WIP: stop persisting empty assistant messages, tolerate corrupt tool_use args

[birdayz]

What

Two related changes that together unwedge sessions hit by the empty-content / corrupt-tool-args failure modes around #116.

1. Agent loop skips persistence when len(resp.Message.Content) == 0. Empty assistant messages no longer land in session state, where Anthropic and most other providers reject every subsequent replay with messages.N.content: Field required.
2. ToolRequest.ArgumentsAsObject() collapses nil / invalid Arguments to an empty object, so request mappers can feed provider tool_use APIs without faceplanting on truncated streaming state. Bedrock and Gemini request mappers adopt it. Anthropic and OpenAI Compatible request mappers still need to switch over — that is the WIP part.

Why

#116 didn't introduce a new bug — it exposed a pre-existing one. Pre-#116, a stream cut off mid-tool-args left partial JSON in Content, so the message was non-empty and the persistence path was fine; the corruption surfaced differently (the JSON-parse-error wedge that #116 already covers). Post-#116, the partial tool_use is correctly dropped, but if it was the only block, Content is empty — and the agent loop appends that empty Message to session state, locking the session forever.

Sources of empty content observed in production:

• max_tokens hit before any content block was emitted
• the only content block was a partial tool_use that stream finalisation correctly dropped (fix: surface truncation instead of corrupting state when tool_use is cut off #116)
• refusal / safety filter with no text emitted

In all of these the FinishReason carries the truth (Length, ContentFilter, etc.) and the agent loop's existing terminal-reason handling fires. The bug was that persistence ran before the terminal-reason check, so the empty Message landed in session state regardless of how cleanly the loop terminated. Stayed wedged across pod restarts because the empty Message was already persisted.

The corrupt-tool-args side of the same problem: a session that already has truncated JSON in its history needs to be replayable. Erroring at json.Unmarshal time means the model can never recover, even though the paired tool_result carries the original parse error and the model has enough context to retry.

Implementation details

Persistence guard at the session-store boundary (agent/llmagent/llmagent.go). One-line check: skip the append when content is empty. The MessageEvent still fires (observers see what happened) and the FinishReason still propagates (terminal-reason handling fires below). Single chokepoint for "what gets persisted" — guard it there, not in every provider. Same shape as adk-go's AppendEvent honouring Event.Partial.

Provider response mappers do not error on empty content. Empty content is a real wire state; the FinishReason already encodes it correctly. Erroring at the mapper would swallow the Length / ContentFilter signal that the agent loop needs to terminate cleanly. Earlier iteration of this PR enforced the invariant at the mapper — reverted in favour of the session-store guard.

ToolRequest.ArgumentsAsObject() in llm/part.go. Decodes Arguments into a JSON object, falling back to an empty object on nil bytes or invalid JSON. The paired tool_result already carries the original parse error so the model has context to retry instead of every replay dying in mapping.

WIP scope. Anthropic and OpenAI Compatible request mappers still call json.Unmarshal directly on tool arguments and will fail on truncated state. The wedged-session reproducer in providers/anthropic/request_mapper_test.go is currently red and pins the next step.

References

• fix: surface truncation instead of corrupting state when tool_use is cut off #116 — surface truncation instead of corrupting state when tool_use is cut off
• docs/empty-content-persistence.md — design notes on where the invariant should live

[blacksmith-sh]

Found 2 test failures on Blacksmith runners:

Failures

Test	View Logs
github.com/redpanda-data/ai-sdk-go/plugins/retry/TestRequestMapper_EmptyToolArguments	View Logs
github.com/redpanda-data/ai-sdk-go/plugins/retry/TestRequestMapper_WedgedSessionReplay	View Logs