Anthropic's newest models think whether or not they are asked to, so their thinking goes back to them within a turn as it does when an effort turns it on; and a provider error the retries outlast, or one inside a stream, is said in the provider's words, trimmed and without the key¶
Decision 805 · accepted · 2026-10-06 · issue #427
With no reasoning_effort the Anthropic adapter sends no thinking field (780), and it
handed a thinking block back only when the request turned thinking on (658). Anthropic's
newest models think with no thinking field at all, so for them every thinking block was
dropped: on the chunk's tip, the second call of a tool-use turn to claude-opus-5-5 went
out with the first call's tool_use and none of the thinking that led to it.
- Which: keep the blocks, never turn thinking off. Turning it off is not on offer for
most of these models (a 400 on Opus 5.5, Fable and Mythos, a form of its own on
Sonnet 5.5), and where it is, on Opus 5, the guide names what it costs: tool calls that
never run. Keeping the blocks is what keeps the model's quality: within a tool-use turn
each call goes on from the reasoning the last one ended with rather than starting
over. That is what the maintainer leaned to, and what an effort already got (658, 780).
- Which model thinks unasked is read from its name, as 780 reads the form:
Catalog.thinks_unasked?/1 says yes to Opus and Sonnet from 5 on, Fable and Mythos,
inside a gateway's renaming too (eu.anthropic.claude-opus-5). The model list has no
field for it: capabilities.thinking.types says which forms a model takes, and Opus
4.8 takes adaptive thinking without using it unasked. The adapter hands Anthropic's
blocks back when the request turns thinking on, as before, or when the model thinks
unasked; the request is otherwise as it was: no thinking field, no output_config,
the model's own default effort. An effort of none or off sends nothing either, so
a model that thinks anyway gets its thinking back the same way.
- A block the conversation no longer vouches for. Troupe's prompt changes between
turns (the task list as each turn begins, 792; the instruction files and brief read as
it begins, 798; the goal) and compaction rewrites the history, so on an account the
check is enforced for, a block from before such a change is refused. That 400 is
answered by sending the call once more with no thinking in it, the API's own way on and
what every one of these requests was before this decision; a second refusal is the
turn's error, in the provider's words. With an effort configured it is the same. The
cost is one refused call and the earlier reasoning, where without it the turn failed.
- A failure the retries outlast says what the provider said. with_retries carried
only the status, so a 5xx or a 429 past its retries read gave up after retrying:
{:http_status, 503}. Both adapters now give the retry the status with the provider's
message, read and cleaned as an error response's is (791), {:http_status, status,
detail}; describe_error says the last one as it would have been said unretried
(gave up after retrying: the provider answered 529 (Overloaded)), and a rate limit
the backoff outlasted names what the provider said about it. A transport failure is
said as before.
- An error inside an Anthropic stream ({:api_error, message}, an error event after
a 200) is trimmed and has the key masked, as an error response's message is since 791.
Governs: apps/troupe_core/lib/troupe/llm/providers/anthropic.ex, apps/troupe_core/lib/troupe/llm/providers/openai.ex, apps/troupe_core/lib/troupe/llm/provider.ex, apps/troupe_core/lib/troupe/llm/catalog.ex