* fix(gather): clear a finalized deepgram interim word so UtteranceEnd is not deferred forever
The #1088 check defers UtteranceEnd while an interim word newer than
last_word_end is pending. The marker was only cleared by a non-empty
is_final, so an interim word that Deepgram later dropped (e.g. line
noise, followed by an empty final) left it set: the gather waited for
the caller to speak again and merged both utterances.
Clear the marker on any Deepgram is_final result whose window
(start + duration) covers it. A word that Deepgram carries into the
next segment still defers UtteranceEnd.
Port of jambonz/feature-server#223.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(gather): return the buffer when a deferred deepgram UtteranceEnd is satisfied
Clearing the unprocessed-word marker only helps if the covering final
arrives before UtteranceEnd. If it arrives after, UtteranceEnd has
already been deferred and Deepgram sends no further UtteranceEnd until
new speech, so the next utterance is still merged in.
Remember a deferred UtteranceEnd, and when a later final clears the
marker, return the buffered transcript once that final has been
handled (setImmediate, since _onTranscription has early returns).
Port of jambonz/feature-server#225.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(gather): only an empty final ends a gather early after a deferred UtteranceEnd
A final with words means the caller resumed talking after the pause;
Deepgram will send a new UtteranceEnd after those words, so keep the
existing flow instead of cutting the caller off mid-sentence (or
skipping the asrTimeout wait). Only a final that drops the pending word
(empty) leaves nothing else to end the gather.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The ws v8 library throws a plain Error('Opening handshake has timed out') with no
.code and name 'Error' when the handshake timeout (JAMBONES_WS_HANDSHAKE_TIMEOUT_MS,
default 1500ms) elapses. BaseRequestor._shouldRetry classified retryable errors only
by .code/.name/.statusCode, so this error matched neither the ct nor rt bucket and was
never retried under the default rp=ct policy — the call failed immediately on the first
attempt, silently bypassing maxReconnects.
Match the handshake-timeout message in the ct bucket so a transient slow WS upgrade
(e.g. a serverless/edge backend cold start) is retried as intended. Adds tape coverage
for both the default-ct retry path and the rp=4xx no-retry path.
_lccTtsTokens silently returned on a missing `id`, leaving the WS client with
no feedback — no tts:tokens-result, no audio, only a server-side info log.
Mirror the existing missing-`tokens` handling and reply with status 'failed' /
reason 'missing id' so callers that omit `id` (e.g. the Python SDK) get an
actionable result.
* fix: fire sip:refer actionHook when the far end BYEs before any NOTIFY
When a REFER is accepted with 202 the verb waits for a NOTIFY carrying the
final status of the referred call, and only fired the actionHook from the
15s timeout callback. If the far end sends BYE straight after the 202 the
call session kills the task, awaitTaskDone() resolves, and exec returned
without ever performing the action.
Perform the action once from exec on every exit path - before the session
tears the requestor down - and de-duplicate it against the final-NOTIFY path.
Fix#1258
Generated by Ora Studio
Vibe coded by ousamabenyounes
Co-Authored-By: Ora Agent <noreply@oratelecom.net>
* fix: keep final_referred_call_status when a BYE races the final NOTIFY (#1580)
A final NOTIFY (status >= 200) whose eventHook round trip is still in flight
when the far end BYEs lost final_referred_call_status on the actionHook: kill()
woke exec(), which memoised the action via _performReferAction({refer_status})
without the final status, so the NOTIFY path's later call carrying it was
deduped away. On main the app received final_referred_call_status here.
Record the final status synchronously in _handleNotify before the eventHook
await, and include it from exec() so the memoised action carries it regardless
of which exit path wins the race. Extract the 200 threshold into a named
constant while touching those lines.
Addresses davehorton's review on #1580.
