dcdpr / jp / 27565777572

67%

push

github

feat(llm, cli): Surface Anthropic model refusals to the user (#752)

Add a `FinishReason::Refused { category, explanation }` variant to the
`FinishReason` enum in `jp_llm`. This represents a terminal state where
a safety classifier declined the request rather than the model failing.

On the Anthropic provider side, `stop_reason: "refusal"` is now mapped
to this new variant, pulling `category` and `explanation` out of the
`stop_details` field when present. Both fields are `Option<String>` to
handle uncategorized refusals gracefully.

In the CLI coordinator, the `finish_notice` function now handles this
variant, producing a human-readable notice like:

    The model declined this request (cyber): This request was declined.

When neither category nor explanation is present, the notice falls back
to the simpler form:

    The model declined this request.

---------

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

46 of 46 new or added lines in 4 files covered. (100.0%)

34650 of 52073 relevant lines covered (66.54%)

208.8 hits per line