Asterisk is the registrar for devices, not Hold Slayer. A softphone REGISTERs
to the lab and the gateway transfers a live call to it by dialling extension
2001. This is deliberate: Hold Slayer's own SIP listener answers 200 OK to any
REGISTER with no digest challenge, so anything on the network could register
as a device and receive transferred calls. Keeping registration in Asterisk
means the lab does not exercise or depend on that path, and the device is
authenticated.
The pjsua CLI built alongside the Python bindings is the test device — same
library stack as the gateway, so no new dependency. Verified end to end: a
gateway call to 2001 produces two channels Up under one bridge id.
Three things that cost time and are now written down:
- `--realm=asterisk`, not `--realm='*'`: the wildcard fails against Asterisk's
digest challenge with PJSIP_EFAILEDCREDENTIAL.
- pjsua is an interactive console app and exits ~8s after start if stdin is
closed or /dev/null. `script -qfc` and `setsid </dev/null` both appear to
work — registration succeeds — and then the process dies, leaving a stale
contact in Asterisk that routes INVITEs to a port nobody is listening on.
Hold a fifo open on stdin instead, and verify the port is actually bound
rather than trusting `pjsip show contacts`.
- Qualify is off for this AOR: the pjsua console does not answer OPTIONS, so
polling marks a working softphone Unavail and the dialplan refuses to ring
it. The 2001 guard therefore tests PJSIP_AOR(softphone,contact) rather than
DEVICE_STATE. A real hardphone answers OPTIONS and can have it re-enabled.
The identify block now matches source address *and port*. A host-only match
claims every packet from that address, so a co-located softphone's REGISTER
was attributed to the gateway endpoint and checked against the gateway's
password — surfacing as "Failed to authenticate" on a correct password.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The speech fixture classified as LIVE_HUMAN on its first 3s window and drifted
to MUSIC for every window after. I had validated only the first window and
reported the fixture as verified, which overstated it: any lab result resting
on that fixture — the hold-slayer scenarios above all — was proving less than
it appeared to.
The cause was one modelling error, not a tuning problem. `_detect_tonality`
looks for an autocorrelation peak above 0.5 in the 50-1000 Hz lag range, and
each syllable used a *constant* f0, which is perfectly periodic there. That
scored is_tonal=True, handing the music score a free 0.3 that speech could not
outrun — and the decision requires speech_score to strictly exceed
music_score, so ties went to music.
Real voices glide and jitter, so the periodicity never locks. The fundamental
now follows a per-syllable pitch contour (rise or fall, plus ~2% cycle-to-cycle
jitter), with the frequency integrated to phase rather than multiplied by t —
`2*pi*f*t` is only a chirp when f is the instantaneous rate, which it is not
once f0 itself moves. is_tonal is now False in every window.
Two smaller fixes fell out of that:
- Aspiration noise is high-passed rather than broadband. Flat noise puts
energy in every Goertzel bin, so the strongest DTMF row and column both
clear the detector's `total_power * 0.1` threshold and each syllable reads
as a keypress. A first-difference filter leaves the 697-1633 Hz bands
comparatively empty. The level is set for margin — spectral flatness lands
at ~0.46, mid-way through the 0.1-0.5 band, not on an edge.
- The music fixture gained two more harmonics and a recording-style noise
floor. Windows straddling a chord change had a momentarily sparse spectrum
and fell *below* the music score's 0.05 flatness floor, scoring as speech.
All three fixtures now classify correctly in 100% of windows (music 27/27,
speech 5/5, silence 2/2), and remain correct when the window is stepped by
half a window — a fixture that only works on aligned boundaries would still
be a trap in a live call, where the analysis window has no relationship to
where the audio began.
Confirmed on a real call through the lab: scenario 1003 now shows the whole
hold-slayer arc, speech -> sustained music -> speech, matching the dialplan.
tests/test_lab_fixtures.py guards this: it sweeps every window rather than
sampling the first, which is exactly what the original validation missed, and
checks the generator is byte-for-byte deterministic. It skips when the
fixtures have not been generated, since they are gitignored.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An Asterisk instance that answers calls, plays an IVR, holds with music and
connects a "human", so the gateway has something real to dial that is not the
PSTN: no charges, no strangers, no E911 exposure.
Asterisk rather than Kamailio because the unproven risks are media risks.
Kamailio is a proxy — it routes signalling and answers nothing, so it would
forward the INVITE and find nobody home. Asterisk is a B2BUA: it answers,
plays prompts and collects DTMF, which is the hold-slayer scenario itself.
Kamailio remains the better model for trunk registration/digest auth later.
