fix(agent): nudge past progress-placeholder responses after tool call…

[ithelpm]

…s (#42503)

After executing tool calls, some models (observed cross-provider, most often local/smaller models) close the turn with a short progress note ("Working on it...", "I'll now update the file…") instead of continuing to the next step or delivering the result. The conversation loop treats any non-empty no-tool-call response as the final answer, so the task silently ends unfinished — no error, no hint (#42503).

The existing intermediate-ack continuation deliberately skips turns that already ran tools, and the post-tool EMPTY-response nudge only covers blank content, so this failure mode fell between the two recoveries.

Fix, mirroring the empty-response nudge:

• detect progress placeholders after a tool round (short, non-question, explicit progress phrase OR future commitment with an open ending)
• nudge the model to continue, capped at 2 per tool round
• if the nudges are exhausted, keep the note but mark the turn post_tool_placeholder_response so the turn-completion explainer appends a visible "Incomplete turn" footer instead of ending silently
• flag the synthetic nudge pair as ephemeral scaffolding so it is popped before the real final response and never persisted

What does this PR do?

Related Issue

Fixes #

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

How to Test

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:

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
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

For New Skills

• This skill is broadly useful to most users (if bundled) — see Contributing Guide
• SKILL.md follows the standard format (frontmatter, trigger conditions, steps, pitfalls)
• No external dependencies that aren't already available (prefer stdlib, curl, existing Hermes tools)
• I've tested the skill end-to-end: hermes --toolsets skills -q "Use the X skill to do Y"

Screenshots / Logs

[teknium1]

Thanks for targeting the post-tool placeholder gap from #42503; current main still takes non-empty no-tool replies through the normal final-response path (agent/conversation_loop.py:4941-5272), and the existing intent-ack tests explicitly exclude histories where a tool already ran (tests/agent/test_intent_ack_continuation.py:166-174).

Problems

• agent/conversation_loop.py:4995-5007 adds synthetic scaffold rows, but the cleanup at :5023-5033 removes only a trailing suffix. If the nudge produces another tool call, those rows become interior. agent/turn_finalizer.py:175 saves trajectories before session persistence, and run_agent.py:2034 converts the unfiltered messages, so save_trajectories=True records the supposedly ephemeral rows.

Suggested changes

• Filter/remove this scaffold from trajectory output as well as session persistence.
• Add the tool-call-after-nudge regression; the current persistence test patches _persist_session, so it does not cover the buried-scaffold case.

Automated hermes-sweeper review.

[teknium1]

Blocking: if this nudge leads to another tool call, this flagged pair becomes interior rather than trailing, so the cleanup at lines 5023-5033 cannot remove it. The finalizer saves trajectories before session persistence filtering, so save_trajectories=True records these supposedly ephemeral rows. Filter/remove them from trajectory output and add a tool-call-after-nudge regression.

[ithelpm]

Thanks for the catch - the trailing-suffix pop indeed missed the case where the nudge succeeds via another tool call, leaving the synthetic pair buried under the new tool turns and leaking into save_trajectories=True output.

Addressed in the follow-up commit (branch also rebased onto current main):

• Tool-call branch pop (agent/conversation_loop.py): the _post_tool_placeholder_synthetic pair is now popped in the tool-call path, before the new assistant tool-call turn is appended, so the rows can never become interior in the first place and also stop being replayed to the model in the live context.
• Structural trajectory sweep (agent/agent_runtime_helpers.py::convert_to_trajectory_format): all _EPHEMERAL_SCAFFOLDING_FLAGS rows are filtered at the top of the conversion, mirroring the per-message filter the session-store flush already applies - same philosophy as Rotation compression: _db_persisted marker copied via messages[i].copy() skips bulk flush to child session (cached gateway regression since #50372) #57491's terminal sweep, so trajectory hygiene no longer depends on every loop path cleaning up positionally. This also covers the pre-existing empty-response/prefill scaffolds if they ever get buried the same way.
• Regression tests: test_tool_call_after_nudge_leaves_no_buried_scaffolding runs the tools -> placeholder -> nudge -> tools -> answer sequence end-to-end and asserts no synthetic rows survive in the message list; test_trajectory_conversion_drops_buried_scaffolding feeds an interior scaffold pair directly into the trajectory conversion (no _persist_session patching involved) and asserts it never reaches the output.

[ithelpm]

Self-review addendum: the tool-call-path cleanup had an interleaving gap. With tool round -> thinking-only response (prefill row) -> progress placeholder (nudge pair) -> tool calls, the stack is [prefill, placeholder, nudge] and two sequential single-flag pop loops strand the prefill row beneath the popped nudge pair - it then replays as a fake assistant turn (and a consecutive-assistant pair) in the live context. Reproduced empirically before fixing. This gap only exists because this PR's nudge pair can sit on top of a prefill row, so it is fixed here rather than in a separate PR.

Fixed in 2d541ee by merging the pops into one loop over both flags, mirroring the combined terminal pop on the final-response path (_had_prefill is only set when a prefill row is actually popped, so the retry-counter reset semantics are unchanged). Added test_interleaved_prefill_and_placeholder_scaffolds_all_popped covering the exact sequence.

[yingliang-zhang]

Draft Comment for PR #57610

Status: DRAFT — needs user review before posting

---

Thanks for this PR — it addresses a real gap (#42503) that we've also been working on locally.

We have a complementary implementation that covers an adjacent failure mode not addressed by the progress-placeholder detection in this PR:

Signature A — lost tool-call preamble. When the model emits a short text ending with : or ： (e.g. "接下来读取文件：") with finish_reason=stop and no tool calls, right after a tool round completed. This is distinct from the progress-placeholder case (#57610's scope) — it's a truncated tool-call narration where the tool_calls array was lost. The upstream _dropped_toolcall_nudge path only covers finish_reason=tool_calls with an empty tool_calls array, not finish_reason=stop with a text preamble.

CJK intent marker corpus. Our implementation includes ~30 CJK intent markers (继续/接下来/下一步/然后/开始/...) with done/wait/question exclusion lists that were production-hardened over several weeks of testing with K3 and GLM models. These could extend #57610's English-only phrase detection.

What we'd like to contribute:

1. The Signature A detection (colon-preamble + tool-round gate) as a follow-up to this PR
2. The CJK marker corpus + exclusion lists as a locale extension
3. A force_display parameter on _emit_interim_assistant_message (fixes the streaming-buffer overwrite issue from the closed fix: persist and emit agent response during verification stop loop #62676 — ensures verify-on-stop candidates survive as permanent UI bubbles on messaging gateways)

What we won't do:

• Submit a competing PR — we'd like to coordinate with this PR as the merge vehicle
• Scope-creep this PR — we'll file as a separate follow-up after merge

Would this sequencing work for you? Happy to share the code or more details if useful.