github beamivalice/sushi v1.7.3
sushi v1.7.3

one hour ago

v1.7.3 — Claude Code and tool calls that hold, requests checked end to end

Bugs fixed in this release, by severity and area:

Severity Memory & admission Server & API Models & loads Engine & tests Total
High 3 4 5 1 13
Medium 9 12 5 7 33
Low 7 9 3 9 28
Total 19 25 13 17 74
  • Claude Code sessions run again, and long tool calls keep streams alive: /v1/messages takes a system
    message inside messages again and reads it into the prompt where it sits, as 1.7.1 did
    (#115). While the model writes a long tool call, every surface
    now sends the call's opening and an empty arguments event every 5 s, so clients whose idle watchdog counts events
    (not keepalive comments) no longer give up and retry mid-call.
  • Every request field is checked and read from one table, per surface: a field a surface would have dropped, or
    one of the wrong shape, is a 400 naming it, where it was silently ignored (a top_k of 2.5 is refused, not served
    as 2; a Responses stop is read). Newly refused: OpenAI's legacy functions and function_call, the grammar
    and guided_* constraint fields, and Messages tools of any type but custom. Responses now serves
    stream_options; Codex's store, service_tier, safety_identifier and metadata stay accepted.
  • Tool calls come out right in a dozen more shapes: calls cut short or left open by the model's own end,
    backticked, spelling their own closer or </think> in a value, bare JSON after prose, and MiniCPM5, <invoke>,
    <tool> and DSML markup are each shipped as written or declined by name, never sent half-made or leaked into the
    reply's text; a reply of hundreds of calls settles in milliseconds again.
  • Memory bills match what runs: cache regrows, KV rebuild copies, MiMo's ring compaction and DeepSeek's row
    stores are billed where they happen, so a request never holds more than it was admitted for. A resident DeepSeek
    STQ 8,192-row chunk bills 9.30 GB, not 10.78; a DeepSeek request makes its row stores at its last row (a 1M-token
    request bills 2.51 GB of state, not 2.91); with the prefix cache off no SSM checkpoints are billed; a Sushi Cooler
    refusal names a context that is then admitted.
  • Concurrent speculation holds up: Qwen3.8-Flash-Next streams keep their grouped MTP speculation beside a
    DeepSeek-V4.1 request or a row that cannot join them, and concurrent speculative requests on Sushi Cooler no longer
    fail with PoolOverBudget.
  • The model list, downloads and web UI stay current: Rescan drops a deleted or unservable model and refreshes a
    replaced one (#114); a finished download shows within seconds,
    and Stop or shutdown during its last step no longer misreports or crashes; the web UI never shows an older answer
    over a newer one. --expert-cache-bias (0 to 2) now serves streamed Qwen STQ packs, and sushi kld capture always
    captures with exact routing.
  • Thanks @catalinb13 for the debug line that names every term of a prefill's memory bill (--log-level debug).

Don't miss a new sushi release

NewReleases is sending notifications on new releases.