dcdpr / jp / 27491629811

67%

push

github

feat(llm, anthropic): Add `claude-fable-5` and escalating-nudge retry (#745)

Claude Fable 5 runs with adaptive thinking permanently and rejects both
`thinking: disabled` and forced `tool_choice` while thinking is active.
The existing hard retry (disable thinking, then force the tool call) is
therefore unavailable for this model.

This commit introduces a `ForceStrategy` enum with two variants:
`DisableThinking` (the existing path for all other models) and
`EscalatingNudge` (for thinking-always-on models). A new
`supports_disabling_thinking()` method on `ModelDetails` gates which
path is taken, keyed off the new `thinking-always-on` feature flag.

When Fable 5 finishes a soft-forced request without calling the required
tool, `dispatch_force_retry` now delegates to `soft_force_retry`, which
appends the non-compliant assistant turn and re-issues the request with
a progressively firmer user message ("you MUST call this tool now"),
up to `SOFT_FORCE_MAX_RETRIES` times before accepting the response.

The `thinking: disabled` branch in `create_request` is also guarded by
`supports_disabling_thinking()`, so Fable 5 never sends a disabled
thinking config that the API would reject with a 400 error.

Unknown models now source token limits and structured-output capability
directly from the API response rather than leaving them entirely unset.
A reported value of `0` is treated as unspecified and left as `None`.

---------

Signed-off-by: Jean Mertz <git@jeanmertz.com>

44 of 106 new or added lines in 2 files covered. (41.51%)

34510 of 51928 relevant lines covered (66.46%)

221.17 hits per line