fix(agent): nudge once before the housekeeping empty-follow-up shortcut

[0xDevNinja]

What does this PR do?

Stops the fallback_prior_turn_content shortcut in agent/conversation_loop.py from silently ending the turn on the first empty completion after a housekeeping-only tool turn.

Today that shortcut fires before the standard post-tool empty-response nudge ever runs. That's the right trade when the model is genuinely done ("You're welcome!" + memory save → empty), but weak/quantized local models intermittently return a genuinely empty completion right after a housekeeping call because they choked, not because they finished — and they never get the retry every other empty-after-tools path gets. The turn ends with recycled narration and no error, and the session sits idle (#65600's background-review repro).

The fix defers the shortcut past the single nudge retry that already exists a few lines below:

• First empty follow-up after a housekeeping-only turn → standard nudge ("you returned an empty response, please continue"), preserving the housekeeping fallback across it.
• Second consecutive empty → the shortcut fires exactly as before and serves the prior-turn content. Two empties in a row is a much stronger "done" signal than one.

Cost in the genuinely-done case is one extra API call, which is what #7968's original suggested direction called for ("don't let the classification fully bypass retries"). No behavior change for substantive-tool turns, thinking-only responses (prefill path untouched), or the partial-stream recovery path. Unlike #64496, this isn't gated to kanban sessions — it covers the general case.

The nudge previously cleared _last_content_with_tools unconditionally; it now preserves it when the flag says housekeeping-only, so the second empty can still deliver the content instead of dead-ending in the empty-response exhaustion path. Clearing is kept for substantive narration (the #63860 stale-state fix is untouched — its regression test still passes).

Related Issue

Fixes #65600

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/conversation_loop.py: compute nudge availability (_prior_was_tool / _has_inline_thinking / retry flag) before the housekeeping fallback shortcut and gate the shortcut on not _nudge_available; in the nudge branch, preserve the fallback when it came from a housekeeping-only turn (still cleared for substantive narration).
• tests/run_agent/test_conversation_fallback_state.py:
  • new test_housekeeping_empty_follow_up_gets_one_nudge_before_fallback — housekeeping turn → empty (choke) → nudge → real recovery content becomes the final response. Fails on current main with exactly the reported log line and reason=fallback_prior_turn_content, api_calls=2.
  • updated test_housekeeping_only_turn_still_sets_fallback for the deferred contract: empty → nudge → empty again → fallback content served, reason=fallback_prior_turn_content, api_calls=3.

How to Test

1. uv run python -m pytest tests/run_agent/test_conversation_fallback_state.py -q → 3 passed.
2. Revert the agent/conversation_loop.py hunk and re-run → the new test fails with the issue's exact symptom (recycled narration, 2 API calls, fallback_prior_turn_content).
3. Full sibling suites: uv run python -m pytest tests/run_agent/ -q → 2052 passed. tests/agent/ has 127 pre-existing failures on clean main (verified identical with and without this change; the two turn-finalizer files that reference this exit reason pass).

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 14.2

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A

[tonydwb]

Code Review Summary

Verdict: Comment

Overview

• Agent fix: nudge once before the housekeeping empty-follow-up shortcut
• 137 additions, 29 deletions

Assessment

• Improves agent behavior by adding a nudge before the housekeeping shortcut
• No obvious security concerns
• Clean diff

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the focused regression fix. The premise remains present on current main: agent/conversation_loop.py:5141-5155 returns the housekeeping fallback before the post-tool empty-response nudge at agent/conversation_loop.py:5157-5222 can run. The proposed ordering preserves the existing role-safe nudge sequence and retains the fallback only for the immediately following empty completion.

The new regression scenario matches the reported failure mode, and GitHub reports a clean merge state with all required checks passing.

Automated hermes-sweeper review.

[0xDevNinja]

Rebased onto current main — the branch had gone conflicting. agent/conversation_loop.py merged cleanly (the nudge-vs-shortcut ordering is intact); the only conflict was in the test file, where both regression cases landed at a spot upstream had since cleared. Kept both, markers removed. Green: test_conversation_fallback_state.py 3 pass, and both housekeeping cases fail on main without the source change (verified by swapping in upstream's conversation_loop.py — the shortcut fires at 2 API calls instead of deferring one nudge).

[GottZ]

This was generated by AI during triage.

Summary

Six PRs address this empty-after-tool-response complex: #9432 handles substantive-turn fallback for #9400; #63888, #64080, and merged #64348 handle stale housekeeping state across a later substantive tool-only turn for #63860; #64496 covers kanban terminal-guard bypasses; and #65674 directly defers the #65600 housekeeping fallback until after one recovery nudge.

