fix(loop): graceful exit on max iterations with wrap-up nudge + forced text-only

[toanalien]

Summary

• Problem: AgentLoop could exhaust all 50 iterations when the LLM kept calling tools endlessly, resulting in status="failed" with no useful output returned to the user.
• Fix: Mirror the swarm worker's proven approach — inject a wrap-up nudge at 80% of the iteration budget and drop tool definitions on the last iteration to force a text response.
• Test: Added test_force_text_only_on_last_iteration validating that an LLM which always returns tool calls now produces status="success" instead of status="failed".

Changes

agent/src/agent/loop.py

1. Calculate wrap_up_at = max(1, int(self.max_iterations * 0.8)) before the main loop
2. At 80% of iteration budget, inject a [SYSTEM] user message nudging the LLM to stop calling tools and produce a final answer
3. On the last iteration, pass tools=None to stream_chat — the LLM cannot request tool calls and must produce text output, guaranteeing a final answer

agent/tests/test_agent_loop_terminal_state.py

• Added _StubLLMAlwaysToolCalls — LLM stub that returns compact tool calls until tools=None forces text
• Added test_force_text_only_on_last_iteration — verifies the forced text-only last iteration produces status="success" with content

How it works

The swarm worker (agent/src/swarm/worker.py) already had both mechanisms (lines 363-426). This PR brings the same pattern to AgentLoop:

Mechanism	Swarm Worker	AgentLoop (before)	AgentLoop (after)
Wrap-up nudge at 80%	Yes	No	Yes
Force text-only on last iter	Yes	No	Yes

Test plan

• Existing 3 tests in test_agent_loop_terminal_state.py pass unchanged
• New test_force_text_only_on_last_iteration passes
• ruff check passes
• python -m py_compile passes

[warren618]

Merged — thanks @toanalien! 🙏 Mirroring the swarm worker's wrap-up + forced-text-only pattern into AgentLoop is a nice consistency win, and the forced-text last iteration is verified to turn a tool-spamming run into a success with a final answer (confirmed against the real LLM e2e for max_iter).

Heads-up: I pushed a small follow-up on top (7fd8c56). The nudge fired at iteration == wrap_up_at unconditionally, and on tiny budgets (max_iter <= 2 → wrap_up_at == 1) it injected the [SYSTEM] wrap-up message on the first iteration. That displaced the active research-goal context as the most recent user message — breaking test_agent_goal_context — and could tell the agent to stop early while a goal was still continuing. The follow-up gates the nudge to 1 < iteration < max_iterations (never on the first iteration, never on the last, where the forced-text path already guarantees an answer). Full suite is green again. Thanks for the well-structured PR and tests!