Sync public snapshot from freebuff-private

Source: CodebuffAI/freebuff-private@e78db64132a11ed880cea51c541f70281338bdda

Author: github-actions[bot]

bun.lock

@@ -9,9 +9,9 @@ export const FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID = 'deepseek/deepseek-v4-flash'
 export const FREEBUFF_DEEPSEEK_V4_PRO_MODEL_ID = 'deepseek/deepseek-v4-pro'
 export const FREEBUFF_MINIMAX_M3_MODEL_ID = 'minimax/minimax-m3'
 /** DeepSeek V4.1 Flash in FAST MODE: the same model as
- *  FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID, served by DeepSeek's own API (a move
- *  to Luminal, web/src/llm-api/fast-mode-luminal.ts, was rolled back on
- *  2026-10-24 for decode speed) and run on the `base3-fast` harness,
+ *  FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID, served by Pareto with DeepSeek's own
+ *  API behind it (web/src/llm-api/deepseek-pareto.ts, 2026-10-24) and run on
+ *  the `base3-fast` harness,
  *  which gathers context with parallel subagents. Its own wire id because
  *  it is its own entitlement: one wire id per price is the rule, and this
  *  row is priced on DeepSeek's own peak and off-peak hours. */

common/src/constants/freebuff-model-ids.ts

@@ -1818,10 +1818,11 @@ const DEEPSEEK_V4_FLASH_MODEL = {
  *    parallel worker subagents that share its prompt cache, and verifies the
  *    combined result itself. Several requests in flight at once is the whole
  *    point, which is why it cannot ride the Flash row.
- *  - It is served by DeepSeek's own API. A move to Luminal (2026-10-24,
- *    web/src/llm-api/fast-mode-luminal.ts) was rolled back the same night:
- *    under production load Luminal decoded at a p50 of 137 tok/s against
- *    direct's ~220, and this row is sold on speed.
+ *  - It is served by Pareto with DeepSeek's own API behind it (2026-10-24,
+ *    web/src/llm-api/deepseek-pareto.ts): Pareto decodes as fast as direct
+ *    (p50 ~224 tok/s) at about half the cost, and the Flash row is held to
+ *    ~130 tok/s there so this row stays the fast one. A move to Luminal the
+ *    same night was rolled back: under load Luminal decoded at a p50 of 137.
  *
  * Priced on DeepSeek's own clock (freebuff-freebucks.ts): 50 Freebucks in
  * the weekday window where DeepSeek direct doubles, 25 the rest of the time.
@@ -1841,7 +1842,7 @@ const DEEPSEEK_V4_FLASH_FAST_MODEL = {
   displayName: 'DeepSeek V4.1 Flash Fast 10/17',
   tagline: 'Parallel agents',
   taglineTooltip:
-    'Gathers context with parallel subagents. Runs on DeepSeek’s own API.',
+    'Gathers context with parallel subagents.',
   availability: 'always',
   // Inert while `always`; names the row this one is a variant of.
   unavailableFallback: FREEBUFF_DEEPSEEK_V4_FLASH_MODEL_ID,