Related pull requests

• Bugfix: avoid stale tool-turn fallback on empty responses #9432 [closed] best fix — (+87/-17) — n/a: Restricts cached fallback content to housekeeping-only tool turns, preventing substantive-tool narration from ending an unfinished Bug: Empty response after tool calls silently abandons incomplete multi-step tasks #9400 task. It remains relevant as a closed historical implementation; the closure discussion says its core behavior was already on main, while contributor teknium1 attributes the broader empty-response nudge to fix: stale agent timeout, uv venv detection, empty response after tools (#9051, #8620, #9400) #10065.
• fix(conversation): clear stale housekeeping fallback on substantive tool-only turns #63888 [closed] fixes — (+146/-8) — n/a: Clears an older housekeeping fallback when a later substantive tool-only turn occurs, directly addressing [Bug]: Stale housekeeping fallback survives a later substantive tool-only turn #63860, although a COMMENTED review identified adjacent validation-retry and mute-state cases. It remains relevant because it was closed in favor of merged salvage fix(conversation_loop): invalidate stale housekeeping fallback on substantive tool-only turn #64348 with contributor credit preserved.
• fix(conversation_loop): invalidate stale housekeeping fallback on substantive tool-only turn #64080 [closed] fixes — (+201/-7) — n/a: Implements the same [Bug]: Stale housekeeping fallback survives a later substantive tool-only turn #63860 state invalidation by clearing stale fallback fields after a substantive tool-only turn while preserving legitimate housekeeping fallback. It remains relevant as a closed duplicate of fix(conversation): clear stale housekeeping fallback on substantive tool-only turns #63888, superseded by merged fix(conversation_loop): invalidate stale housekeeping fallback on substantive tool-only turn #64348.
• fix(conversation_loop): invalidate stale housekeeping fallback on substantive tool-only turn #64348 [merged] best fix — (+207/-8) — n/a: Merged salvage/reference implementation for [Bug]: Stale housekeeping fallback survives a later substantive tool-only turn #63860; it carries forward fix(conversation): clear stale housekeeping fallback on substantive tool-only turns #63888's invalidation, additionally clears _mute_post_response, and tests both recovery and housekeeping preservation.
• fix(kanban): gate the housekeeping-fallback shortcut on the terminal-tool guard #64496 partial — (+208/-2) — n/a: Gates both housekeeping fallback and partial-stream recovery while a kanban worker still owes kanban_complete/kanban_block, with regressions for both early-break paths. Consistent with its keep_open review, retain it for focused re-review; the visible diff addresses the review's partial-stream concern but remains a kanban-specific fix rather than the general [Bug]: Housekeeping-tool empty-follow-up shortcut wrongly ends the turn when a weak/quantized local model chokes instead of finishing #65600 recovery.
• fix(agent): nudge once before the housekeeping empty-follow-up shortcut #65674 best fix — (+188/-28) — n/a: Defers fallback_prior_turn_content until one existing post-tool nudge has been attempted, then preserves the fallback for a second consecutive empty response, directly matching [Bug]: Housekeeping-tool empty-follow-up shortcut wrongly ends the turn when a weak/quantized local model chokes instead of finishing #65600. Consistent with its high-salvageability keep_open review, the rebased diff retains focused recovery and fallback regressions, and the contributor reports three passing focused tests plus negative checks against main.

Duplicates

#63888 and #64080 are substantially duplicate implementations of the #63860 invalidation; merged #64348 salvages and supersedes both. #64496 and #65674 are not duplicates: the former enforces the kanban terminal protocol, while the latter provides general housekeeping-empty recovery.

Suggested consolidation

Keep #65674 open with a salvage path: return its rebased head to maintainer review as the focused #65600 fix, consistent with the visible keep_open verdict. Keep #64496 open for separate kanban-specific review; #63888 and #64080 are already closed in the explicit duplicate/salvage chain ending at merged #64348, and closed #9432 requires no further action.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I9400(["issue #9400 (closed)"])
    I65600(["issue #65600 (open)"])
    P65674["PR #65674 (open)"]
    P65674 -.->|partial| I9400
    P65674 -->|best fix| I65600
    class I9400 closed
    class I65600 open
    class P65674 open
    class P65674 best
    class P65674 target
    click I9400 "https://github.com/NousResearch/hermes-agent/issues/9400"
    click I65600 "https://github.com/NousResearch/hermes-agent/issues/65600"
    click P65674 "https://github.com/NousResearch/hermes-agent/pull/65674"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 6 pull requests and 3 issues in this complex. Each diff was read against this issue; Assessment working set: 64 kB of PR diffs, 38 kB of issue/PR text, 12 kB of discussion (15 comments), 10 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[0xDevNinja]

Rebased onto current main. agent/conversation_loop.py merged cleanly (nudge-vs-shortcut ordering intact). Test conflict was another interleave — upstream landed a new test_bare_tool_marker_is_not_reused_as_final_response (#78148) at the same insertion point as my two housekeeping cases. Rebuilt the file on upstream's clean version and re-appended my two tests. Green: 4 pass (upstream's 2 + mine); both housekeeping cases still fail on main without the source change (shortcut fires at 2 API calls), verified by swapping in upstream's conversation_loop.py.