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.

feat(missions): speak a voice mission's result without an extra model call

rubenluetke10-beepaf19b64dc888
diff --git a/jarvis/missions/init.py b/jarvis/missions/init.py
index 73aa5bcf4c..5347bdfa9b 100644
--- a/jarvis/missions/init.py
+++ b/jarvis/missions/init.py
@@ -1321,19 +1321,10 @@ async def bootstrap_missions(
# the listener is the direct-TTS fallback.
_readback_mode = _resolve_readback_mode(tts_speak_fn=tts_speak_fn, speech_bus=speech_bus)
- # Context-aware readback composer (maintainer mandate: no fixed stock
- # phrases). One instance shared by whichever readback path is active. Built
- # in fallback-only mode when the flash path is off, so the spoken line is the
- # existing canned/signed text unless [ack_brain] is enabled. Lazy import +
- # best-effort so an import cycle / wiring fault never blocks the bootstrap.
- _readback_composer = None
- try:
- from jarvis.brain.factory import build_readback_composer
-
- _readback_composer = build_readback_composer()
- except Exception as exc: # noqa: BLE001
- logger.warning("Mission readback composer wiring skipped: %s", exc)
-
+ # The spoken line is the mission's own signed summary or a fixed status
+ # phrase — no extra model call rephrases it (2026-09-30: the flash
+ # composer billed an API key for every background result). Inside a live
+ # call the realtime model already words the announcement itself.
voice_listener: MissionVoiceListener | None = None
if _readback_mode == "listener":
voice_listener = MissionVoiceListener(
@@ -1343,7 +1334,6 @@ async def bootstrap_missions(
tts_speak_fn=tts_speak_fn,
announce_critic_loop=voice_announce_critic_loop,
language_default=voice_language_default, # type: ignore[arg-type]
- readback_composer=_readback_composer,
)
await voice_listener.start()
logger.info("Phase-6 voice listener active (direct-TTS readback)")
@@ -1367,7 +1357,6 @@ async def bootstrap_missions(
speech_bus=speech_bus,
announce_critic_loop=voice_announce_critic_loop,
language_default=voice_language_default, # type: ignore[arg-type]
- readback_composer=_readback_composer,
)
await mission_announcer.start()
logger.info("Phase-6 mission-announcer active (mission-bus -> speech-bus)")
diff --git a/jarvis/missions/voice/announcer.py b/jarvis/missions/voice/announcer.py
index a3716b5f9e..e9e475d249 100644
--- a/jarvis/missions/voice/announcer.py
+++ b/jarvis/missions/voice/announcer.py
@@ -32,7 +32,7 @@ from __future__ import annotations
import json
import logging
-from typing import TYPE_CHECKING, Literal
+from typing import Literal
from jarvis.brain.output_filter import scrub_for_voice
from jarvis.brain.scrub_verdict import is_harmless_scrub_residue
@@ -43,10 +43,6 @@ from jarvis.missions.voice.readback import (
failure_phrase_key,
render_agent_brand,
)
-from jarvis.voice.contextual_readback import render_readback
-
-if TYPE_CHECKING:
- from jarvis.voice.contextual_readback import ReadbackComposer
from ..event_bus import MissionBus
from ..event_store import MissionEventStore
@@ -100,7 +96,6 @@ class MissionAnnouncer:
scrub: bool = True,
announce_critic_loop: bool = False,
language_default: _Lang = "de",
- readback_composer: ReadbackComposer | None = None,
) -> None:
self._bus = bus
self._store = store
@@ -108,12 +103,6 @@ class MissionAnnouncer:
self._scrub = scrub
self._announce_critic_loop = announce_critic_loop
self._lang_default: _Lang = language_default
- # Context-aware readbacks (maintainer mandate: no fixed stock phrases).
