Pedido do usuário: "adiciona upload de áudio pro prompt do IVR" — até
aqui o prompt era só texto livre (na prática, sempre um tom padrão,
nunca voz de verdade).
`POST /ivr-menus/:id/prompt` (multipart via @fastify/multipart —
primeiro upload de arquivo binário desta API) só aceita WAV (cabeçalho
RIFF/WAVE validado antes de gravar; esta implantação do FreeSWITCH não
tem mod_shout, então MP3 nunca funcionaria de qualquer forma). Gravado
num bind mount NOVO (./data/ivr-prompts no host ↔ /ivr-prompts no
container freeswitch) — mesma convenção já usada 3x neste projeto pra
arquivo que o FreeSWITCH precisa enxergar de verdade (gateways externos,
filas do callcenter, spool de gravação), mas na direção contrária:
apps/api (host) escreve o que o usuário sobe, o FreeSWITCH lê ao vivo
durante play_and_get_digits. Um fetch em rede (S3/HTTP) durante uma
chamada ativa foi descartado de propósito — latência/confiabilidade
desnecessárias pra um prompt de poucos segundos.
GET /ivr-menus/:id/prompt (autenticado) serve o preview — mesmo
princípio do player de gravações, nunca uma URL direta pro storage.
Achado real corrigido antes de commitar: minha primeira versão do
delete de menu deixava o .wav órfão no disco — agora deletar o menu ou
trocar/remover o prompt sempre limpa o arquivo, confirmado com um teste
real de upload+delete.
Tela "Telefonia > IVR" ganhou upload/troca/remoção de áudio por menu +
player de preview.
Testado ponta a ponta com um WAV real de 44.1kHz/mono (não o formato
"nativo" de telefonia, de propósito, pra confirmar que funciona com o
que uma pessoa qualquer gravaria): softphone externo discou o DID,
play_and_get_digits abriu e tocou o arquivo até o fim duas vezes
(mod_sndfile resample automático, sem erro no log), colheu o dígito
real e bridged corretamente com o ramal de destino.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaBaBSQGhyXGEgtTYZGV8
Usuário tentou registrar um ramal e não conseguiu — nenhuma porta SIP/RTP
estava publicada no host, só o Event Socket (interno). O teste ponta a
ponta da fase anterior só funcionava porque os softphones de teste
rodavam dentro da mesma rede Docker.
RTP restrito a um range fixo de 200 portas (16384-16584, via sed em
switch.conf.xml) — o default vanilla (~16k portas) é inviável de publicar
uma a uma. docker-compose.yml publica 5060/udp+tcp (SIP) e
16384-16584/udp (RTP); Event Socket continua nunca publicado.
Verificado: iptables -t nat -L DOCKER confirma DNAT correto pras 201
portas; sofia status confirma que o STUN já configurado (external_rtp_ip/
external_sip_ip) resolve pro IP público real da VM, então o SDP vai
anunciar o IP certo — não só o registro, o áudio também deve funcionar.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaBaBSQGhyXGEgtTYZGV8
Sub-fase B do modulo de IA: novo servico apps/ai-worker (poll + FOR UPDATE
SKIP LOCKED) processa AIJob de TRANSCRIPTION/ANALYSIS disparados
automaticamente apos uma gravacao ficar disponivel, respeitando a cascata
de privacidade Tenant>Queue>Campaign e o entitlement do Plan. Transcricao
separa o WAV estereo em 2 canais (parser proprio, sem ffmpeg) e transcreve
cada perna independente; analise sempre redige dados sensiveis antes de
sair pro provider e valida o resultado contra o schema antes de persistir.
CRUD de AIPromptTemplate/AIPromptVersion em apps/api.
Testado ponta a ponta contra o worker real em Docker e Postgres real com
RLS (cascata de privacidade em 3 cenarios, WAV sintetico real no object
storage, claim/retry/dead-letter reais) — chamada de rede contra
OpenAI/Anthropic continua nunca exercitada (mesma restricao de rede desde
o Provider Layer). Detalhes em docs/AI_PIPELINE.md.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X1HxY46WGU4G1zmVDNKcWw
Fecha agente.md secao 90-94. A especificacao lista "Recording" e "Object
Storage" como dois passos separados na ordem de implementacao (secao
232), mas ficaram numa unica fase — sao acoplados o suficiente (Recording
precisa de um lugar pra guardar bytes) pra fazer sentido construir juntos.
