v0.24.0

Minor Changes

• 650b4c1: Apply Synthetic's 80% cache-read discount in the model catalog instead of the streamSimple wrapper. The wrapper is removed and Pi's built-in cost calculation now uses the discounted cacheRead rate directly.

• a6f10fa: Add Kimi-K3, drop Kimi-K2.7-Code, and store cache-read price as reported by the API.

  • Add hf:moonshotai/Kimi-K3 (text+image, 524k context). Verified reasoning
    behavior via the Synthetic connector: backend accepts none/low/high/
    max, rejects medium; low emits no reasoning content. thinkingLevelMap
    exposes off/high/max.
  • Remove hf:moonshotai/Kimi-K2.7-Code (no longer in the Synthetic catalog).
  • Update syn:large:vision to its current backing model pricing/context
    (524k context, $3/$15) and the Kimi-K3 reasoning map.
  • Remove the automatic cache-read discount. The catalog and API-sourced
    models now store the input_cache_reads price as reported by the Synthetic
    API directly, with no 80% discount applied in code. syn:large:text and
    hf:zai-org/GLM-5.2 cacheRead updated from 0.2 to 0.16 to match the API.
    Verified against live billing: cache-read calls are charged at the API's
    input_cache_reads rate (e.g. $0.45/M for Kimi-K3), cache writes are free.
  • Add a thinkingLevelMap to hf:openai/gpt-oss-120b. Probed all efforts:
    the backend accepts none/minimal/low/medium/high/xhigh and
    rejects max (400). The model reasons at every accepted effort including
    none, so reasoning cannot be disabled (off hidden). Exposes
    minimal/low/medium/high/xhigh; hides max to avoid a 400 on the top level.

Patch Changes

• 496de92: Update GLM-5.2 pricing in the hardcoded fallback catalog to match the Synthetic API: input and cache-read drop from $1.40 to $1.00 per million tokens, output drops from $4.40 to $3.00 per million tokens.