- # The signed/canned line from _render is the deterministic ground truth;
- # the composer only rephrases it naturally, honesty-bound for the signed
- # MissionApproved summary (ADR-0009 — rephrase, never invent). None =>
- # the canned line is spoken unchanged (risk-free when unwired).
- self._readback_composer = readback_composer
# Cache (mission_id -> (is_voice_source, language))
self._mission_voice_cache: dict[str, tuple[bool, _Lang]] = {}
self._unsubscribe = None # set by start()
@@ -142,30 +131,13 @@ class MissionAnnouncer:
if not is_voice:
return
+ # The runtime-signed summary or a fixed status phrase, spoken as it
+ # is: no model call rephrases it (2026-09-30). Inside a live call the
+ # realtime model words it from this line; otherwise classic TTS reads it.
text, priority = self._render(env, lang)
if not text:
return
- # Context-aware rephrasing of the signed/canned line. honesty_bound for
- # the MissionApproved summary so the spoken surface stays a faithful
- # rephrasing of the Kontrollierer-signed observation (ADR-0009); failure/
- # timeout/cancel are status, not observations, so they phrase more freely
- # (digit + vocab guards + strict persona still apply). Falls back to the
- # exact canned line on any miss (AD-OE6, zero silent drops).
- instruction, honesty_bound = self._situation(env.payload)
- canned_line = text
- text = await render_readback(
- self._readback_composer,
- instruction=instruction,
- language=lang,
- canned=lambda: canned_line,
- facts={"result": canned_line},
- honesty_bound=honesty_bound,
- latency_budget_ms=2500,
- )
- if not text:
- return
-
if self._scrub:
scrubbed = scrub_for_voice(text, language=lang)
if scrubbed.actions:
@@ -225,32 +197,6 @@ class MissionAnnouncer:
detail["reason"] = payload.reason
return json.dumps(detail, ensure_ascii=True, separators=(",", ":"))
- @staticmethod
- def _situation(payload: object) -> tuple[str, bool]:
- """English (instruction, honesty_bound) for a terminal mission event.
-
- ``honesty_bound`` is True ONLY for the signed success summary, so its
- spoken surface stays a faithful rephrasing (ADR-0009). Failure / timeout
- / cancel are status lines, not observations, so they phrase freely.
- """
- if isinstance(payload, MissionApproved):
- return (
- "A background task the user asked for has finished successfully; "
- "tell them naturally, keeping the reported result faithfully.",
- True,
- )
- if isinstance(payload, MissionFailed):
- return (
- "A background task the user asked for did not succeed; tell them "
- "plainly and kindly, keeping any reason given.",
- False,
- )
- if isinstance(payload, MissionTimedOut):
- return ("A background task the user asked for ran out of time.", False)
- if isinstance(payload, MissionCancelled):
- return ("A background task the user asked for was cancelled.", False)
- return ("A background task the user asked for has an update.", False)
-
def _render(
self, env: EventEnvelope, lang: _Lang,
) -> tuple[str, Literal["normal", "interrupt"]]:
diff --git a/jarvis/missions/voice/listener.py b/jarvis/missions/voice/listener.py
index 716ffa0319..8d0a9b1df0 100644
--- a/jarvis/missions/voice/listener.py
+++ b/jarvis/missions/voice/listener.py
@@ -21,11 +21,9 @@ from __future__ import annotations
import logging
from collections.abc import Awaitable, Callable
-from typing import TYPE_CHECKING
from jarvis.brain.output_filter import scrub_for_voice
from jarvis.brain.scrub_verdict import is_harmless_scrub_residue
-from jarvis.voice.contextual_readback import render_readback
from ..event_bus import MissionBus
from ..event_store import MissionEventStore
@@ -41,10 +39,6 @@ from ..events import (
)
from .readback import Lang, MissionReadback
-if TYPE_CHECKING:
- from jarvis.voice.contextual_readback import ReadbackComposer
-
-
logger = logging.getLogger(__name__)
@@ -66,7 +60,6 @@ class MissionVoiceListener:
tts_speak_fn: TTSSpeakFn,
announce_critic_loop: bool = False,
language_default: Lang = "de",
- readback_composer: ReadbackComposer | None = None,
) -> None:
self._bus = bus
self._store = store
@@ -74,11 +67,6 @@ class MissionVoiceListener:
self._tts = tts_speak_fn
self._announce_critic_loop = announce_critic_loop
self._lang_default = language_default
- # Context-aware readbacks (maintainer mandate: no fixed stock phrases).
