fix(tools): don't mask empty streaming tool-selection as %!w(<nil>)

[localai-bot]

Problem

Image turns with a reasoning model (e.g. gemma-4-e2b via LocalAI) fail with the meaningless error:

failed to select tool: failed to pick tool: tool selection failed: failed to make a streaming decision after 3 attempts: %!w(<nil>)

%!w(<nil>) is Go telling us fmt.Errorf("...: %w", …, lastErr) was called with a nil lastErr.

Root cause (reproduced against LocalAI v4.6.2 + gemma-4-e2b)

decisionWithStreaming's empty-content retry path (content == "" && len(toolCalls) == 0) continued without setting lastErr. When every attempt hit it, the loop exhausted with lastErr == nil and rendered %!w(<nil>).

This reliably happens on image turns: the vision model streams reasoning whose length exhausts the output-token budget (vision tokens crowd the context), finishing with finish_reason=length and no visible content and no tool call. The loop then retried three identical truncations and reported the nil wrap. Verified deterministically: with a constrained budget the model emits reasoning only, content_len=0, tools=0, finish=length.

Fix

Capture finish_reason from StreamEventDone, and in the empty-content branch:

• finish_reason=length → fail fast with an actionable error (retrying truncates identically), never a nil wrap.
• otherwise → keep retrying, but always record a real lastErr so the final error is a real cause.

The non-streaming decision() path already accepted empty content without this retry, so only the streaming path was affected.

Tests

decision_empty_content_test.go:

• TestDecisionWithStreamingTruncatedEmptyContent — reproduces the exact %!w(<nil>); asserts a meaningful, length-naming error and fail-fast (1 attempt).
• TestDecisionWithStreamingEmptyContentNoFinishReason — genuinely-empty responses still error out, never nil-wrapped.

Both fail on main (RED) and pass with the fix. Existing streaming tests unchanged.

🤖 Generated with Claude Code