No application changes are needed to use it. SIP_TRUNK_HOST is just an
address, so the production code path runs unmodified — there is no test-only
branch anywhere in the gateway. It also means safety is structural: while the
trunk points at the lab there is no route to the PSTN at all, an absence of
route rather than a policy that could be misconfigured.
Nine scenarios (1001-1008 plus an echo test) cover the baseline call, the
IVR/DTMF path, hold-then-human, long hold, busy, no-answer, remote hangup and
silence.
The image ships no sound files, so sounds/generate.py synthesises three
fixtures from fixed seeds — byte-identical on every run, which is what makes
a classifier regression distinguishable from noise. Verified against
AudioClassifier: music→MUSIC 0.85, speech→LIVE_HUMAN 0.75,
silence→SILENCE 1.00. The speech formants deliberately avoid the DTMF bands;
the first version landed on a valid pair and classified as a keypress.
Anonymous inbound calls are refused, and endpoint matching is by source
address — Asterisk's default matches the From-header domain, which Hold
Slayer populates from its SIP bind address (0.0.0.0 on a wildcard bind).
Generated audio and the rendered per-host configs are gitignored: the former
is reproducible from a fixed seed, the latter carry a host-specific IP and
the lab password.
Known limit, documented in the README: MediaPipeline.create_tap is a stub, so
the classifier receives no audio on a live call. RTP flows and Asterisk plays
audio, but the tap is never fed — the fixture results above were measured by
feeding the classifier directly. This blocks scenarios 1002/1003/1004.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Add comprehensive deployment and validation plan documenting a staged
bring-up approach that gates each layer on its predecessor and defers
PSTN testing until everything else is proven.
Update .env.example to reflect current configuration:
- Replace static API_TOKEN with Casdoor SSO + owner-minted PAT auth
- Add Rhema TTS settings with port-collision warning
- Add AI Receptionist settings for inbound calls
- Reconcile GATEWAY_SIP_PORT to 5060 and default SIP domain
Add comprehensive rule documentation for AI-assisted development covering
authentication surfaces, outbound-call safety invariants, and other project
conventions to guide Claude's understanding of critical system behaviors.
The UI never sent the bearer token, so with API_TOKEN set every data
call 401'd and /ws/events was rejected pre-accept (the 403s in the
uvicorn log). The API client now keeps the token in localStorage,
attaches Authorization to every request, prompts once on a 401 and
retries, and appends ?token= to the WebSocket connect — the page's
reconnect loop picks the token up after the first prompt.
require_token also accepts a ?token= query parameter (same convention
as the WebSocket) because <audio> elements fetching recordings can't
set headers; recordingUrl() rides the token there.
The dashboard header's status call moved to a new authenticated
GET /api/v1/status — its old source was the JSON root endpoint that
the dashboard itself replaced at /.
Two new auth tests (query-param accepted / wrong query-param 401);
dashboard rebuilt. Verified live: WS rejected without token and
connected with ?token=, status 200, ?token=wrong 401.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The SvelteKit build was always made for the root (no base path); it
now mounts at / — registered last so /api/v1, /ws, /health, and /mcp
match first — and the JSON root endpoint is gone (its info lives in
/health and gateway_status). REST routers move from /api/* to
/api/v1/*; /ws and /health stay put; /mcp/ unchanged. Dashboard API
client, tests, README, and docs updated; dashboard rebuilt (build/ is
gitignored).
Verified live: / serves the UI, /api/v1 answers 200/401, the old
/api paths 404, MCP still lists 15 tools.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The call-flow learner finally gets fed: exploration mode records its
IVR discoveries on the call (ActiveCall.exploration_steps) instead of
throwing them away, persistence stores them in the call record's
metadata, and the rebuilt learn_call_flow MCP tool turns a completed
exploration call into a stored flow via CallFlowLearner — correct
constructor (llm_client from get_llm, heuristic labels when the LLM
is unavailable), build for a new number, merge/refine when a flow
already exists. save_learned_flow/update_flow_from_model keep the
CallFlow↔row mapping in call_persistence.
Test gaps closed: tests/test_learner.py (discoveries→linked steps,
exploration persistence, learn-then-refine through the in-memory MCP
client, no-data and unknown-call answers) and tests/test_websocket.py
(4401 without token, trunk-status-then-replay on connect, per-call
stream filtering).
