Recognizes the DeepSeek/OpenAI-compatible relay wording
max_tokens (98304) exceeds model's maximum output tokens (65536)
in both parse_available_output_tokens_from_error (returns the cap) and
is_output_cap_error (keeps the 400 out of the compression death-loop).
Salvaged from PR #72283; the conversation_loop early-clamp block was
dropped in favor of routing through the existing output-cap handler
(follow-up commit).
Fixes the retry loop that spins forever when a vLLM server rejects a
request for having a max_tokens too big for what is left of the context
window.
The catch is that vLLM does not tell you how big your prompt actually is
in that situation. It works the number backwards from the constraint it
just failed, so you get:
"requested 65536 output tokens and your prompt contains at least
36865 input tokens, for a total of at least 102401 tokens"
That 36865 is just window + 1 - requested, and the total is always
exactly window + 1. Subtracting it from the window hands back
requested - 1 every single time, whatever the real prompt size is.
parse_available_output_tokens_from_error believed it and returned
requested - 1. conversation_loop then takes off its 64 token safety
margin and retries, which walks the cap down 65 tokens at a time while
the reported input walks up by the same 65:
65536 -> 65471 -> 65406 -> 65341
Three attempts is the default budget, so the session gives up with
"Context length exceeded" having closed 195 tokens of a roughly 28000
token gap. Compression cannot save it either, because the input was
never the problem, which is why the compressor keeps refusing with
"summary would have GROWN".
This is also what is behind the unexplained "input-token drift" in
issue #61761. The input is not drifting. It is a derived number, and it
moves because we moved max_tokens.
So when that shape shows up (the "at least" wording, plus a budget that
works out to exactly requested - 1), halve the requested cap instead. It
is still guaranteed to sit under whatever was just rejected, and it
converges on the first retry: 65536 -> 32768, which next to a real 36865
token prompt comes to 69633 against a 102400 window.
Nothing else moves. A measured input is still trusted, and a genuine
input overflow still returns None so the caller falls through to
compression the way it always did.
The existing test asserted the bogus 65535, so it is updated. Added
tests for the measured input path, and for the retry actually
converging.
vLLM (and other OpenAI-compatible servers) report context overflow with
both the window and the prompt in tokens:
"This model's maximum context length is 131072 tokens. However, you
requested 65536 output tokens and your prompt contains at least 65537
input tokens, for a total of at least 131073 tokens."
parse_available_output_tokens_from_error() already classified this as an
output-cap error (the "requested N output tokens" gate), but none of the
extraction patterns matched the "prompt contains [at least] N input
tokens" phrasing, so it returned None. The recovery path then
misclassified the failure as prompt-too-long and looped through
compression — which frees little while each retry keeps requesting the
same oversized max_tokens — terminating in "cannot compress further"
even though simply lowering the output cap would have succeeded.
Add an extraction branch for the token-based phrasing: available output
= window - reported input. When the input alone is at or over the
window it still returns None, so the caller correctly falls through to
compression.
Relates to #43547.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
An over-cap model.max_tokens produces a provider 400 that mentions
max_tokens, which trips _CONTEXT_OVERFLOW_PATTERNS and is classified as
context_overflow. On providers whose wording isn't recognized by
parse_available_output_tokens_from_error() (e.g. DashScope/Qwen:
"Range of max_tokens should be [1, 65536]") the smart-retry is skipped
and the error falls into the compression fallback, which re-sends the
same oversized max_tokens, fails identically, and loops until
"cannot compress further" on a tiny conversation (#55546).
Root-cause fix for the whole class, not just DashScope:
- parse_available_output_tokens_from_error(): recognize the DashScope
"Range of max_tokens should be [1, N]" form and return N (smart-retry
then caps output and retries WITHOUT compressing).
- new is_output_cap_error(): broader yes/no gate for output-cap 400s.
In the loop, when the error is output-cap-shaped but unparseable, fail
fast with an actionable message (lower model.max_tokens) instead of
routing into compression. Mirrors the existing GPT-5 max_tokens guard.
Real input overflows and GPT-5 unsupported-param 400s are unchanged.
Add TestParseCharBasedOutputCap for the LM Studio / llama.cpp phrasing
(context in tokens, prompt in characters): the reported error resolves to
the available output budget, the retried cap plus the estimated input
stays inside the window, and a prompt larger than the window falls through
to None so the prompt-too-long/compression path still owns that case.
Two isolated reliability fixes:
- chat_completion_helpers: raise on a zero-chunk stream (no finish_reason,
no content/reasoning/tool_calls) so retry handles it instead of
fabricating a successful empty turn.
- model_metadata: parse the OpenRouter/Nous output-cap error phrasing
("maximum context length is N ... (A of text input, B of tool input,
C in the output)") so parse_available_output_tokens_from_error returns
a real cap and the caller stops looping on it.
Salvaged from #40405 (@ashishpatel26) — took the two stream/error-parsing
fixes. The PR also bundled compression-state changes (on_session_start
clearing _previous_summary; cron session-id prefix preservation, #38788);
those touch the compression hot path and are split out for separate review.
Co-authored-by: ashishpatel26 <ashishpatel26@users.noreply.github.com>
ekinnee