fix(tts): strip leading reasoning labels before speech by KoNit-K · Pull Request #107047 · NousResearch/hermes-agent

KoNit-K · 2026-09-10T02:32:49Z

What does this PR do?

Gateway auto-TTS (and the shared TTS tool path) already strips tagged <think> blocks (#34213), but a reply that starts with a visible section label such as Reasoning: or thinking： still reaches prepare_spoken_text() and gets spoken aloud (observed on Telegram).

This PR adds an allowlisted leading-label strip in tools/tts_text_normalize.py on the shared prepare_spoken_text() pipeline, immediately after strip_nonspoken_blocks and before markdown cleanup. Only reasoning / thinking / analysis / / / `` at the very start of the text, followed by :, `：`, or a newline, are removed. Mid-prose uses and non-allowlisted headers stay spoken.

Related Issue

Fixes #107044

Type of Change

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

Changes Made

tools/tts_text_normalize.py — add and wire it into
tests/tools/test_tts_prepare_spoken.py — focused RED/GREEN coverage for leading-label strip + fail-open mid-prose / non-allowlisted labels

How to Test

scripts/run_tests.sh on the files in Changes Made — focused verification
scripts/run_tests.sh tests/tools/test_tts_prepare_spoken.py -q
scripts/run_tests.sh tests/tools/test_tts_text_normalize.py -q
Result: 19 passed

Evidence

BEFORE: the focused probe was RED on origin/main before this change
AFTER: - [x] scripts/run_tests.sh on the files in Changes Made — focused verification — -> 19 passed
CONTROL: neighboring paths listed in How to Test still pass

Checklist

Code

✅ I've read the Contributing Guide
✅ My commit messages follow Conventional Commits
✅ I searched for existing PRs to make sure this isn't a duplicate
✅ My PR contains only changes related to this fix
✅ I've run relevant tests locally (see How to Test)
✅ I've added tests for my changes
✅ I've tested on my platform: macOS

Documentation & Housekeeping

✅ Documentation update: N/A unless noted in Changes Made
✅ cli-config.yaml.example: N/A
✅ CONTRIBUTING.md or AGENTS.md: N/A
✅ Cross-platform impact considered
✅ Tool descriptions/schemas: N/A

Gateway auto-TTS was reading visible section labels like Reasoning: or thinking： aloud. Strip allowlisted labels only at text start so mid-prose uses stay spoken. Co-authored-by: Cursor <cursoragent@cursor.com>

kvnloo · 2026-09-10T04:00:34Z

Exact-head comment on #107047 head 6f9eb4e. KEEP strip_leading_reasoning_labels on the shared prepare_spoken_text() pipeline (after strip_nonspoken_blocks, before markdown) so gateway auto-TTS no longer speaks a visible leading Reasoning: / thinking： / Analysis: / 分析： / 推理： / 思考： section label (#107044; <think> tags already handled by #34213). KEEP the allowlist + start-anchored regex (^\s*(?:reasoning|thinking|analysis|推理|思考|分析)[ \t]*(?:[:：]|\r?\n), IGNORECASE) and count=1 so mid-prose (The reasoning is clear.) and non-allowlisted headers (Summary:) stay spoken. KEEP the RED/GREEN cases in tests/tools/test_tts_prepare_spoken.py (leading EN/fullwidth/CN labels, newline form, mid-prose fail-open, Summary preserved). CHECK: label-only input (Reasoning:) collapses to empty spoken text — confirm TTS callers treat empty as skip, not a broken utterance. CHECK: a second leading allowlisted label after the first strip is intentionally left (count=1) — fine unless models emit stacked Reasoning:/Analysis: openers. CHECK: Chinese labels are case-sensitive codepoints (no IGNORECASE analogue) — OK for CJK. CHECK: no separate gateway adapter path that bypasses prepare_spoken_text (Telegram path author cites should already share this cleaner). Author KoNit-K not kvnloo. 0 reviews / 0 comments. effort: factory-easy.

KoNit-K · 2026-09-10T06:19:30Z

Checked the KEEP/CHECK notes against 6f9eb4e1 — no code change needed.