Docs aligned to code: README (15 tools incl. learn_call_flow, HTTP
not SSE, Python 3.12+, PostgreSQL+Alembic — no SQLite fallback, media
pipeline marked stub-mode until pjsua2 installed, Alembic and honest
/health checked off); docs/mcp-server.md rewritten against the actual
tool surface (hangup not end_call, real params, 3 real resources,
/mcp/ streamable HTTP + bearer auth); architecture/development/
configuration drift fixed.
pyproject: pruned never-imported deps (websockets, librosa,
soundfile, python-multipart).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Alembic replaces create_all as the schema authority: async env.py
against Base.metadata (CLI and in-app entry paths share it via
config.attributes["connection"]), an autogenerated baseline of the
create_all-era schema, and init_db now runs upgrade head — stamping
the baseline first on a pre-Alembic database so existing deployments
adopt cleanly. create_all remains for tests only.
Calls are durable from the start: CallManager gains an
on_call_created hook (wired to persist_call_on_create) that inserts
an in_progress CallRecord the moment a call is created;
persist_call_on_end finalizes that same row. A SIGKILL mid-call now
leaves an in_progress row instead of erasing the call from history
(verified live against the dev database).
One transcript representation: ActiveCall.transcript_chunks holds
TranscriptEntry (t_offset_ms, speaker, text) — add_transcript stamps
real offsets from connect time, receptionist passes speaker instead
of encoding it into "caller: ..." strings, persisted chunks carry
real seek offsets, and the dead CallRecord.transcript Text column is
dropped by migration. Device.is_online migrates String → Boolean
(with a USING cast for existing rows).
Model de-triplication: CallResponse/CallStatusResponse build via
from_call classmethods (one ActiveCall→response mapping);
DeviceStatus deleted — can_receive_call is a computed field on
Device and the list endpoint returns the domain model; all row↔dict
and row↔domain mapping now lives in call_persistence.py
(record_summary/record_detail/chunk_to_dict + device row functions).
New tests/test_data_layer.py: upgrade-head-matches-models,
pre-Alembic adoption, durable in_progress rows, end-without-create
fallback, transcript offsets, consolidated response models.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Engine mode is now explicit: USE_MOCK_SIP=true is the only way to get
the mock engine; an unconfigured trunk fails startup with guidance
instead of silently degrading. Root-caused why the engine always ran
mock: nested pydantic-settings never read .env (no env_file on the
sub-settings classes) — all 8 now declare it.
/health stops lying: reports engine mode (sippy|mock), a live DB
SELECT 1, trunk registration state with reason, and TTS/STT
availability from their last real request; "healthy" now requires
ready + db + sippy + registered trunk.
Error policy: leaf services (tts/transcription/llm_client) raise and
track availability; call-loop callers catch, publish EventType.ERROR
naming the failed service, and apply an explicit fallback. Persistence
writes get one bounded 3x exponential retry, then an ERROR log — no
more silent data loss.
Event bus: a full subscriber queue drops its oldest event (counted)
instead of silently evicting the subscription; subscribe(replay_last=N)
delivers the advertised history replay, used by /ws/events (25).
Receptionist correctness: a matched TAKE_MESSAGE rule beats the LLM;
voicemail polls for early hangup and stops/transcribes/hangs up in
finally; RecordingSession finally keeps its leg_ids so taps detach.
Dead code removed: models/contact.py + Contact table, dtmf_buffer,
transcribe_stream stub, SMS stub in notification.py.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The gateway was the composition root, device registry, inbound-call
policy, and call-operations service in one class, with core↔services
circular imports papered over by function-local imports, wiring done
by assigning private attributes, and MCP tools duplicating REST query
logic against their own sessions.
Composition:
- main.py's lifespan now builds every service and wires them by
constructor/registration. gateway.from_config() is gone; core/ no
longer imports services/ anywhere — the cycle is dead.
- Inbound-call policy moved to ReceptionistService.on_inbound_call
(routing evaluation, reject/answer, screening dispatch); wired as
the engine's on_incoming_call by the lifespan. Receptionist deps
(tts/transcription/recording/routing) are constructor-injected —
no more gateway._tts reach-through or importing hold_slayer's
private _get_llm (now services.llm_client.get_llm, shared).
- Hold-slayer launch goes through a mode-handler registry
(register_mode_handler); the gateway no longer knows the service's
type. CallManager takes on_call_ended in its constructor.
- build_sip_engine() is a pure function taking explicit callbacks.
- api/routing.py uses the routing service from app.state via a
proper dependency instead of gateway._routing.
Shared data layer:
- db.session_scope() is the one session convention (get_db wraps it).
- services/call_persistence.py gains the query/write functions and
the single StoredCallFlow→CallFlow mapper; api/call_flows.py,
api/call_history.py, and the six DB-touching MCP tools are thin
wrappers over them — the two surfaces can't drift.
- legs_for_call() replaces the three private _call_legs scans
(gateway transfer/hangup, REST dtmf, MCP dtmf).
