feat(agent): configurable content_response_policy for content-without-tool-call responses (#3992)

[Code-weaver1]

HUMAN:

• A human has tested these changes.

AGENT:

Investigated, implemented, and unit-tested by an AI agent (Claude Code) operating under
Solomon Okello's direction; every claim below is backed by the included tests.

---

Why

#3992: ResponseDispatchMixin dispatches asymmetrically — an EMPTY response gets a
corrective nudge and the loop continues, but a CONTENT response (prose, no tool call)
sets execution_status = FINISHED unconditionally. Weaker/local models narrate a step
in prose before emitting its tool call, so they get silently terminated mid-task with
exit 0. Three earlier attempts to change the default (#1303, #1304, #1385) didn't merge —
plain content is a legitimate completion signal for strong models. So this PR follows
the issue's second suggested direction: make the policy configurable, default
unchanged.

Summary

• New AgentBase field content_response_policy: Literal["finish", "nudge"], default
  "finish" — zero behavior change unless opted in.
• Under "nudge": the prose message is emitted (preserved in history), a corrective
  nudge is sent (pointing at the finish tool as the explicit completion path), and the
  loop continues. Runaway prose loops are bounded by conversation-level stuck detection —
  the same mechanism that already bounds the EMPTY-response nudge (per the existing
  docstring on _send_corrective_nudge).
• _send_corrective_nudge gains an optional text kwarg so the content-case feedback
  can differ from the empty-case feedback (the model did produce a message; telling it
  otherwise would be false).

Issue Number

Fixes #3992

How to Test

uv run pytest tests/sdk/agent/test_response_dispatch.py -q

Three new tests:

• test_content_response_nudge_policy_continues_loop — prose turn under "nudge": not
  FINISHED, agent message preserved, nudge follows, nudge text points at finish
• test_content_response_default_policy_is_finish — guards the no-behavior-change
  contract
• test_finish_tool_still_finishes_under_nudge_policy — explicit completion still works

The pre-existing test_content_response_sets_finished continues to pass unchanged,
which is the other half of the compatibility guarantee.

[all-hands-bot]

[Automatic Post]: I have assigned @neubig as a reviewer based on the repository MAINTAINERS file. Thanks in advance for the help!

This comment was posted by an AI agent (OpenHands) on behalf of the user.