026-08-21',  350000,  750000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/openai/gpt-oss-20b',                       '2026-08-21',  200000,  300000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/ibm-granite/granite-4.0-h-micro',          '2026-08-21',   17000,  112000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/moonshotai/kimi-k2.6',                     '2026-08-21',  950000, 4000000, 160000, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/moonshotai/kimi-k2.7-code',                '2026-08-21',  950000, 4000000, 190000, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/meta/llama-3.1-8b-instruct-fp8',           '2026-08-21',  152000,  287000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/meta/llama-3.2-11b-vision-instruct',       '2026-08-21',   49000,  676000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/meta/llama-3.2-1b-instruct',               '2026-08-21',   27000,  201000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/meta/llama-3.2-3b-instruct',               '2026-08-21',   51000,  335000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/meta/llama-4-scout-17b-16e-instruct',      '2026-08-21',  270000,  850000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/mistralai/mistral-small-3.1-24b-instruct', '2026-08-21',  351000,  555000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/nvidia/nemotron-3-120b-a12b',              '2026-08-21',  500000, 1500000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/qwen/qwen2.5-coder-32b-instruct',          '2026-08-21',  660000, 1000000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/qwen/qwen3-30b-a3b-fp8',                   '2026-08-21',   51000,  335000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/qwen/qwen3.8-27b',                         '2026-08-21',  450000, 3200000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/'),
+  ('workers-ai/@cf/qwen/qwq-32b',                             '2026-08-21',  660000, 1000000,   NULL, 'https://developers.cloudflare.com/workers-ai/platform/pricing/');

