Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

fix(dictation): keep the STT dictionary out of the decoder prompt (BUG-211)

rubenluetke10-beep27110f04a1d8
diff --git a/docs/BUGS.md b/docs/BUGS.md
index 43c73b9b27..1f0c8e4378 100644
--- a/docs/BUGS.md
+++ b/docs/BUGS.md
@@ -14825,3 +14825,45 @@ off-Windows no-op), `tests/unit/platform/test_deescalate.py`
**Related.** BUG-204 (the VRAM holder this amplified), BUG-189 (the cold-boot
start storm the priority made fatal), AP-17 (the task stays `RunLevel
Limited` — priority is orthogonal to elevation).
+
+## BUG-211: dictation pasted dictionary words nobody said ("IDE, gone.") (HIGH, FIXED 2026-09-01)
+
+**Symptom.** A dictation lands in the document with a dictionary term the
+user never spoke — "IDE." on a 0.8 s near-silent press, "IDE, gone.",
+"IDE, noFableForCent, …" — and longer recitations ("IDE, Agentic IDE,
+Agentic IDE, Agentic") show up in the log as stripped echoes five times in
+sixteen minutes. Words that only resemble a dictionary term also arrive as
+the term, and the whole lane feels laggy.
+
+**Root cause.** Two things stacked. A user-scope environment variable
+(`JARVIS__STT__PROVIDER=groq-api`) overrode the `faster-whisper` the config
+file names, so every dictation ran on cloud Whisper — 208 of 208 history
+rows that day — even though the local engine was installed in the app's
+interpreter. The dictation lane then handed the STT dictionary
+("GitHub, Grok, Claude, Agentic IDE, Agentic") to that provider as the
+Whisper `prompt`. Whisper answers a pause or silence by reciting the primed
+list; `strip_prompt_echo` (BUG-185) removes a RUN of items but must leave a
+single one alone, because "Claude." is something a person says. The local
+engine never receives a prompt by design, which is why the defect only
+appeared once the crossover put a prompt-capable provider in front. The lag
+is the cloud round-trip per window (median 440 ms, p90 1.6 s, max 10.9 s).
+
+**Fix.** `build_stt_from_config` / `build_named_stt_provider` grew a
+`dictionary_bias` keyword; `_dictation_stt` builds its primary AND every
+fallback alternate with `dictionary_bias=False`. The dictation lane now sends
+NO decoder prompt at all — neither the voice bias paragraph (F14) nor the
+dictionary — on any provider, so there is nothing to recite. The dictionary
+keeps correcting the finished transcript through `DictionaryCorrectingSTT`,
+the half of the feature that cannot invent a word. The voice lane is
+unchanged. The environment override is a per-machine setting and was left
+in place; the history row's `metadata.stt_providers` names the recognizer
+that really ran.
+
+**Guard.** `tests/unit/dictation/test_stt_retry_and_isolation.py`
+(`test_the_dictation_provider_is_built_without_the_voice_bias_prompt`
+asserts `dictionary_bias=False`),
+`tests/unit/plugins/stt/test_dictionary_bias_opt_out.py`.
+
+**Related.** BUG-185 (the echo guard, which stays as the net for the voice
+lane), F14 (dictation builds its own prompt-free provider), AP-21/22 (the
+opt-out rides the same cross-family chain).
diff --git a/jarvis/plugins/stt/__init__.py b/jarvis/plugins/stt/__init__.py
index 717b9905e5..a92cd3238b 100644
--- a/jarvis/plugins/stt/__init__.py
+++ b/jarvis/plugins/stt/__init__.py
@@ -491,11 +491,19 @@ def _load_provider_class(name: str) -> type | None:
return None
-def build_stt_from_config(stt_cfg: Any) -> Any:
+def build_stt_from_config(stt_cfg: Any, *, dictionary_bias: bool = True) -> Any:
"""Return an STTProvider instance for ``stt_cfg.provider``.
Falls back to a local FasterWhisperProvider if the configured provider has
no entry-point or raises on construction.
+
+ ``dictionary_bias=False`` keeps the user's STT dictionary OUT of the
+ decoder prompt. The dictation lane asks for this (BUG-211): a
+ prompt-capable Whisper answers a pause or a near-silent press by reciting
+ the primed word list ("IDE, Agentic IDE, Agentic"), and the echo guard can
+ only strip a RUN of items — a lone "IDE" reads as a word and lands in the
+ document. The dictionary still applies to every transcript through the
+ post-STT corrector, which is the half of the feature that cannot invent.