Generated by Ora Studio
Vibe coded by ousamabenyounes
Co-Authored-By: Ora Agent <noreply@oratelecom.net>
---------
Co-authored-by: Ora Agent <noreply@oratelecom.net>
Co-authored-by: Ben Younes <2910651+ousamabenyounes@users.noreply.github.com>
A call transferred between feature servers arrives at the receiving server as a
fresh INVITE, so the session runs through trying, ringing, early-media and
answered again as it accepts the REFER. The application was already told all of
that by the server that first handled the call, so it sees a second round of
events for a call it believes is long since answered.
Skip only the application-facing status callback for those four statuses when
the session is a transferred call. callInfo, the redis call record and
record-all-calls still run exactly as before, so the call record stays accurate
and answering a transferred call still starts recording.
Fixes#1392
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
_onTtsStreamingResume asked for "streaming-event" while stream_open,
stream_paused, stream_closed and user_interruption all ask for
"/streaming-event".
That one missing slash decides whether the event reaches an HTTP application at
all. HttpRequestor joins a hook path to the app's baseUrl only when it is
relative, and relative means starting with a slash:
_isRelativeUrl(u) { return typeof u === 'string' && u.startsWith('/'); }
const absUrl = this._isRelativeUrl(url) ? `${this.baseUrl}${url}` : url;
"streaming-event" is neither relative nor absolute, so it is passed through
unjoined and never lands on the application. The call site catches and logs the
failure rather than raising it, so an application simply sees stream_paused with
no matching stream_resumed and nothing anywhere says why.
Present since the verb was introduced in #994. Found by reading the file, so
there is no issue to close.
Tests: the two new cases fail against the unpatched line and pass with it.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This repo is public and held a long-lived AWS access key as repository secrets
(set 2023-11-22). It is replaced with short-lived credentials from the GitHub OIDC
provider; no AWS key and no account id remain in the repo.
create-test-db.js wrote {access_key_id, secret_access_key, aws_region} into the test
database as the aws speech credential, which sends speech-utils' getAwsAuthToken down
its access-key branch and calls GetSessionToken -- rejected by AWS for session
credentials. The role_arn branch calls AssumeRole instead, which accepts them, and is
already plumbed through db-utils.js, call-session.js and stt-task.js. The pinned
speech-utils 0.2.30 already supports it, so no dependency change is needed.
Fork pull requests receive neither secrets nor an OIDC token, so the credentials step
is guarded by a condition; the AWS tests then skip for forks exactly as they do today.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* feat: surface the SIP Reason header on call status events
Carriers fronting ISDN/E1 PRI trunks put the authoritative disconnect cause in
an RFC 3326 Reason header rather than in the SIP status line, e.g.
SIP/2.0 408 Request Timeout
Reason: Q.850 ;cause=18
Q.850 cause 18 is "no user responding" - nobody answered, not a platform fault.
Different causes also arrive under the same SIP status (503 with cause=38
network out of order, or cause=41 temporary failure), so the status code alone
cannot classify the outcome of the call.
drachtio relays the header intact and it is present on the response object, but
the feature server only read the status line from it, so the cause was lost at
the application boundary and never reached the call status webhook.
Rather than extract the header at each emit site, carry the SIP message that
caused the status change on the callStatusChange event and derive from it in one
place, so provisional responses, the 200, final failures, BYE and CANCEL are all
covered by the same code and future headers cost one line.
Note this partly overlaps the existing _extractCustomHeaders/sip_headers
passthrough: that already exposes a Reason header arriving on a BYE, but it only
runs on the hangup path, so nothing covered the outbound INVITE failure
responses where a Q.850 cause matters most. sip_reason_header is a dedicated,
documented field that behaves the same on every status event.
sip_reason keeps meaning the status line phrase, and no key is added when there
is no Reason header, so existing consumers see an unchanged payload.
