Blanket per-endpoint limits would have been the wrong shape here. Every
REST/WS/MCP surface is owner-only — an unauthenticated request is rejected by
resolve_bearer/is_owner before any handler runs — so limiting them would
mostly throttle the single legitimate operator, and real spend control for
outbound calls is already max_concurrent_calls in gateway.make_call.
What is genuinely exposed is the handful of /auth/* routes that must answer
before an identity exists. /auth/callback and /auth/refresh-callback each make
an outbound token exchange with Casdoor on every request; /auth/me opens a DB
session and runs a token lookup. All are free to trigger and none are cheap to
serve. /auth/logout is left unlimited — it builds a redirect URL and does no
I/O.
Not a defence against credential guessing: PATs are secrets.token_urlsafe(32)
(256 bits) compared by SHA-256 digest, so brute force was never the threat.
This is about unauthenticated work an attacker controls.
Fixed-window, in-process, no new dependency — one operator and one process
make a shared counter store infrastructure without a purpose. The bucket store
is bounded and evicts oldest-first, since an unbounded map keyed by source
address would itself be the exhaustion vector.
The limiter keys on the socket peer and deliberately ignores X-Forwarded-For.
That header is attacker-controlled unless a trusted proxy overwrites it, and
this app establishes no such trust; keying on it would let one client present
as thousands and make the limiter worse than useless. Behind the estate's
reverse proxy the limit is therefore per-proxy, not per-caller — correct for
exhaustion and honest about what it can enforce. Per-caller limits need an
explicit trusted-proxy config, noted in CLAUDE.md so it isn't added silently.
Verified against a real server: exactly 30 requests pass, then 429 with
Retry-After: 60, while an owner-gated route serves 40/40. The 429s appear in
the JSON access log with queryable status_code and client_addr, so an attack
is visible in Loki. The wiring test identifies the dependency by qualname
rather than string search, and was mutation-checked by removing the limit from
/auth/me.
Also documents 401/403/429 in the API reference — 401 and 403 have existed
since auth landed but were never in the status-code table. Phase 4 is now
complete.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Started as the SIP_ENGINE row flagged in the last commit. Cross-checking the
tables against config.py mechanically (rather than by eye) turned up more,
including two entries that were actively wrong.
Corrections:
- GATEWAY_RTP_PORT_MIN/MAX and GATEWAY_HOST are documented in
configuration.md but do not exist — no code reads them and they are absent
from .env.example. Setting them today does nothing. Replaced with the real
GATEWAY_SIP_ fields (host/port/domain).
- GATEWAY_SIP_PORT was documented as 5080 in two places; the code and
.env.example both say 5060.
- DATABASE_URL was documented with a SQLite default. There is none, and
startup exits if it is unset.
Additions — every env var the models accept is now documented somewhere
(verified bidirectionally: nothing in the models undocumented, nothing
documented that the models reject):
- Server section: HOST, PORT, DEBUG, LOG_LEVEL, LOG_FORMAT
- Safety section: MAX_CONCURRENT_CALLS, USE_MOCK_SIP, SIP_ENGINE
- Receptionist section (configuration.md had none, though seven vars exist)
Structural staleness, from the PR #8 media-plane work:
- core/pjsua_engine.py was absent from the component list and file tree; so
were dial_plan.py (the emergency guard) and sip_engine.py.
- architecture.md's banner still read "media plane in transition". The engine
landed; it is now two selectable engines with the audio consequence stated.
- Tech Stack described "single-process async architecture" — the
simplification CLAUDE.md explicitly calls out. Now points at the threading
model, since there are three execution contexts.
- The Asterisk lab shipped in PR #8 with its own README but nothing linked to
it. Linked from the test section and both doc indexes.
- CLAUDE.md's "no structured JSON logging" gap is closed; test count was 146
across 16 files, now 189 across 19. The other listed gaps (no /metrics, no
rate limiting, no health-probe log filter) were re-verified and still hold.
Deliberately not hardcoding a test count in the README — that is the same
staleness this commit is clearing up. All internal links and anchors verified
to resolve; 189 tests pass; lint unchanged at its 216 baseline.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Implements the tap half of the media path and records why the other half
requires moving call placement into PJSUA2.
MediaPipeline.create_tap was a stub: it logged "🎤 Audio tap created" and
returned a tap that nothing ever fed, so the classifier received no audio on
a live call. It now builds a real pj.AudioMediaPort subclass whose
onFrameReceived converts the SWIG ByteVector to PCM bytes and fans it out to
every tap on the stream.
One capture port per stream, shared by all taps: a second port on the same
stream would be mixed back into the conference bridge and the call would echo.
Thread safety is the constraint here. onFrameReceived runs on a PJSUA2 worker
thread — a third execution context beside the asyncio loop and the Sippy ED
thread — and touches nothing but AudioTap.feed, which hops to the owning loop
via call_soon_threadsafe. An exception escaping into PJSUA2's C++ callback
would tear down the worker thread and silently kill media for every call, so
the handler catches and logs once per port rather than on every 20ms frame.
Also fixes a hard crash found while testing this against real PJSUA2: a media
port finalised after Endpoint.libDestroy() calls pjmedia_conf_remove_port
against a freed conference bridge and aborts the process on a native
assertion. Ports are now released in remove_stream while the bridge still
exists, and stop() forces a collection before libDestroy — dropping the last
Python reference is not sufficient on its own.
Verified against the real bindings: frames fan out to multiple taps, cross the
thread boundary intact, and shutdown is clean.
add_remote_stream remains a stub, and deliberately so. PJSUA2 exposes no
standalone RTP media object — every AudioMedia subclass in the Python
bindings is a file player, recorder, tone generator or capture port, and RTP
is reachable only via pj.Call.getAudioMedia() on a dialog PJSUA2 itself owns.
A design where Sippy owns the dialog can never obtain media from PJSUA2, so
that function cannot be written against this API. docs/architecture.md now
explains this and records the resolution: PJSUA2 places the trunk call while
Sippy keeps the SBC roles (device registration, routing, leg bridging), with
the emergency guard and concurrency cap staying first in gateway.make_call
regardless of which library dials.
The architecture doc also had drift unrelated to media: it described the
thread boundary as asyncio.run_in_executor() when the real mechanism is
run_coroutine_threadsafe / ED2.callFromThread, claimed two execution contexts
where there are three, and cited MediaPipeline.add_stream() and
SippyEngine.bridge() — neither of which exists. Corrected, with the data flow
now showing the emergency guard and concurrency cap in their real positions.
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>
Update deployment validation plan to reflect Phase 1b completion — pjsua2
built from pjproject 2.17 on caliban without sudo. Document the non-obvious
RPATH/patchelf step, the status() method name correction, and two remaining
caveats (not captured by pip install, Docker still runs stub media).
Update README to point at the new docs/pjsua2-build.md and clarify stub-mode
behavior.
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 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>
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