"""
configured_name = (getattr(stt_cfg, "provider", "") or "").strip()
# Both crossings — the on-device engine missing from THIS interpreter, and a
@@ -522,13 +530,14 @@ def build_stt_from_config(stt_cfg: Any) -> Any:
# still gets the dictionary's post-STT corrections — AP-21/22). The local
# utterance fwhisper deliberately receives NO initial_prompt (silence-
# hallucination risk, see _build_local_fallback).
- try:
- from jarvis.speech.stt_dictionary import dictionary_bias_words
+ vocab: list[str] = []
+ if dictionary_bias:
+ try:
+ from jarvis.speech.stt_dictionary import dictionary_bias_words
- vocab = dictionary_bias_words()
- except Exception as exc: # noqa: BLE001 — the dictionary must never break STT build
- logger.debug("STT dictionary bias words unavailable: {}", exc)
- vocab = []
+ vocab = dictionary_bias_words()
+ except Exception as exc: # noqa: BLE001 — the dictionary must never break STT build
+ logger.debug("STT dictionary bias words unavailable: {}", exc)
if vocab:
joined = ", ".join(vocab)
bias_prompt = f"{bias_prompt}, {joined}" if bias_prompt else joined
@@ -688,7 +697,9 @@ def available_stt_provider_names() -> list[str]:
return names
-def build_named_stt_provider(name: str, stt_cfg: Any) -> Any:
+def build_named_stt_provider(
+ name: str, stt_cfg: Any, *, dictionary_bias: bool = True
+) -> Any:
"""Build the provider called ``name`` using ``stt_cfg`` for everything else.
Language, bias prompt, dictionary vocabulary and the team-proxy handling all
@@ -702,7 +713,7 @@ def build_named_stt_provider(name: str, stt_cfg: Any) -> Any:
except AttributeError:
# Not a pydantic model (test doubles); a tiny view is enough.
patched = _StttConfigView(stt_cfg, name)
- return build_stt_from_config(patched)
+ return build_stt_from_config(patched, dictionary_bias=dictionary_bias)
class _StttConfigView:
diff --git a/jarvis/speech/pipeline.py b/jarvis/speech/pipeline.py
index 1fe00a5ab0..8a91467659 100644
--- a/jarvis/speech/pipeline.py
+++ b/jarvis/speech/pipeline.py
@@ -10781,7 +10781,11 @@ class SpeechPipeline:
# reads as empty, which is the shipped behaviour we want anyway.
own_prompt = str(getattr(dictation_cfg, "bias_prompt", "") or "").strip()
patched = stt_cfg.model_copy(update={"bias_prompt": own_prompt})
- instance = build_stt_from_config(patched)
+ # No dictionary words in the decoder prompt either (BUG-211): a
+ # prompt-capable Whisper recites the primed list over every pause
+ # and a lone recited item survives the echo guard. The dictionary
+ # still corrects the finished transcript — see the wrapper below.
+ instance = build_stt_from_config(patched, dictionary_bias=False)
except Exception as exc: # noqa: BLE001 — never lose dictation over this
log.warning(
"Dictation-specific STT could not be built (%s); reusing the "
@@ -10804,7 +10808,9 @@ class SpeechPipeline:
from jarvis.speech.stt_fallback import FallbackSTT
def _build(name: str) -> Any:
- return build_named_stt_provider(name, patched)
+ return build_named_stt_provider(
+ name, patched, dictionary_bias=False
+ )
instance = FallbackSTT(
instance, alternates, _build, primary_name=configured
diff --git a/tests/unit/dictation/test_stt_retry_and_isolation.py b/tests/unit/dictation/test_stt_retry_and_isolation.py
index 17966e0351..489887f1a6 100644
--- a/tests/unit/dictation/test_stt_retry_and_isolation.py
+++ b/tests/unit/dictation/test_stt_retry_and_isolation.py
@@ -386,8 +386,8 @@ def test_the_dictation_provider_is_built_without_the_voice_bias_prompt(
seen: list[Any] = []
- def _build(cfg: Any) -> Any:
- seen.append(cfg)
+ def _build(cfg: Any, **kwargs: Any) -> Any:
+ seen.append((cfg, kwargs))
return SimpleNamespace(name="built")
monkeypatch.setattr(stt_plugins, "build_stt_from_config", _build)
@@ -399,7 +399,11 @@ def test_the_dictation_provider_is_built_without_the_voice_bias_prompt(
instance = pipe._dictation_stt()
assert instance is not None
- assert seen and seen[0].bias_prompt == ""
+ assert seen and seen[0][0].bias_prompt == ""
+ # ...and the STT dictionary stays out of the decoder prompt too (BUG-211):
+ # Whisper recites a primed word list over pauses, and a lone recited item
+ # passes the echo guard and lands in the document.