Adds a test:unit script so the smoke test runs without the docker testbed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs: note that the Reason header arrives re-serialized, not verbatim
Verified live end-to-end against a cluster, capturing both the external leg and
the leg into the feature server:
14.226.234.142 -> 10.0.197.31:5060 Reason: Q.850 ;cause=31 (as sent)
10.0.197.31:5060 -> :5070 Reason: Q.850;cause=31 (to fs)
Proxying re-serializes the header, normalizing the optional whitespace RFC 3326
permits around ';'. The same normalization appears in customer captures from an
unrelated deployment, so this is drachtio behaviour, not cluster-specific.
sip_reason_header is therefore the header as this process received it, not the
carrier's exact bytes - worth stating outright, since the spacing inconsistency
is exactly what consumers ask about, and a consumer who string-matches on the
spaced form would silently never match.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: clear a stale Reason header in redis, pass it on the alloc path, run the tests in CI
Three problems found reviewing the earlier commits.
1. The redis call record kept a stale header forever. updateCallStatus assigns
unconditionally, but toJSON only emits truthy values, and the same object is
written to the redis hash with hmset - a MERGE. An absent key therefore left
the previous status change's header in place, readable via GET /Calls/:sid:
a leg that got 183 + "Reason: Q.850;cause=31", then answered on a clean 200
and completed on a plain BYE, reported a Q.850 temporary-failure cause for a
call that ended normally. The webhooks were right; only the call record was
wrong. Storing absence as '' overwrites it. The existing "does not linger"
test passed throughout because it only exercised toJSON in memory, so this
adds one that pins what survives the redis filter (verified: it fails
without the fix).
2. The endpoint-allocation failure path already extracted the Reason header and
relayed it to the SBC, but called _notifyCallStatusChange without it - so a
FreeSWITCH 488 with "Reason: Q.850;cause=88 INCOMPATIBLE_DESTINATION" told
the SBC the cause and the application nothing, which is exactly the case
this feature exists to expose. That error is an fsmrf object rather than a
SipMessage, so it cannot go through msg (the guard in
reasonHeaderFromSipMessage would silently return undefined); the event now
takes an explicit sipReasonHeader for callers holding the value already.
3. The test:unit script added with these tests was never wired into CI - the
workflow runs jslint and npm test, and npm test enumerates its files
explicitly and skips test/unit entirely, so the suite would have rotted
unnoticed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: clear the stale header on the CallSession path too, and keep the CANCEL
Both from PR review.
1. The previous fix only worked for SingleDialer. Storing '' on the instance is
enough for callers that write the instance itself, but CallSession writes
toJSON(), and toJSON() drops falsy values on purpose so the key stays out of
the webhook payload - so '' never reached hmset and the stale header
survived. Reproduced:
after 183, toJSON has: Q.850;cause=31
raw instance value: ""
instance -> redis has key: true (SingleDialer, clears)
toJSON() -> redis has key: false (CallSession, stale persists)
The split is by session class, not call direction: place-outdial is required
only by dial.js, so SingleDialer covers dial-verb child legs while inbound,
REST-created and adulting all inherit CallSession._notifyCallStatusChange -
including the REST outdial path this feature was written for.
Adds CallInfo.toRedisJSON(), a named projection for the merge-semantics
store, so the divergence from toJSON() lives in one documented place rather
than being rediscovered at each call site. The unit test asserted against the
raw instance, which is why it passed while the real path was broken; it now
goes through both writers and fails if either regresses.
2. The caller-abandoned race dropped the CANCEL. middleware.js had it in hand
and discarded it, so the constructor's _onCancel() passed nothing whenever
the CANCEL beat the application fetch - the header landed or not depending on
timing. Worth closing because the 487 and its 'Request Terminated' phrase are
both ours, so every abandoned inbound call looks identical: a
Reason: SIP;cause=200;text="Call completed elsewhere" is what separates a
forked branch losing the race from a caller who gave up, and today both are
just no-answer.
Confirmed on a deployed srf (5.0.27) that this is not a no-op:
copyUASHeaderToUACForOnlyCancel forwards a hardcoded ['Reason', 'X-Reason']
when proxying a CANCEL, so the header does reach us.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* refactor: keep the Reason header out of the redis call record
Reversing the earlier approach after a closer look at who reads what.
