feat(zarvis): steer the agent mid-turn between tool steps

[edwin-zvs]

Problem

In a zarvis session, a prompt typed while a turn is running is enqueued but only consumed after the entire turn finishes — so you can't redirect the agent between tool calls while it's working on a multi-step task.

Change

At each step boundary of a turn (after that step's tool results are appended to the conversation, before the next model request), drain anything the user typed during the step and fold it in as a User message. The model then incorporates the steering on its very next step — no need to wait for the whole turn to end.

assistant(tool_calls) → tool_result(s) → user(your steering) → next model step

This is additive steering, not an interrupt — in-flight model/tool work is never aborted. Stop / Interrupt (Ctrl-C) remain the way to abort. Input typed during the silent provider stream / tool execution is back-filled into the chat history (PTY echo + Message event) exactly like the outer loop's queued-input path, so it shows up where you'd expect.

Provider-safe: injecting a user message after tool results is the standard interjection pattern; the Anthropic adapter already emits consecutive user-role messages for multi-tool turns, and OpenAI/Ollama accept user-after-tool.

Notes

• Always-on (no flag). Steering only applies when the turn continues to another step; a turn that ends with a text-only response still hands a queued message to the next turn as before.
• Text-only steering during a long single tool call is picked up after that tool finishes (the next step boundary); to abort the tool itself, use Interrupt.

Tests

• New drain_steering helper + unit tests: empty → none; multi-entry → newline-joined and queue emptied; and the coalesced multi-submit case (enqueue_line combines consecutive submits, whitespace ignored).
• Full agentd-adapter-zarvis suite (126) green; --locked workspace build clean.

🤖 Generated with Claude Code