package/migrations/0026_lanes.sql

@@ -0,0 +1,83 @@
+-- Lanes: a model id that names a job, allocated to a concrete model at request
+-- time.
+--
+-- auto-mode.md designed this as `auto/code:saver`. It ships as `bb/<job>`, a
+-- row in `/v1/models` like any other, because that is the form any
+-- OpenAI-compatible client can consume with no new code: the model is an opaque
+-- string in one field, and nothing parses it. The consuming side (agents#574)
+-- treats any row carrying `x-cruise.lane: true` as a lane and renders its
+-- members.
+--
+-- A lane is data, not a constant, for the reason this repository keeps
+-- recording: adding a model to a lane is a data change, the same as adding a
+-- model. `lane_members` is the candidate set — not a priority. Allocation is by
+-- policy: the cheapest member that passes every hard gate for *this* request
+-- (`src/lanes.ts`). `position` keeps the listing stable and means nothing else.
+--
+-- Members may be unmeasured. Conformance admits them to the lane the same way
+-- it admits them to `/v1/models`; a lane lists and allocates only what is
+-- routable for the asking key.
+--
+-- `role` is the board's job vocabulary (naderu.com/models-board), so the word
+-- on the board and the word in the API are the same word.
+
+CREATE TABLE lanes (
+  id          TEXT PRIMARY KEY CHECK (id LIKE 'bb/%'),
+  role        TEXT NOT NULL,
+  description TEXT NOT NULL,
+  enabled     INTEGER NOT NULL DEFAULT 1,
+  created_at  TEXT NOT NULL DEFAULT (datetime('now'))
+);
+
+CREATE TABLE lane_members (
+  lane_id  TEXT NOT NULL REFERENCES lanes(id),
+  model_id TEXT NOT NULL REFERENCES models(id),
+  position INTEGER NOT NULL,
+  PRIMARY KEY (lane_id, model_id)
+);
+CREATE INDEX idx_lane_members_lane ON lane_members (lane_id, position);
+
+-- The first lanes. Members are the models measured on staging on 2026-08-21
+-- that fit the job; the five strongest Workers AI models sit behind the paid
+-- plan and join when that changes.
+
+INSERT INTO lanes (id, role, description) VALUES
+  ('bb/security-review', 'code-review',
+   'Reads a diff or a repository and reasons about what is wrong with it. Tool calling and a large window; quality over price.'),
+  ('bb/agentic-coding', 'agentic-coding',
+   'Long tool-use loops over a repository. Output tokens dominate the bill.'),
+  ('bb/extraction', 'bulk-extraction',
+   'Classification and extraction across many rows. Cheapest that clears the bar.'),
+  ('bb/cheap-fast', 'lowest-latency',
+   'Interactive turns where a small model is enough.'),
+  ('bb/long-context', 'long-context',
+   'Whole documents and codebases in one prompt.');
+
+INSERT INTO lane_members (lane_id, model_id, position) VALUES
+  ('bb/security-review', 'deepseek/deepseek-v4-pro',                  1),
+  ('bb/security-review', 'mistral/mistral-large-2512',                2),
+  ('bb/security-review', 'mistral/mistral-medium-2604',               3),
+  ('bb/security-review', 'workers-ai/@cf/openai/gpt-oss-120b',        4),
+  ('bb/security-review', 'deepseek/deepseek-v4-flash',                5),
+
+  ('bb/agentic-coding',  'deepseek/deepseek-v4-pro',                  1),
+  ('bb/agentic-coding',  'mistral/mistral-medium-2604',               2),
+  ('bb/agentic-coding',  'mistral/codestral-2508',                    3),
+  ('bb/agentic-coding',  'workers-ai/@cf/qwen/qwen3.8-27b',           4),
+  ('bb/agentic-coding',  'deepseek/deepseek-v4-flash',                5),
+
+  ('bb/extraction',      'deepseek/deepseek-v4-flash',                1),
+  ('bb/extraction',      'mistral/mistral-small-latest',              2),
+  ('bb/extraction',      'mistral/ministral-8b-2512',                 3),
+  ('bb/extraction',      'workers-ai/@cf/zai-org/glm-4.7-flash',      4),
+  ('bb/extraction',      'workers-ai/@cf/ibm-granite/granite-4.0-h-micro', 5),
+
+  ('bb/cheap-fast',      'mistral/ministral-3b-2512',                 1),
+  ('bb/cheap-fast',      'workers-ai/@cf/meta/llama-3.2-3b-instruct', 2),
+  ('bb/cheap-fast',      'workers-ai/@cf/meta/llama-3.2-1b-instruct', 3),
+  ('bb/cheap-fast',      'workers-ai/@cf/ibm-granite/granite-4.0-h-micro', 4),
+
+  ('bb/long-context',    'deepseek/deepseek-v4-flash',                1),
+  ('bb/long-context',    'deepseek/deepseek-v4-pro',                  2),
+  ('bb/long-context',    'mistral/mistral-medium-2604',               3),
+  ('bb/long-context',    'mistral/ministral-14b-2512',                4);

package/migrations/0027_lane_alternatives.sql

@@ -0,0 +1,36 @@
+-- What each other member would have cost, for the usage that occurred.
+--
+-- A lane allocates the cheapest member passing the request's gates (0026). The
+-- ledger records what that member cost. It does not record what the lane
+-- *saved*, and a number that is not recorded per request cannot be reported
+-- later without being reconstructed — from prices that will have changed and
+-- candidate sets that will have changed with them.
+--
+-- So each lane request writes one row per member that could have served it,
+-- the served one included at rank 1, priced at that member's rate at that hour
+-- for the tokens the request actually used. A report then answers "what did
+-- the lane save against the dearest member", "against the member a human
+-- would have pinned", or "against the median" from rows that were true when
+-- they were written.
+--
+-- The approximation, stated once: the alternative's cost is priced on the
+-- *served* member's token counts. A different model tokenises differently and
+-- would have written a different completion. This is the cost of the same
+-- work at the other member's rate, which is the number a report can defend;
+-- it is not a claim about what the other member would have done.
+--
+-- `position` is the allocation rank: 1 served, 2 would have served had 1
+-- failed, and so on. A member the gates rejected has no row — it could not
+-- have served, so it is not an alternative.
+
+CREATE TABLE lane_alternatives (
+  ledger_id   TEXT    NOT NULL REFERENCES ledger(id) ON DELETE CASCADE,
+  model       TEXT    NOT NULL,
+  position    INTEGER NOT NULL,
+  cost_micros INTEGER NOT NULL,
+  -- 0 when the member had no price in force, in which case cost_micros is 0
+  -- and the row is excluded from savings rather than counted as free.
+  priced      INTEGER NOT NULL,
+  PRIMARY KEY (ledger_id, model)
+);
+CREATE INDEX idx_lane_alternatives_ledger ON lane_alternatives (ledger_id);