+ assert seen[0][1] == {"dictionary_bias": False}
# The VOICE config is untouched — this is a copy, not a mutation.
assert pipe._config.stt.bias_prompt == "Jarvis, Adex, Ruben, Vokando"
@@ -413,7 +417,7 @@ def test_the_dictation_provider_is_built_once_and_cached(
monkeypatch.setattr(
stt_plugins,
"build_stt_from_config",
- lambda cfg: builds.append(cfg) or SimpleNamespace(),
+ lambda cfg, **_k: builds.append(cfg) or SimpleNamespace(),
)
monkeypatch.setattr(stt_plugins, "resolve_keyed_stt_fallback", lambda *_a, **_k: ())
pipe = _pipeline_with_config(provider="groq-api")
@@ -442,7 +446,7 @@ def test_the_dictation_provider_arms_the_cross_family_chain(
from jarvis.speech.stt_fallback import FallbackSTT
monkeypatch.setattr(
- stt_plugins, "build_stt_from_config", lambda cfg: SimpleNamespace()
+ stt_plugins, "build_stt_from_config", lambda cfg, **_k: SimpleNamespace()
)
monkeypatch.setattr(
pipeline_mod,
@@ -467,7 +471,7 @@ def test_an_unbuildable_dictation_provider_falls_back_to_the_voice_one(
"""A dictation with the wrong prompt beats a dictation with no provider."""
import jarvis.plugins.stt as stt_plugins
- def _boom(cfg: Any) -> Any:
+ def _boom(cfg: Any, **_kwargs: Any) -> Any:
raise RuntimeError("no entry-point")
monkeypatch.setattr(stt_plugins, "build_stt_from_config", _boom)
diff --git a/tests/unit/plugins/stt/test_dictionary_bias_opt_out.py b/tests/unit/plugins/stt/test_dictionary_bias_opt_out.py
new file mode 100644
index 0000000000..b4b4d614ec
--- /dev/null
+++ b/tests/unit/plugins/stt/test_dictionary_bias_opt_out.py
@@ -0,0 +1,78 @@
+"""BUG-211: the STT dictionary stays out of the decoder prompt on request.
+
+A prompt-capable Whisper recites its primed word list over a pause, and a
+lone recited item passes the echo guard. The dictation lane therefore builds
+its provider with ``dictionary_bias=False``; this pins that the flag really
+keeps the words out of the ``prompt`` kwarg, and that the default still
+merges them for the voice lane.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from typing import Any
+
+import pytest
+
+import jarvis.plugins.stt as stt_plugins
+import jarvis.speech.stt_dictionary as dictionary
+
+
+class _RecordingProvider:
+ """Stands in for a cloud plugin; keeps the kwargs it was built with."""
+
+ built: list[dict[str, Any]] = []
+
+ def __init__(self, **kwargs: Any) -> None:
+ type(self).built.append(kwargs)
+
+
+def _cloud_provider(monkeypatch: pytest.MonkeyPatch) -> type[_RecordingProvider]:
+ _RecordingProvider.built = []
+ monkeypatch.setattr(stt_plugins, "_resolve_effective_stt", lambda name: (name, ""))
+ monkeypatch.setattr(stt_plugins, "_load_provider_class", lambda name: _RecordingProvider)
+ monkeypatch.setattr(
+ dictionary, "dictionary_bias_words", lambda *_a, **_k: ["GitHub", "Agentic IDE"]
+ )
+ return _RecordingProvider
+
+
+def _cfg(**overrides: Any) -> Any:
+ base = {"provider": "groq-api", "language": "auto", "bias_prompt": "", "model": ""}
+ base.update(overrides)
+ return SimpleNamespace(**base)
+
+
+def test_the_dictionary_is_merged_into_the_prompt_by_default(
+ _cloud_provider: type[_RecordingProvider],
+) -> None:
+ stt_plugins.build_stt_from_config(_cfg(bias_prompt="Jarvis"))
+
+ assert _cloud_provider.built[0]["prompt"] == "Jarvis, GitHub, Agentic IDE"
+
+
+def test_the_opt_out_sends_no_prompt_when_the_dictionary_is_the_only_bias(
+ _cloud_provider: type[_RecordingProvider],
+) -> None:
+ stt_plugins.build_stt_from_config(_cfg(), dictionary_bias=False)
+
+ assert "prompt" not in _cloud_provider.built[0]
+
+
+def test_the_opt_out_keeps_an_explicit_bias_prompt_but_not_the_dictionary(
+ _cloud_provider: type[_RecordingProvider],
+) -> None:
+ stt_plugins.build_stt_from_config(_cfg(bias_prompt="Jarvis"), dictionary_bias=False)
+
+ assert _cloud_provider.built[0]["prompt"] == "Jarvis"
+
+
+def test_a_named_fallback_honours_the_opt_out_too(
+ _cloud_provider: type[_RecordingProvider],
+) -> None:
+ """A fallback that quietly re-added the words would recite them the
+ moment it took over — the chain must match the primary."""
+ stt_plugins.build_named_stt_provider("openai-api", _cfg(), dictionary_bias=False)
+
+ assert "prompt" not in _cloud_provider.built[0]