7 new tests (mode-handler launch, on_call_ended hook, receptionist
inbound answer/reject, call-flow CRUD round-trip and history routes
against real SQLite through the shared layer). aiosqlite added to dev
deps for that.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three thread domains were mutating shared dicts with no locks: Sippy's
ED thread wrote _legs/_registered_devices directly from SIP handlers,
the asyncio loop wrote them from make_call/hangup, and
run_in_executor(None, ...) had default-pool threads driving sippy UA
objects. AudioTap.feed() pushed into an asyncio.Queue (not
thread-safe) from the PJSUA2 thread.
New ownership rule, enforced structurally:
- The asyncio loop owns all app-visible state; the only mutator is the
new _on_engine_event funnel. Sippy handlers extract plain strings on
the ED thread and post via run_coroutine_threadsafe.
- The ED thread owns sippy objects plus _ed_ua_to_leg/_ed_leg_to_ua;
loop-side commands (INVITE/BYE/DTMF/trunk register) hop over via
ED2.callFromThread. UA references no longer live on SipCallLeg.
- AudioTap captures its loop and feed() hops via call_soon_threadsafe.
- Fix ED import: installed sippy 2.x exposes ED2, not ED — the old
import could never start the event loop.
Also:
- Wire the never-connected on_leg_state_change callback: outbound
ringing/connected/terminated now reaches CallManager; a call ends
when its last leg terminates (transfers keep it alive). Adds
CallManager.unmap_leg/legs_for_call.
- AudioClassifier.classify(): async entry that runs the FFT work in
asyncio.to_thread and updates history on the loop — all four
hold_slayer call sites now route through it, fixing both the
loop-blocking and the 2-of-4 history gap. DTMF Goertzel loop
replaced by the equivalent vectorized DFT-bin power.
- Task hygiene: gateway.spawn() tracks hold-slayer/receptionist tasks
and stop() cancels them; recording safety-timeout task is retained
and cancelled on stop_recording; engine tracks incoming-call
dispatch tasks.
10 new tests: funnel events from a foreign thread, auto-answer
fallback, AudioTap cross-thread feed, classifier history, leg-state →
call status (including no stomping of ON_HOLD), stop() cancellation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The MCP server was created but never mounted — no client could reach
it. Mount it at /mcp/ over streamable HTTP with a combined lifespan,
resolving the gateway lazily so mounting happens at app construction.
Security and safety for the agent surface:
- One static API_TOKEN (SecretStr) enforced across REST (dependency),
WebSocket (query param/header before accept), and MCP
(StaticTokenVerifier). Startup refuses tokenless non-loopback binds.
- Emergency numbers (911/9911/112) always refused on make_call, plus a
MAX_CONCURRENT_CALLS cap; ValueError surfaces as 400/ToolError.
- Safe defaults: debug off, no credential in default DATABASE_URL,
SIP/LLM/TTS secrets as SecretStr.
Cleanups:
- Delete broken learn_call_flow tool (wrong ctor args, nonexistent
method) and the never-fed CallAnalytics service; keep
call_flow_learner for proper wiring later.
- Trim dial_plan to what is actually used (emergency guard, extension
allocation); delete the unreferenced matcher/normaliser.
- Register call_history before calls so /api/calls/history is no
longer shadowed by /api/calls/{call_id}.
- fastmcp pinned >=3.0 (http_app + StaticTokenVerifier).
New tests: MCP in-memory client (tool surface, lazy gateway, emergency
refusal, call cap) and API security (401 paths, route order, mount).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds read-only access to persisted call records for the dashboard
and implements a client for the Rhema text-to-speech service.
- api/call_history.py: New router providing paged call lists
and detailed call records with transcript metadata.
- services/tts.py: Async client for OpenAI-compatible TTS
endpoints (Rhema/Kokoro) used for call-flow steps.
Complete project scaffolding and core implementation of an AI-powered
telephony system that calls companies, navigates IVR menus, waits on
hold, and transfers to the user when a human answers.
Key components:
- FastAPI server with REST API, WebSocket, and MCP (SSE) interfaces
- SIP/VoIP call management via PJSUA2 with RTP audio streaming
- LLM-powered IVR navigation using OpenAI/Anthropic with tool calling
- Hold detection service combining audio analysis and silence detection
- Real-time STT (Whisper/Deepgram) and TTS (OpenAI/Piper) pipelines
- Call recording with per-channel and mixed audio capture
- Event bus (asyncio pub/sub) for real-time client updates
- Web dashboard with live call monitoring
- SQLite persistence via SQLAlchemy with call history and analytics
- Notification support (email, SMS, webhook, desktop)
- Docker Compose deployment with Opal VoIP and Opal Media containers
- Comprehensive test suite with unit, integration, and E2E tests
- Simplified .gitignore and full project documentation in README