## packages/storage — ObjectStorageProvider (secao 92)
Abstracao pequena: putObject/getObjectStream/deleteObject. Dois backends:
LocalObjectStorageProvider (filesystem, com checagem de path traversal
mesmo a key sendo sempre montada no servidor) e S3ObjectStorageProvider
(@aws-sdk/client-s3, preparado pra AWS S3 e MinIO via endpoint/
forcePathStyle customizaveis — nunca exercitado nesta sessao, sem
servidor S3 disponivel neste laboratorio). Escolhido por STORAGE_PROVIDER
env.
buildRecordingObjectKey (secao 93):
tenants/{tenant_id}/recordings/YYYY/MM/DD/{call_id}.wav, sempre montada
no servidor a partir de dados confiaveis.
## Bind mounts, nao volumes nomeados
/recordings e /data/object-storage usam bind mount pra um diretorio real
do host — apps/api roda no host, nao em Docker, e precisa enxergar os
mesmos arquivos que fs-events escreve. LOCAL_STORAGE_ROOT tem valores
diferentes por ambiente (mesmo padrao ja usado pra REDIS_URL).
## Quem grava: apps/predictive-dialer
So' chamadas originadas pelo discador com Campaign.recordingEnabled sao
gravadas nesta fase (unico caminho de originate que o sistema controla
hoje). RECORD_STEREO=true + execute_on_answer='record_session ...'
adicionados ao originate; origination_uuid pre-gerado (em vez de deixar o
provider sortear) porque o path de gravacao precisa dele antes do
comando de originate ser montado — o mesmo uuid vira Call.id no CDR.
## Quem sobe: apps/freeswitch-events/src/recording.ts
Em CALL_ENDED, encadeado depois do persistCallEvent terminar (nao em
paralelo) — uploadRecordingIfPresent le Call.talkTime/durationSeconds,
que e' exatamente o que persistCallEvent acabou de calcular no mesmo
evento (mesma classe de corrida ja corrigida uma vez na fase CDR, aqui
evitada por ordenacao). Sobe pro storage, cria Recording (retentionUntil
a partir de Plan.recordingRetentionDays), apaga o spool local.
## API + retencao
GET /recordings, GET /recordings/:id, GET /recordings/:id/audio (stream
autenticado, nunca URL direta pro storage). runRetentionSweep (secao 94)
no boot do apps/api + a cada hora — apaga o objeto, marca status=DELETED
(linha nunca apagada, fica como auditoria).
## Bug real achado testando esta fase
Recording.sizeBytes (BigInt) quebrava GET /recordings com 500 — Fastify
nao serializa BigInt nativamente (mesma classe de bug ja corrigida uma
vez no logger, fase Event Socket). Corrigido convertendo pra number na
resposta.
Verificado ponta a ponta: gravacao real criada (RIFF WAVE, PCM 16-bit,
ESTEREO 8000Hz — RECORD_STEREO confirmado), upload com path exato da
secao 93, download via API com md5 identico ao objeto original, varredura
de retencao apagando objeto + status DELETED + list/download bloqueados
depois. typecheck do workspace inteiro limpo.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X1HxY46WGU4G1zmVDNKcWw
Fecha agente.md secao 72-86 (motor preditivo) e 77-79 (CPS distribuido,
reserva de leads, lock de campanha). Uma campanha RUNNING agora origina
chamadas sozinha, respeitando capacidade de agentes, CPS hierarquico e
taxa de abandono — sem intervencao manual.
Deliberadamente fora do escopo (agente.md secao 72: "nao e' so' `for lead
-> originate`"): mod_avmd (opcional), callbacks agendados, disposicoes de
agente — ficam pra fase CDR.
## Novo servico apps/predictive-dialer
Mesmo padrao arquitetural de fs-events/fs-config: Node standalone em
Docker, ESL propria, tick a cada 2s sobre tenants ativos x campanhas
RUNNING/WAITING_SCHEDULE.
- Lock de campanha (dialer:campaign:{id}, secao 79): TTL/ownership/
renewal/safe-release via Lua compare-and-delete.