package/migrations/0028_lanes_tenant_scope.sql

@@ -0,0 +1,66 @@
+-- A lane belongs to a tenant, or to everyone.
+--
+-- Every lane so far is ours and global: `bb/extraction` is the same for every
+-- tenant and changing it is a migration. The value of a lane to a user is that
+-- it is config they can change at runtime (#66), which needs a lane to have an
+-- owner. This is the schema half; the routes come in #68 and nothing
+-- user-visible changes here.
+--
+-- Namespace. `bb/` stays ours. A tenant's lanes are prefixed with the tenant id
+-- — `acme/review` — which is what the key already carries, so a lane id says
+-- who owns it. Tenants have no slug, so the id is the namespace and must stay
+-- immutable.
+--
+-- The CHECK compares the prefix exactly, with substr() and `=`, not LIKE.
+-- LIKE is case-insensitive and reads `%` and `_` as wildcards, so under LIKE a
+-- tenant id of `a%` or `Acme` would satisfy the check for `acme/review` —
+-- another tenant's namespace. Raised in review of PR #73. What the CHECK still
+-- cannot enforce is that a tenant id is not also an upstream id (`deepseek/...`
+-- is a model): the route that creates tenants must refuse those and a strict
+-- charset, and #68 records that. The request path is defensive regardless — a
+-- lane lookup that misses falls through to the model lookup, so a prefix can
+-- never hide a model.
+--
+-- SQLite cannot alter a CHECK, and `lane_members` references `lanes`, so both
+-- are rebuilt: new tables, copy, drop the old in child-then-parent order, and
+-- rename — which rewrites the foreign key in `lane_members_new` to follow the
+-- parent's new name.
+
+CREATE TABLE lanes_new (
+  id          TEXT PRIMARY KEY,
+  -- NULL is a global lane, visible to every tenant and writable by none.
+  tenant_id   TEXT REFERENCES tenants(id),
+  role        TEXT NOT NULL,
+  description TEXT NOT NULL,
+  enabled     INTEGER NOT NULL DEFAULT 1,
+  created_by  TEXT NOT NULL DEFAULT 'migration',
+  created_at  TEXT NOT NULL DEFAULT (datetime('now')),
+  updated_at  TEXT NOT NULL DEFAULT (datetime('now')),
+  CHECK (
+    (tenant_id IS NULL     AND substr(id, 1, 3) = 'bb/') OR
+    (tenant_id IS NOT NULL AND substr(id, 1, length(tenant_id) + 1) = tenant_id || '/'
+                           AND substr(id, 1, 3) <> 'bb/')
+  )
+);
+
+INSERT INTO lanes_new (id, tenant_id, role, description, enabled, created_at, updated_at)
+SELECT id, NULL, role, description, enabled, created_at, created_at FROM lanes;
+
+CREATE TABLE lane_members_new (
+  lane_id  TEXT NOT NULL REFERENCES lanes_new(id),
+  model_id TEXT NOT NULL REFERENCES models(id),
+  position INTEGER NOT NULL,
+  PRIMARY KEY (lane_id, model_id)
+);
+
+INSERT INTO lane_members_new (lane_id, model_id, position)
+SELECT lane_id, model_id, position FROM lane_members;
+
+DROP INDEX idx_lane_members_lane;
+DROP TABLE lane_members;
+DROP TABLE lanes;
+ALTER TABLE lanes_new RENAME TO lanes;
+ALTER TABLE lane_members_new RENAME TO lane_members;
+
+CREATE INDEX idx_lane_members_lane ON lane_members (lane_id, position);
+CREATE INDEX idx_lanes_tenant ON lanes (tenant_id, enabled);