- # The signed/canned line from _render is the ground truth; the composer
- # rephrases it, honesty-bound for the signed MissionApproved summary
- # (ADR-0009). None => canned line spoken unchanged (risk-free).
- self._readback_composer = readback_composer
# Cache: mission_id -> (is_voice, language). Filled lazily on the
# first event per mission. None = not yet resolved.
self._mission_voice_cache: dict[str, tuple[bool, Lang]] = {}
@@ -101,28 +89,12 @@ class MissionVoiceListener:
if not is_voice:
return
+ # The signed summary or a fixed status phrase, spoken as it is: no
+ # model call rephrases it (2026-09-30).
text = self._render(env, lang)
if not text:
return
- # Context-aware rephrasing of the signed/canned line (maintainer mandate).
- # honesty_bound only for the signed success summary (ADR-0009 — rephrase,
- # never invent); other statuses phrase more freely. Falls back to the
- # exact canned line on any miss (AD-OE6).
- instruction, honesty_bound = self._situation(env.payload)
- canned_line = text
- text = await render_readback(
- self._readback_composer,
- instruction=instruction,
- language=lang,
- canned=lambda: canned_line,
- facts={"result": canned_line},
- honesty_bound=honesty_bound,
- latency_budget_ms=2500,
- )
- if not text:
- return
-
# Audit F-AUDIT-2: send the mission readback through the output
# filter before handing it to TTS. Otherwise mission outputs with
# tool-use markup / "Sir" address / engineering jargon reached the
@@ -151,38 +123,6 @@ class MissionVoiceListener:
await self._tts(scrubbed.cleaned, lang)
- @staticmethod
- def _situation(payload: object) -> tuple[str, bool]:
- """English (instruction, honesty_bound) for a mission event.
-
- ``honesty_bound`` is True ONLY for the signed success summary, so its
- spoken surface stays a faithful rephrasing (ADR-0009). Everything else is
- status, not an observation, so it phrases freely.
- """
- if isinstance(payload, MissionApproved):
- return (
- "A background task the user asked for has finished successfully; "
- "tell them naturally, keeping the reported result faithfully.",
- True,
- )
- if isinstance(payload, MissionFailed):
- return (
- "A background task the user asked for did not succeed; tell them "
- "plainly and kindly, keeping any reason given.",
- False,
- )
- if isinstance(payload, MissionTimedOut):
- return ("A background task the user asked for ran out of time.", False)
- if isinstance(payload, MissionCancelled):
- return ("A background task the user asked for was cancelled.", False)
- if isinstance(payload, MissionBudgetWarning):
- return ("A background task is approaching its budget limit.", False)
- if isinstance(payload, WorkerKilled):
- return ("A background task was stopped for a safety reason.", False)
- if isinstance(payload, WorkerCorrectionRequired):
- return ("A background task is still being worked on.", False)
- return ("A background task the user asked for has an update.", False)
-
def _render(self, env: EventEnvelope, lang: Lang) -> str:
"""Map event payload -> voice text. Empty string if the event type
has no voice response.
diff --git a/tests/missions/test_voice_announcer.py b/tests/missions/test_voice_announcer.py
index c89fd8216f..08566a5072 100644
--- a/tests/missions/test_voice_announcer.py
+++ b/tests/missions/test_voice_announcer.py
@@ -30,28 +30,6 @@ from jarvis.missions.ids import uuid7_str
from jarvis.missions.voice.announcer import MissionAnnouncer
-class _FakeAckProvider:
- """Minimal flash-provider stand-in for ReadbackComposer wiring tests."""
