fix(tools): stop leaking raw tool results and model content through the status callback

[localai-bot]

Problem

The status callback is documented as short human one-liners, but ExecuteTools pushed raw payloads through it, flooding consumer chat UIs:

• every raw tool result (statusCallback(execResult.result)) — dante-desktop rendered each tool result twice: once as an unbounded, unstyled "status" wall of text directly in the chat, once as its styled collapsed tool block. notetaker carries a looksLikeToolResultStatus() heuristic filter working around this exact leak.
• the no-tool reply/reasoning in toolSelection — an exact duplicate of the reasoningCallback call on the next line, and can be a full-length model reply.
• selectedToolFragment.LastMessage().Content — empty in the normal path (toolSelection strips content), pure noise.
• the "Selected N tool(s)" counter line.

All four are removed. Raw results keep flowing on their dedicated channel (WithToolCallResultCallback), reasoning on WithReasoningCallback.

New: WithStepContentCallback

The assistant content accompanying a tool selection ("I'll search for X now…", decisionResult.message) was previously discarded outright — never in history, never in a callback (only visible as live stream deltas). It now has a dedicated channel: WithStepContentCallback fires at the step boundary, before the selected tools execute, so consumers can render the commentary chronologically (text → tool result → … → final answer).

It never fires for empty content, for the guided (forceReasoning) path (which carries no message), or for the turn's final reply.

Tests

status_channel_test.go (2 new specs): the status channel receives neither raw results nor step content; step content arrives on its own channel ordered before the tool result; silent steps don't fire.

Unit suite green: 73/73 (--label-filter='!e2e'). Note: agent_test.go "spawn agent in background and return ID" is flaky on clean main too (parent/background sub-agent race on the shared mock queue) — unrelated, worth a separate fix.

Consumers

• nib / dante-desktop: follow-up PRs wire OnStepContent through to the chat UI (interleaved assistant bubbles) — this fixes the raw-dump + ordering UX in Dante.
• notetaker: after bumping, can delete the looksLikeToolResultStatus workaround (internal/notary/agent/chat.go).
• Same-class leaks exist in extractors.go:77 / reviewer.go:59 (ExtractStructure/ContentReview flows) — left out of scope here.

🤖 Generated with Claude Code