The header was reaching the redis call record simply because CallInfo feeds two
sinks with opposite semantics: the status webhook is an EVENT (each POST is
independent, an absent key means "not in this event") while the redis record is
STATE written with hmset, a MERGE (an absent key means "keep what was there").
sipReasonHeader is the first field here that can legitimately go from set back
to unset - sipStatus and sipReason are only ever overwritten - so it was the
first to expose the mismatch, and two rounds of fixes had to chase it because
the two writers project from different bases.
Rather than keep managing that, exclude it: redis holds calls that are still
live, where there is usually no interesting cause yet, and by the time there is
one the call is over. Call history is served from RecentCalls (the CDR), which
is what the webapp reads - not GET /Calls. So the field earned very little there
while costing an invariant that has to be remembered forever.
Worth being explicit that this is NOT simply a revert: dropping the '' would
only have stopped the empty value being written. When the header is PRESENT it
still reached redis through both writers - toJSON includes it, and SingleDialer
writes the instance's own properties - and was then never cleared. Keeping it out
takes the same machinery as clearing it, just inverted, so this is a choice about
where the field belongs rather than a saving.
WEBHOOK_ONLY_FIELDS names that intent in one place, with the reasoning, so the
next field with the same property has somewhere obvious to go. The test asserts
both writers, since they project from different bases and checking one would
pass while the other still wrote the field (verified: it fails if the exclusion
is removed).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: apply the redis projection at the boundary, and drop dead adulting plumbing
From PR review.
1. There was a THIRD call-record writer. Asking each call site to remember the
projection was the wrong shape, and review found the proof: this class writes
the record from the status change, the recording flag AND (private only)
_persistConferenceState, and the last one still passed toJSON() straight
through. A conferenced caller whose BYE carried Reason: Q.850;cause=16 would
have that written into the call hash on the next conference-state persist,
where hmset merges and nothing can ever clear it.
Fixed by wrapping updateCallStatus once where it is bound, so every write is
projected and a newly added writer cannot bypass it by forgetting to ask. The
call sites go back to passing plain shapes. This is a class of miss that a
unit test cannot catch - the contract test passed the whole time - so the fix
is structural rather than another assertion.
2. The msg/byeReq parameters added to AdultingCallSession were dead code. The
only path that reacts to the far-end BYE there is the inline
sd.dlg.on('destroy') handler, which calls _callReleased() and discards the
request; _hangup is reachable only from CallSession.hangup() (LCC), which
passes nothing. The Reason header on that leg does reach the webhook, via
SingleDialer's own destroy handler - so the plumbing was not just unused but
misleading about which path carries it. Removed, with a comment at the handler
recording where the event actually comes from so it does not get re-added.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* docs: correct the projection comment in SingleDialer
Copied verbatim from CallSession, where the list of writers (status change,
recording flag, conference state) is accurate. SingleDialer has one writer, so
the comment described code that is not there.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Applications currently have no way to observe a mid-call hold/un-hold.
The feature-server handles the in-dialog re-INVITE internally (re-modifies
the endpoint, returns 200) but never notifies the app, so a transcription
or agent-assist app can't pause/resume when a caller is held or retrieved.
This detects the hold state transition on a re-INVITE and delivers it over
the existing sipRequestWithinDialogHook channel as a verb:hook, plus emits
a local 'hold'/'unhold' event so in-process tasks (e.g. transcribe) can react.
- call-session.js: ordinary calls — reuse isOnhold() (a=sendonly/a=inactive).
- siprec-call-session.js: SIPREC overrides _onReinvite, so a separate handler
is required. SIPREC recording streams are sendonly by nature, so a stream
flipping to a=inactive is treated as the hold signal; the full multipart
body (SDP + rs-metadata) is always forwarded. Fires on transition only;
other re-INVITEs are delivered as event:"reinvite".