- CPS distribuido (secao 77, 62): token bucket janela 1s, hierarquia
GLOBAL/TENANT/TRUNK/CAMPAIGN numa unica chamada Lua atomica — nivel
esgotado bloqueia todos SEM incremento parcial dos que passariam.
- Reserva atomica de leads (secao 78): FOR UPDATE SKIP LOCKED dentro da
mesma transacao withTenantContext.
- CallAttempt/CampaignStats (schema novo): state machine da chamada
(secao 82) + EWMA (secao 75) de answer_probability/average_answer_delay/
average_talk_time/abandon_rate por campanha.
- Capacidade em tempo real + pacing (secao 73-76, 84-85): conta agentes
por estado via Tier->Agent.state, previsao de liberacao (horizonte
unico de 15s, simplificacao documentada dos 4 buckets da especificacao),
controle de abandono reduz pacing progressivamente, nunca origina sem
capacidade prevista.
## Modo simulacao (secao 185-186)
DIALER_SIMULATION=true (default, ja estava no .env desde o inicio da
sessao) sorteia ANSWER/BUSY/NO_ANSWER/FAILED em software, sem PSTN real.
So' quando ANSWERED e' que uma chamada sintetica (null/dummy, sem PSTN)
entra na fila real via mod_callcenter de verdade — escolha deliberada pra
maximizar codigo real exercitado em vez de simular tudo em memoria. Os
identificadores da secao 81 (b2bcall_tenant_id/call_id/attempt_id/
campaign_id/lead_id) vao como channel variables nessa perna, entregando
tenantId real no WebSocket sem fan-out.
Real Outbound Safety (secao 186): as duas flags checadas no boot, nunca
ativadas automaticamente — caminho PSTN real implementado mas nunca
exercitado (sem trunk/operadora real neste laboratorio).
## Dois bugs reais achados e corrigidos testando esta fase
- Perna sintetica (null/dummy) nao tem midia do outro lado — nunca
desligava sozinha depois de bridgear com um agente. Corrigido com
hangup agendado via uuid_kill no talk_time simulado.
- Corrida entre queue:sync e tier:sync (dois canais Redis independentes,
sem ordem garantida): atribuir tier logo depois de criar a fila podia
rodar tier add antes do queue reload terminar ("-ERR Queue not found!",
erro real, diferente do ja conhecido "already exist"). Corrigido com
retry curto (ate 3 tentativas) em agent-sync.ts::addTierWithRetry.
## GET /campaigns/:id/stats
Secao 227.7 "visualizar pacing" — CampaignStats + agentes por estado +
calls em andamento, sem esperar a fase Frontend.
Verificado ponta a ponta: campanha RUNNING originando 3 tentativas por
tick, outcomes simulados corretos com retry agendado (BUSY 15min/
NO_ANSWER 60min/FAILED 30min), uma tentativa ANSWERED completando o ciclo
real inteiro (fila -> agente -> bridge -> hangup -> EWMA atualizada),
stop nao derruba chamada ativa (secao 66), calls_answered=3 confirmado no
`queue list` do FreeSWITCH. CPS limiter e lock de campanha testados
isoladamente (hierarquia sem incremento parcial, ownership nunca
roubado). typecheck do workspace inteiro limpo.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X1HxY46WGU4G1zmVDNKcWw
- Investigated the real callcenter_config command surface via
'help callcenter_config' on the running FreeSWITCH before writing any
code: queues only have load/unload/reload (static XML + reload, no
'queue add' exists), while agents and tiers are fully dynamic via ESL
commands (agent add, tier add) -- no file involved. This shapes the next
phase (Agents/Tiers) differently from this one.
