feat(orchestrator): S8 shared speech service and sealed model transport

Requirement 001 S8: the hub can use a shared OpenAI-style speech service (KVB_ORCHESTRATOR_SPEECH_URL, key from a mounted file) instead of the local piper and whisper engines, and the image takes its pinned artefacts through an optional INFINITO_ARTEFACT_BASE cache. A new transport serves speech and the LLM: one wall-clock deadline over DNS, connect, TLS handshake and body, certificate and hostname always verified, no redirect followed. A key travels only over https or to loopback; both the speech and the LLM start refuse it otherwise. Key files are capped at 4096 bytes. Streamed WAV answers with placeholder sizes are accepted only when the data chunk runs to the end, and are rewritten before caching. Accepted MEDUSA false positives (41, independent review, operator pick): raw client socket patterns in transport.py, AIC-EXFIL-005 on the sealed key path, JB-SCAN-2237 on the hub's own system prompt, STG009 on the RIFF pad byte and LLM01 at language.py:111, where agent-written titles reach the NL prompt but the model only proposes a plan the operator confirms.

🎯 Fix: 93 %
💥 Break: 4 %
💸 Cost: –
⏱️ Time: –
🛡️ Security: 100 %
🏆 Score: 89
🚦 Gate: ✅ 90/5, cost and time <= 100, security 100 %

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Kevin Veen-BirkenbachandClaude Opus 5.5 committed 2026-11-13 16:57:52 +00:00
1 parent 590b430d37
commit 6d2bca77f7
16 files changed
+1253 -56

No files matched your search

+23 -7
View File
@@ -39,8 +39,8 @@
--llm-url: base URL of an OpenAI-compatible chat endpoint; empty turns the
natural-language command line off.
--llm-model: model name sent to that endpoint; empty turns it off.
--llm-key-file: file holding the endpoint's bearer key; empty file or
unreadable means no key.
--llm-key-file: file holding the endpoint's bearer key; an empty file
means no key, an unreadable one stops the start.
--piper: the Piper executable, run as a separate process; missing turns
speech output off.
--voices: directory holding the Piper voices as <model>.onnx with
@@ -51,6 +51,12 @@
input off.
--whisper-model: the ggml model whisper.cpp loads.
--scratch: writable directory the engines use as HOME and TMPDIR.
--speech-url: base URL with /v1 of an OpenAI-compatible audio API such as
the core role svc-ai-speech; /api/tts and /api/stt then render there
instead of in Piper and whisper.cpp, behind the same CPU gate and
transcription limits. Empty keeps the in-container engines.
--speech-key-file: file holding that service's bearer key; an empty file
means no key, an unreadable one stops the start.
--approvers: the pinned passkeys, mounted read-only; the hub lists their
ids and algorithms for the page, never pins, and verifies a signature
only for the operator messages, denials and rules it acts on itself,
@@ -70,9 +76,11 @@ from kvb_orchestrator.errors import ToolError
from kvb_orchestrator.exchange import HubServer
from kvb_orchestrator.handler import Handler
from kvb_orchestrator.hub import Places, open_hub
from kvb_orchestrator.language import Model, read_key
from kvb_orchestrator.keys import key_from
from kvb_orchestrator.language import Model
from kvb_orchestrator.redirects import TIMEOUT, Redirects, dial
from kvb_orchestrator.skill_scripts import Scripts
from kvb_orchestrator.speechservice import Service
def main() -> int:
@@ -81,9 +89,9 @@ def main() -> int:
for name in ("--listen-port", "--public-port"):
parser.add_argument(name, type=int, required=True)
paths = ("--council", "--pages", "--state", "--host-state", "--skills", "--transcripts", "--templates", "--tree-file", "--timers", "--kept", "--vendor", "--cleanup", "--offices", "--llm-key-file", "--approvers")
for name in (*paths, "--piper", "--voices", "--whisper", "--whisper-model", "--scratch"):
for name in (*paths, "--piper", "--voices", "--whisper", "--whisper-model", "--scratch", "--speech-key-file"):
parser.add_argument(name, type=Path, required=True)
for name in ("--llm-url", "--llm-model", "--operator-voice"):
for name in ("--llm-url", "--llm-model", "--operator-voice", "--speech-url"):
parser.add_argument(name, required=True)
args = parser.parse_args()
try:
@@ -97,8 +105,16 @@ def main() -> int:
scripts=Scripts(skills=args.skills, council=args.council, home=args.state),
templates=args.templates,
places=Places(tree_file=args.tree_file, timers=args.timers, vendor=args.vendor, cleanup=args.cleanup, offices=args.offices, kept=args.kept),
model=Model(url=args.llm_url.strip(), model=args.llm_model.strip(), key=read_key(args.llm_key_file)),
engines=Engines(piper=args.piper, voices=args.voices, operator_voice=args.operator_voice, whisper=args.whisper, whisper_model=args.whisper_model, scratch=args.scratch),
model=Model(url=args.llm_url.strip(), model=args.llm_model.strip(), key=key_from(args.llm_key_file, "model endpoint")),
engines=Engines(
piper=args.piper,
voices=args.voices,
operator_voice=args.operator_voice,
whisper=args.whisper,
whisper_model=args.whisper_model,
scratch=args.scratch,
service=Service(url=args.speech_url.strip(), key=key_from(args.speech_key_file, "speech service")),
),
approvers=args.approvers,
redirects=Redirects(tls=ssl.create_default_context(), dial=dial, timeout=TIMEOUT),
)