[Bug]: transform_llm_output plugin output silently dropped in CLI streaming mode

[kweiner]

Bug Description

Content appended by a transform_llm_output plugin hook is silently dropped when Hermes CLI streams a response token-by-token. The appended content is present in the final response text and stored in conversation history, but the user never sees it on screen.

Steps to Reproduce

1. Install a plugin that implements transform_llm_output and appends a suffix to the response (e.g. a disclaimer, citation, or additional content).
2. Run hermes in interactive CLI mode (streaming enabled by default).
3. Send any message that triggers the plugin to append content.
4. Observe that only the streamed response is shown — the appended content is missing from the terminal output.

To confirm the content is actually there: check conversation history or run in a non-streaming surface (Dashboard, Desktop, TUI). The appended content appears correctly on all non-streaming surfaces.

Expected Behavior

After streaming completes and transform_llm_output fires, any content the plugin appended to the response should be printed to the terminal.

Actual Behavior

The CLI enters the already_streamed branch after streaming and skips all output with pass. Since transform_llm_output fires after the stream is closed, the plugin's appended content is never displayed.

Relevant code in cli.py:

elif already_streamed:
    # Response was already streamed token-by-token with box framing;
    # _flush_stream() already closed the box. Skip Rich Panel.
    pass  # ← appended content is silently dropped here

Root Cause Analysis

The sequence is:

1. CLI streams response token-by-token and sets already_streamed = True.
2. After the tool loop, finalize_turn() fires transform_llm_output hooks.
3. A plugin appends content, returning the full new response string.
4. finalize_turn() returns the transformed response in the result dict.
5. Back in the CLI, the already_streamed branch executes pass — the transformed response with appended content is discarded from the display path.

The fix requires two small changes:

agent/turn_finalizer.py — capture the pre-transform response so the CLI knows what was already streamed:

_pre_transform_response = None

# ... inside the transform loop ...
for _hook_result in _transform_results:
    if isinstance(_hook_result, str) and _hook_result:
        _pre_transform_response = final_response   # save before overwrite
        final_response = _hook_result
        _response_transformed = True
        break

# ... in the result dict ...
"response_transformed": _response_transformed,
"pre_transform_response": _pre_transform_response,   # new field

cli.py — in the already_streamed branch, print only the appended suffix:

elif already_streamed:
    if result and result.get("response_transformed"):
        _pre = result.get("pre_transform_response") or ""
        _appended = response[len(_pre):]
        if _appended.strip():
            _cprint(_appended)

A PR with this fix is ready: https://github.com/kweiner/hermes-agent/tree/fix/transform-llm-output-streaming

Debug Report

N/A — this is a code-level bug with a clear root cause, not a configuration or environment issue.

Operating System

Ubuntu 24.04.4 LTS

Python Version

3.11.15

Hermes Version

0.18.0

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR

[GottZ]

This was generated by AI during triage.

Summary

One PR, #57270, directly addresses #57269. Its diff preserves the pre-transform response and displays either only an appended suffix or the full labeled replacement after CLI streaming, eliminating the reported silent drop.

Related pull requests

• fix: transform_llm_output appended content silently dropped in CLI streaming mode #57270 best fix — (+100/-1) — fixes: Tracks the pre-transform response in finalize_turn(), then renders an appended suffix without duplicating streamed content or renders an arbitrary replacement in full; regression tests cover both paths. The diff also addresses the contributor's keep_open review concerns by supporting replacement transforms, forwarding both model and provider, and adding focused CLI and gateway tests.

Suggested consolidation

Merge #57270 — it is the only candidate, directly closes #57269, fixes both append-only and replacement transform delivery after streaming, and addresses the documented contributor review concerns; there are no duplicate PRs to close.

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
    I57269(["issue #57269 (open)"])
    P57270["PR #57270 (open)"]
    P57270 -->|best fix| I57269
    class I57269 open
    class I57269 target
    class P57270 open
    class P57270 best
    click I57269 "https://github.com/NousResearch/hermes-agent/issues/57269"
    click P57270 "https://github.com/NousResearch/hermes-agent/pull/57270"

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 or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 1 pull request and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 8 kB of PR diffs, 7 kB of issue/PR text, 2 kB of discussion (2 comments), 2 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.