Empty spoken text (Reasoning: / thinking： / 分析：): prepare_spoken_text() collapses label-only input to "". Callers already treat that as skip, not a broken utterance:

text_to_speech_tool: returns tool_error("Text is empty after TTS cleanup") before any provider call
gateway auto-TTS (_synthesize_auto_tts): if not speech_text: raise ValueError(...), caught and logged, no playback
run_voice._send_voice_reply, CLI _speak_whole_file / _voice_speak_response, and streaming TTS (_strip_markdown_for_tts, which is prepare_spoken_text): all return early on empty cleaned text

Stacked labels (count=1): confirmed leftover (Reasoning:\nAnalysis: leftover → Analysis: leftover). Leaving the second opener is intentional; no evidence models emit stacked allowlisted headers as the only spoken content. Will loop only if that shows up in the wild.

CJK labels: agree — 推理 / 思考 / 分析 are exact codepoints; re.IGNORECASE is only meaningful for the Latin allowlist.

Gateway / Telegram path: Telegram does not override prepare_tts_text. Auto-TTS goes BasePlatformAdapter._synthesize_auto_tts → prepare_tts_text() → shared prepare_spoken_text(). The voice-reply and streaming-TTS paths use _strip_markdown_for_tts, which is the same pipeline (prepare_spoken_text(..., max_chars=None)). No adapter bypass.

KEEP items (shared pipeline placement, start-anchored allowlist, count=1, existing RED/GREEN cases) stay as-is. Focused tests: tests/tools/test_tts_prepare_spoken.py + test_tts_text_normalize.py → 21 passed.

kvnloo · 2026-09-10T14:36:04Z

Thanks — ACK on the no-code path. Agreed KEEP/CHECK covered; empty spoken text / reasoning-label handling looks fine at head 6f9eb4e without further change from me.

KoNit-K · 2026-09-10T16:58:57Z

Duplicate comparison vs #34213 (triage mark)

Classifier treated this as a duplicate because a comment said <think> tags are “already handled by #34213”. That is not a close instruction. Comparison:

	#34213 (issue, closed) + landed path	this PR #107047
Kind	Issue: strip tagged reasoning/`<think>` from TTS	Bugfix for visible leading labels (#107044)
Implementing PR	Landed in #73513 as `strip_nonspoken_blocks`	this branch
Files	`tools/tts_text_normalize.py` tagged-block regex only	same file, new `strip_leading_reasoning_labels` + tests
Input	`<think>…</think>` / unterminated `<think>` / verifier footer	`Reasoning:` / `thinking：` / `Analysis:` / `分析：` / `推理：` / `思考：` at start only
Fail-open	N/A for plain labels (they are not tags)	mid-prose (`The reasoning is clear.`) and non-allowlisted `Summary:` stay spoken

Not the same change. #34213 does not remove a visible leading section label. On current prepare_spoken_text() without this PR, Reasoning: hello is still spoken.

Independent delta — keep on #107047

Shared-pipeline step after strip_nonspoken_blocks, before markdown
Start-anchored allowlist + count=1
RED/GREEN cases in tests/tools/test_tts_prepare_spoken.py

Do not fold this into #34213 (already closed/landed) or into #77079 (open; expands tag variants only — no leading-label strip). Closed sibling #107048 targeted the same #107044 label case and is not live; this PR is the remaining candidate.

No code change this round. Focused tests: test_tts_prepare_spoken.py + test_tts_text_normalize.py → 21 passed. Leaving open.

KoNit-K · 2026-09-11T00:20:23Z

Duplicate re-check vs #34213 (this maintenance round)

Classifier still marks this as a duplicate of #34213. That is a triage mark, not a close instruction. Re-compared against current origin/main and the related PRs.