- queues table (tenant-scoped, RLS): strategy, moh/announce, wait times,
tier rules, discard/abandoned handling, skip-agents-with-external-calls,
recording_enabled
- packages/telephony: buildQueueXml()
- infrastructure/freeswitch: our own callcenter.conf.xml override (empties
the vanilla static agents/tiers -- those become fully dynamic in the next
phase) that includes callcenter_queues.conf.d/*.xml via X-PRE-PROCESS,
same pattern as the Sofia gateway directory
- apps/api/src/queues: CRUD (POST/GET/GET:id/DELETE) using the queues.view/
.manage permissions already in the seed
- b2bcall-fs-config (queue-sync.ts): one XML file per queue on a shared
volume, synced via Redis pub/sub (b2bcall:queues:sync) on create/delete
and once at boot -- same shape as trunk-sync.ts
- confirmed manually against the real FreeSWITCH, before coding the sync
logic: 'queue load <name>' fails ('Invalid Queue not found!') for a file
added after boot -- needs 'reloadxml' first to repopulate the in-memory
XML tree from disk; after that, 'queue reload <name>' alone handles both
create and update, no need to distinguish load vs reload
- verified end-to-end: created a queue (ROUND_ROBIN, maxWaitTime=120,
discardAbandonedAfter=90), 'callcenter_config queue list' showed the
correct values on the FreeSWITCH side; deleted it, list went back empty
docs/QUEUES.md
- trunks table (tenant-scoped, RLS): host/proxy/realm, register,
username/password_enc (AES-256-GCM), dtmf_mode, ping, transport, and a
status/status_updated_at pair meant to be driven by FreeSWITCH events
- apps/api/src/trunks: CRUD (POST/GET/GET:id/DELETE), same RBAC/tenant
pattern as Extensions, password never exposed in any GET
- packages/telephony: buildGatewayXml() generates a Sofia gateway XML file
- b2bcall-fs-config now writes sip_profiles/external/<trunk_id>.xml (shared
Docker volume with FreeSWITCH -- the vanilla external profile already
includes external/*.xml) and runs 'sofia profile external rescan' over
ESL; syncs on boot and on demand via Redis pub/sub
(b2bcall:trunks:sync), since apps/api runs on the host and fs-config has
no port published to reach directly
- added FreeSwitchTelephonyProvider.waitUntilConnected() to fix a startup
race: the first sync ran before the ESL connection had settled, logging
a harmless but noisy error
- verified end-to-end with a fake host: create trunk -> gateway file
written -> FreeSWITCH shows the real gateway (FAIL_WAIT, expected) ->
delete -> file removed (cleanup also correctly swept the stale
'example.com' gateway that had been copied into the volume from the
vanilla image)
- apps/freeswitch-events/src/trunk-status.ts: written to update Trunk.status
from sofia::gateway_state events, using the same normalizeEslEvent path
already proven for CHANNEL_* events
- KNOWN GAP, documented rather than glossed over: monitored fs-events for
~90s while the gateway visibly transitioned states in FreeSWITCH
(FAIL_WAIT/DOWN) and no sofia::gateway_state event was observed arriving.
CUSTOM/sofia::* events have not actually been proven working end-to-end
in this session -- only CHANNEL_* events have been. Needs verification
against a real SIP target before the status auto-update can be trusted
in production. See docs/TRUNKS.md and TODO.md.
- docs/TRUNKS.md
- extensions table (tenant-scoped, RLS): number, sip_password_enc
(AES-256-GCM via packages/shared/src/crypto.ts), caller_id, context,
sofia_profile, codecs, max_registrations
- apps/api/src/extensions: CRUD (POST/GET/GET:id/DELETE), protected by a
new generic PermissionGuard (@RequirePermission decorator), tenant
resolved only from the JWT (never trusted from the client)
- SIP password is returned in plaintext only once, in the create response;
toPublicExtension() explicitly destructures the encrypted field out
(not a spread) so it can't leak by accident
- b2bcall-fs-config now resolves real directory data: Tenant.telephonyDomain
-> Extension.number, decrypts the password, builds proper directory XML
including a dial-string param (missing it caused originate to fail with
MANDATORY_IE_MISSING instead of the expected USER_NOT_REGISTERED)
- pinned FreeSWITCH's 357737{domain} to a stable value (b2bcall.local) via a
vars.xml patch in the Dockerfile -- it previously used the container's
dynamic IP, which could never match a stored telephony_domain
- added HTTP Basic auth between FreeSWITCH and fs-config
(gateway-credentials, timingSafeEqual comparison) now that the service