-
- def __init__(self, reply: str | None) -> None:
- self.reply = reply
-
- async def run(self, utterance: str, language: str, *, persona_prompt: str) -> str | None:
- return self.reply
-
-
-def _composer_with(reply: str | None):
- from jarvis.brain.ack_brain import CircuitBreaker
- from jarvis.brain.ack_brain.config import AckBrainConfig
- from jarvis.voice.contextual_readback import ReadbackComposer
-
- return ReadbackComposer(
- provider=_FakeAckProvider(reply),
- config=AckBrainConfig(timeout_ms=1500),
- breaker=CircuitBreaker(threshold=3, cooldown_s=60),
- )
-
-
@pytest.fixture
async def store_and_bus(tmp_missions_db: Path):
bus = MissionBus()
@@ -154,56 +132,20 @@ async def test_approved_emits_announcement(store_and_bus) -> None:
@pytest.mark.asyncio
-async def test_approved_readback_is_contextually_rephrased(store_and_bus) -> None:
- """With a composer wired, the signed summary is spoken in a fresh phrasing
- (maintainer mandate: no fixed stock phrases) — but still faithfully."""
- store, bus = store_and_bus
- speech_bus = EventBus()
- captured = _collect_announcements(speech_bus)
-
- announcer = MissionAnnouncer(
- bus=bus, store=store, speech_bus=speech_bus,
- readback_composer=_composer_with("All done — the mission completed."),
- )
- await announcer.start()
-
- mid = await _seed_voice_mission(store, language="en")
- await store.append_and_publish(
- EventEnvelope(
- mission_id=mid,
- source_actor="kontrollierer",
- ts_ms=now_ms(),
- payload=MissionApproved(
- result_uri=f"mission://{mid}", tokens_used=100, cost_usd=0.05,
- wall_ms=1000, summary_de="Mission abgeschlossen.",
- summary_en="Mission completed.",
- ),
- )
- )
-
- assert len(captured) == 1
- text = captured[0].text
- assert text != "Mission completed." # rephrased, not the canned summary
- assert "completed" in text.lower() # but still faithful to the signed fact
+async def test_the_signed_summary_is_spoken_without_a_model_call(store_and_bus) -> None:
+ """No flash composer rephrases a mission result any more (2026-09-30): the
+ announcer has no composer to wire, and the Kontrollierer-signed summary is
+ the spoken line, verbatim (ADR-0009)."""
+ import inspect
-
-async def test_approved_falls_back_to_signed_summary_when_generation_fails(
- store_and_bus,
-) -> None:
- """A composer whose provider yields nothing must fall back to the exact
- Kontrollierer-signed summary (ADR-0009 + AD-OE6 zero silent drops)."""
+ assert "readback_composer" not in inspect.signature(MissionAnnouncer).parameters
store, bus = store_and_bus
speech_bus = EventBus()
captured = _collect_announcements(speech_bus)
-
- announcer = MissionAnnouncer(
- bus=bus, store=store, speech_bus=speech_bus,
- readback_composer=_composer_with(""), # provider returns empty -> fallback
- )
+ announcer = MissionAnnouncer(bus=bus, store=store, speech_bus=speech_bus)
await announcer.start()
- mid = await _seed_voice_mission(store, language="de")
+ mid = await _seed_voice_mission(store, language="en")
await store.append_and_publish(
EventEnvelope(
mission_id=mid,
@@ -217,8 +159,8 @@ async def test_approved_falls_back_to_signed_summary_when_generation_fails(
)
)
- assert len(captured) == 1
- assert captured[0].text == "Mission abgeschlossen."
+ assert [a.text for a in captured] == ["Mission completed."]
+ assert captured[0].kind == "subagent", "still owed to the user who asked"
@pytest.mark.asyncio