v0.3.172-alpha

Fixed

• Phi models stop at the end of their turn instead of inventing the next one.
  Every Phi-3, Phi-3.5 and Phi-4 model declares one token as its end of sequence
  and then ends each turn with a different one, and nothing said so, so nothing
  stopped the reply there. Asked "say exactly: hello", a Phi model answered, then
  invented a second answer, then made up a question from the user and began
  answering that too, until it ran out of room. What you saw depended on the
  model's vocabulary: some carried a stray <|end|> marker into the visible text,
  others hid the seam entirely and simply rambled. A model's turn-ending tokens
  are now found in its own vocabulary rather than taken on trust, so both
  symptoms are gone, and models whose format uses that same token to separate
  parts of one long answer are unaffected.
• Tool calls written in Phi-4's format are recognised. Phi-4-mini is built by
  its publisher for function calling and writes its calls in its own documented
  format, which nothing here could read — so every call it made came back as
  ordinary text with no error, and a client driving it would carry out nothing and
  have no way to tell. Both spellings it uses are now understood. Note that this
  model still cannot be served from shards at all, for a separate reason found
  while testing this release and not yet fixed, so the change is verified against
  its documented formats rather than against the model itself.
• A model whose template mentions tools without using them is told about them
  anyway. One template asks for its tools in a place they were never offered,
  so its own tool section never appeared, and the written description was skipped
  as redundant — leaving the model told nothing at all about tools that were
  attached to the request. A reply built with tools that shows none of them is
  now treated as a failed attempt and the written description is used instead.
• Tool calls from Qwen2.5-Coder are recognised. Its own prompt names two
  similar tags a few lines apart — one for the list of tools it has, one for
  writing a call — and the model reaches for the wrong one. The call it wrote was
  perfectly well formed and came back as plain text, so that model made no working
  tool call at all. Both tags are now understood.
• A model repeating its tool list back is no longer mistaken for calling one.
  A tool's description and a call to it are nearly the same shape, and the
  description was being accepted as a call — with the tool's own parameter
  definition passed along where the argument values belong. A client would have
  received a call it never asked for, carrying nothing it could use. Zero-argument
  calls, which look similar, still work.
• A tool call inside a code block followed by an explanation is read
  correctly. Reading the block required the whole reply to be nothing but that
  block, so the commonest thing a chatty model does — write the call, then
  explain it — defeated it, and a perfectly good call was handed back as text. A
  leftover code-fence marker no longer appears as the reply's visible content
  either.

Changed

• Models are asked for a simpler shape when they have no tool format of their
  own. Two of the fields being requested were discarded on arrival, so asking
  for them only added nesting for a small model to miscount. Both the old and the
  new shape are still understood.
• A request whose model has no readable chat template now says so in the log.
  That is the condition behind a whole class of confusing replies, and nothing
  previously named it.

Internal

• Every step in the build that reaches the network is retried. A stale package
  index had already published one release incomplete; the retry added then
  covered the quick half of the work and not the half that downloads gigabytes.
• The build cache no longer fills with entries nothing can use. Caches belonging
  to finished pull requests and past releases are reclaimed, and the cleaner no
  longer risks deleting a live cache in favour of an unusable one.
• The build workflows are themselves checked now, which they were not. Every job
  has a time limit, no two runs of the same work race each other, and the shell
  inside each step is linted — including the retry loops the point above depends
  on.
• A new pre-release check runs several model families and asserts that each
  produces a sane reply: that it answers the question, stops on its own, leaks no
  internal markers, and has its tool calls read back properly. Every
  model-specific fault fixed in the last four releases was reported by a user
  rather than found here, and each one would have passed the checks that existed.

---

Full commit list: v0.3.171-alpha...v0.3.172-alpha