	#34213 + landed path	this PR #107047
Kind	Issue (closed): strip tagged reasoning/`<think>` from TTS	Bugfix for visible leading labels (#107044)
Implementing PR	Merged in #73513 as `strip_nonspoken_blocks`	this branch (`6f9eb4e1`)
Files	`tools/tts_text_normalize.py` tagged-block regex only	same file, new `strip_leading_reasoning_labels` + tests
Input	`<think>…</think>` / unterminated `<think>` / verifier footer	`Reasoning:` / `thinking：` / `Analysis:` / `分析：` / `推理：` / `思考：` at start only
On current `origin/main` (no this PR)	`prepare_spoken_text("Reasoning: hello world")` → `"Reasoning: hello world"` (label still spoken)	`"hello world"`

Not the same change. #34213 / strip_nonspoken_blocks never matches a plain leading section label. Live check on origin/main vs this HEAD:

Reasoning: hello world → main speaks it; this PR strips to hello world
thinking：请继续。 / 分析：可见回答。 → same (main keeps the label)
mid-prose The reasoning is clear. and non-allowlisted Summary: hello stay spoken (fail-open)
<think>hidden</think>Visible already clean on both (that is the Feature Request: Strip reasoning/<think> blocks from TTS preprocessing #34213 path)

Independent delta — keep on #107047

Shared-pipeline step after strip_nonspoken_blocks, before markdown
Start-anchored allowlist + count=1
RED/GREEN cases in tests/tools/test_tts_prepare_spoken.py

Do not fold this into #34213 (closed issue; landed). Do not fold into #77079 (still open; expands tag variants only — no strip_leading_*). Closed sibling #107048 targeted the same #107044 case and was closed as a duplicate of this PR; #107047 remains the live candidate.

No review threads, no CI failures, mergeable, no code change this round. Focused tests: test_tts_prepare_spoken.py + test_tts_text_normalize.py → 21 passed.

KoNit-K · 2026-10-01T10:45:03Z

Synced to the current integration base in 14207f1. TTS normalization now strips leading reasoning labels before speech while preserving the remaining spoken text. Focused spoken-text preparation tests pass (19 tests). There are no unresolved review threads to resolve.

fix(tts): strip leading reasoning labels before speech

bcfc139

Gateway auto-TTS was reading visible section labels like Reasoning: or thinking： aloud. Strip allowlisted labels only at text start so mid-prose uses stay spoken. Co-authored-by: Cursor <cursoragent@cursor.com>

alt-glitch added type/bug Something isn't working P3 Low — cosmetic, nice to have tool/tts Text-to-speech and transcription labels Sep 10, 2026

alt-glitch mentioned this pull request Sep 10, 2026

fix(tts): strip a leading visible reasoning label before speech #107048

Closed

This was referenced Sep 14, 2026

fix(tts): Discord voice replies — Opus routing + strip reasoning from speech #103790

Open

fix(matrix): restore voice bubbles — send_voice rejected is_voice, dr… #99712

Open

KoNit-K force-pushed the fix/tts-strip-leading-reasoning-labels branch from 6f9eb4e to bcfc139 Compare September 17, 2026 13:56

fix(tts): strip leading reasoning labels before speech

14207f1

Navigation Menu

Repository navigation

fix(tts): strip leading reasoning labels before speech - #107047

fix(tts): strip leading reasoning labels before speech#107047
KoNit-K wants to merge 2 commits into
NousResearch:mainfrom
KoNit-K:fix/tts-strip-leading-reasoning-labels

KoNit-K commented Sep 10, 2026 •

edited

Loading

Uh oh!

kvnloo commented Sep 10, 2026

Uh oh!

KoNit-K commented Sep 10, 2026

Uh oh!

kvnloo commented Sep 10, 2026

Uh oh!

KoNit-K commented Sep 10, 2026

Uh oh!

KoNit-K commented Sep 11, 2026

Uh oh!

KoNit-K commented Oct 1, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

KoNit-K commented Sep 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Related Issue

Type of Change

Changes Made

How to Test

Evidence

Checklist

Code

Documentation & Housekeeping

Uh oh!

kvnloo commented Sep 10, 2026

Uh oh!

KoNit-K commented Sep 10, 2026

Uh oh!

kvnloo commented Sep 10, 2026

Uh oh!

KoNit-K commented Sep 10, 2026

Duplicate comparison vs #34213 (triage mark)

Independent delta — keep on #107047

Uh oh!

KoNit-K commented Sep 11, 2026

Duplicate re-check vs #34213 (this maintenance round)

Independent delta — keep on #107047

Uh oh!

KoNit-K commented Oct 1, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

KoNit-K commented Sep 10, 2026 •

edited

Loading