returns real secret data, closing the gap flagged as pending in the XML
Curl phase instead of leaving it open
- found and fixed: PermissionGuard's constructor-injected Reflector came
back undefined at runtime under tsx/esbuild (unreliable cross-file
decorator metadata emission) -- fixed with an explicit @Inject(Reflector);
worth watching for in future guards/services run via tsx
- verified end-to-end: create extension -> originate user/<ext> reports
USER_NOT_REGISTERED (found, not registered) -> delete -> back to
SUBSCRIBER_ABSENT (not found); password never reappears in any GET;
unauthenticated fs-config requests get 401
- docs/EXTENSIONS.md
- apps/freeswitch-config (b2bcall-fs-config): Fastify service implementing
the mod_xml_curl HTTP protocol (form-encoded POST -> XML response),
containerized, no host port published
- reactivated mod_xml_curl in FreeSWITCH, binding restricted to
directory|dialplan only (configuration was removed after testing showed
it firing several unnecessary HTTP round-trips at boot for module
configs we don't need dynamic — matches agente.md's own 'don't put every
critical config through XML Curl' guidance)
- no extensions/dialplan tables exist yet (next phases), so the service
always answers 'not found' for now — this phase only proves the wire
protocol works without breaking the static vanilla config fallback
- verified end-to-end: user/8888 (nowhere) -> SUBSCRIBER_ABSENT via
fs-config; user/1000 (static vanilla extension) -> USER_NOT_REGISTERED,
proving FreeSWITCH correctly falls through to static XML when xml_curl
says not found
- docs/XML_CURL.md, including the not-yet-authenticated endpoint note (fine
while it only returns not-found; needs gateway-credentials before serving
real directory/dialplan data)
- packages/telephony: TelephonyProvider interface (agente.md secao 25) and
FreeSwitchTelephonyProvider implementation over the 'esl' library
(actively maintained, TypeScript-native, built-in reconnect-with-backoff
satisfying secao 195); normalizeEslEvent() translates raw ESL events into
the internal vocabulary (secao 24)
- apps/freeswitch-events (b2bcall-fs-events): permanent ESL connection,
resubscribes on every reconnect, publishes normalized events to the
'b2bcall:events' Redis pub/sub channel; containerized (Dockerfile +
docker-compose service) since its whole job is reaching the freeswitch
container by internal hostname
- packages/shared: reusable createLogger() (structured JSON per secao 189),
fixed a BigInt serialization crash surfaced by the esl library's error
stats
- found and fixed a real FreeSWITCH 1.11 default: without an explicit
apply-inbound-acl, mod_event_socket silently rejects any non-loopback
connection ('Access Denied, go away.') even with the correct password —
added a dedicated ACL (loopback + the Docker Compose network range, never
0.0.0.0/0) in infrastructure/freeswitch/overrides/autoload_configs/
- verified end-to-end with a local loopback test call: CALL_CREATED ->
CALL_ANSWERED -> CALL_ENDED observed on the Redis channel with the
correct callUuid and hangup cause
- docs/EVENT_SOCKET.md
- infrastructure/freeswitch/Dockerfile: debian:trixie-slim + SignalWire
packaged freeswitch-meta-vanilla, avoiding a C/C++ build on a 1.9GB RAM VM
- FREESWITCH_PAT used only via Docker BuildKit secret, apt credentials file
created and deleted within the same RUN — verified absent from the final
image with docker history
- minimal module set (agente.md secao 15): sofia, event_socket, commands,
dptools, callcenter, avmd, curl, local_stream, etc. mod_xml_curl installed
but disabled — it refuses to load without a configured gateway-url, which
will exist once b2bcall-fs-config is built
- entrypoint.sh rotates the Event Socket password away from the 'ClueCon'
default at container runtime (never baked into the image); fails loudly if
ESL_PASSWORD is unset
- port 8021 not published to the host; only reachable from other containers
on the compose network
- found and fixed: freeswitch-conf-vanilla is a Recommends (not a Depends)
of freeswitch-meta-vanilla, so --no-install-recommends silently produced
an empty /etc/freeswitch and a crash loop
- verified end-to-end: fs_cli status via ESL with the custom password,
default password rejected, expected modules loaded, healthcheck green,
~44MB RAM usage
- docs/FREESWITCH.md, docs/NETWORK_ARCHITECTURE.md (network_mode decision
deferred until a real SIP trunk exists)
