fix: surface exhausted empty model responses

[taivu1998]

Summary

• add explicit empty_response_exhausted metadata when the agent exhausts empty-response retries
• keep the legacy final_response == "(empty)" sentinel for compatibility while preventing hooks, external memory sync, background reviews, CLI titles, and TTS from treating it as a real assistant answer
• show a clear classic CLI message instead of rendering the raw (empty) sentinel
• add regressions for replayed tool history, post-tool exhaustion cleanup, hook/memory skipping, successful nudge recovery, and CLI display normalization

Root Cause

Issue #17248 reports sessions that appear to complete with a literal (empty) answer after tool calls or empty model responses. The agent already had retry and cleanup scaffolding, but the terminal sentinel was indistinguishable from meaningful assistant content to downstream surfaces such as plugin hooks, memory sync, title generation, TTS, and CLI display.

Fix

This PR preserves the existing sentinel return value for compatibility, but marks the exhausted-empty path with explicit result metadata:

• empty_response_exhausted: true
• error_code: "empty_response_exhausted"
• a short user-facing error string

The agent then gates post-response integrations on a meaningful-final-response check, while the CLI normalizes exhausted sentinel results into a helpful visible error message.

Validation

• HERMES_HOME=/private/tmp/hermes-agent-17248-publish-home-focused ./scripts/run_tests.sh tests/run_agent/test_run_agent.py -k "empty_response or replayed_tool_history" - 8 passed
• HERMES_HOME=/private/tmp/hermes-agent-17248-publish-home-adjacent ./scripts/run_tests.sh tests/run_agent/test_empty_response_recovery_persistence.py tests/run_agent/test_message_sequence_repair.py tests/cli/test_empty_response_display.py - 18 passed
• HERMES_HOME=/private/tmp/hermes-agent-17248-publish-home-full ./scripts/run_tests.sh tests/run_agent/test_run_agent.py - 326 passed
• uv run ruff check run_agent.py cli.py tests/run_agent/test_run_agent.py tests/cli/test_empty_response_display.py - passed
• git diff --check - passed

Fixes #17248

[teknium1]

Thanks for addressing the exhausted-empty response path. The user-visible raw-sentinel symptom has since been superseded on current main: agent/turn_finalizer.py:299-326 replaces (empty) with an actionable explanation, with end-to-end coverage in tests/run_agent/test_turn_completion_explainer.py:142-163.

Problems

• The PR edits the old post-loop implementation in run_agent.py; current main runs these post-response effects from agent/turn_finalizer.py:337-480, so the patch needs a non-mechanical port.
• Current main still passes the generated explanation to hooks and external-memory sync (agent/turn_finalizer.py:343-378, 462-478), and the CLI still permits auto-title/TTS (cli.py:12542-12560, 12705-12708). This means the side-effect-suppression objective remains relevant, but it must be keyed from the turn-exit reason/metadata rather than the old raw sentinel.

Suggested changes

• Preserve current main's visible explanation, add exhausted-empty metadata from empty_response_exhausted, and use it to gate the affected integrations and CLI title/TTS paths.

Automated hermes-sweeper review.

[teknium1]

Current main's live post-loop path is now agent/turn_finalizer.py, and it replaces (empty) with a user-facing explanation before hooks run. Please port this predicate there using turn_exit_reason == "empty_response_exhausted"; preserving the old raw-sentinel condition would not cover the current execution path.