v1.1.1

Patch release, delivered through pull request #3.

[1.1.1] - 2026-10-31

Added

• Adaptive pacing: whenever the backend answers 429 and a request has
  to be retried, the pacer widens its minimum intervals automatically
  (the penalty doubles from 2 seconds up to a 30 second cap) and then
  relaxes with a 60 second half-life. Set DS_ADAPTIVE_PACING=0 to
  keep the intervals fixed.
• The server version is now reported by the /health endpoint and in
  the startup line, from a single __version__ constant in the
  package.
• /v1/models now declares each model's capabilities (vision,
  reasoning, tools, search) and its input and output modalities,
  so routers and clients can display and route on them instead of
  guessing from the model name. The declaration follows each model's
  own mode: deepseek-chat is the plain mode and declares no
  reasoning, deepseek-reasoner is the thinking mode.
• Regression fixture corpus under tests/fixtures covering the
  historical parser failure shapes: SSE patch sequences for thinking
  routed through the -1 fragment selector, fragment lists appended
  and batched at the response level, transition chunks with no
  operation plus standalone [reference:N] markers, and
  mixed-dialect tool replies including mojibake-damaged DSML.

Changed

• The model labels in the README and the module docs no longer call
  the web chat models DeepSeek V3 and R1. The web chat has run the
  DeepSeek V4 family since April 2026 (the vendor replaces the web
  checkpoint without notice, and the web API does not report which
  one served a request), so the models are now described as the web
  chat model in its plain and thinking modes.

Removed

• The legacy model aliases deepseek-v3 and deepseek-r1 are gone.
  Only deepseek-chat and deepseek-reasoner are served: lookups of
  any other model name return a model_not_found error, and chat
  completions reject unknown model names the same way instead of
  silently treating them as deepseek-chat.

Fixed

• A streamed request whose parsing produced no text now delivers the
  history fallback through the stream as one content piece, matching
  the buffered path; before, streaming clients silently received an
  empty reply in that case.
• Tool replies with fused dialects no longer leak markup into the
  visible text: a DSML invoke block closed by a tool_call tag (or
  by its calls wrapper) is now parsed as a call, and any orphaned
  DSML or tool_call tag debris left after extraction is stripped
  by a final sweep. Long fenced code blocks in tool-carrying
  replies stay capped at 10 lines plus a trimmed note, while plain
  replies are untouched; both behaviors are now covered by tests.
• The stream parser now handles fragment updates that arrive at
  the response level itself: lists appended with APPEND and bare
  fragment dicts inside BATCH updates are routed to the correct
  channel instead of being dropped. Standalone [reference:N]
  markers are now filtered from every channel, not only from
  fragment batch updates.
• Streaming cancellation now reacts even when the upstream stalls
  completely: the completion stream read re-checks the cancel event
  at least every 2 seconds, so a disconnected client no longer
  waits out a silent upstream. Normal streaming behavior and the
  pacing design are unchanged.

Assets

• ds-server.zip: the complete source package

Warning

Unofficial project. Automating the DeepSeek web interface may violate DeepSeek's Terms of Service and can get your account restricted or permanently banned. Use at your own risk. See the Disclaimer in the README.

Full history: CHANGELOG.md