Delivery uses requestor.request('verb:hook', ...) — the same channel
_handleRefer uses — rather than a task-bound performHook, so events are not
dropped if the current task has ended. Enable per call with:
{"verb":"config","sipRequestWithinDialogHook":"/hold-events"}
Confirmed live against Cisco CUBE (Cisco-SIPGateway/IOS-17.12.7b) over SIPREC:
both recording streams flip sendonly<->inactive on hold/resume, each
transition emitted correctly.
The whole feature is opt-in behind the JAMBONES_HOLD_UNHOLD_EVENTS env var;
when it is unset, _notifyHoldState / _notifySipRecReinvite are no-ops (neither
the local event nor the hook fires).
Reinstates the feature previously reverted in #1563. The revert was due to the
unit test failing in the full suite: lib/config.js captures the env var once at
load time, and earlier suite files require config/session modules (flag unset)
before this test could set it, so the enabled-path assertions saw a disabled
feature. The test now force-reloads config + the session modules with the flag
set (and restores the module cache afterward), so it is robust to suite load order.
Applications currently have no way to observe a mid-call hold/un-hold.
The feature-server handles the in-dialog re-INVITE internally (re-modifies
the endpoint, returns 200) but never notifies the app, so a transcription
or agent-assist app can't pause/resume when a caller is held or retrieved.
This detects the hold state transition on a re-INVITE and delivers it over
the existing sipRequestWithinDialogHook channel as a verb:hook, plus emits
a local 'hold'/'unhold' event so in-process tasks (e.g. transcribe) can react.
- call-session.js: ordinary calls — reuse isOnhold() (a=sendonly/a=inactive).
- siprec-call-session.js: SIPREC overrides _onReinvite, so a separate handler
is required. SIPREC recording streams are sendonly by nature, so a stream
flipping to a=inactive is treated as the hold signal; the full multipart
body (SDP + rs-metadata) is always forwarded. Fires on transition only;
other re-INVITEs are delivered as event:"reinvite".
Delivery uses requestor.request('verb:hook', ...) — the same channel
_handleRefer uses — rather than a task-bound performHook, so events are not
dropped if the current task has ended. Enable per call with:
{"verb":"config","sipRequestWithinDialogHook":"/hold-events"}
Confirmed live against Cisco CUBE (Cisco-SIPGateway/IOS-17.12.7b) over SIPREC:
both recording streams flip sendonly<->inactive on hold/resume, each
transition emitted correctly.
The whole feature is opt-in behind the JAMBONES_HOLD_UNHOLD_EVENTS env var;
when it is unset, _notifyHoldState / _notifySipRecReinvite are no-ops (neither
the local event nor the hook fires).
* fix say verb does not close streaming when finish say
* wip
* wip
* ttsStreamingBuffer reset eventHandlerCount after remove listeners
* only send tokens to module if connected
* wip
* sent stream_open when successfully connected to vendor
* Revert "Update dial.js (#1243)"
This reverts commit 259dedcded.
* add to .gitignore
* when we receive a REFER on the parent leg, after adulting the child the dial task in the parent session should end
* fixed dial transcribe is not able to received final transcribe when close call.
* wip
* fix review comment
* support call session delay detroy ep when current task is transcribe
* wip
* wip
* fixed review comments
* fixed review comments
* fix transcribe fixes for speechmatics
* update to verb-specs with fixes for speechmatics
* add support for speechmatics translation
* add handlers for receiving translations
* call translation hookd
* gather: no need to restart speechmatics after a final transcript during continuous asr
* graceful shutdown
* wip
* wip
* wip
* wip
* wip
* feature server should send USER call to the sbc sip that is connect with the user
* feature server should send USER call to the sbc sip that is connect with the user
* feature server should send USER call to the sbc sip that is connect with the user
* fix review comment
* add env variable to enable the feature
* add env variable to enable the feature
* add env variable to enable the feature
* minor test update
---------
Co-authored-by: Dave Horton <daveh@beachdognet.com>
* fix to be more precise about removing custom event handlers so that when we stop a gather we dont also inadvertently stop a background transcribe as well
* test fixes
* fix: endpointing=false was being ignored for Deepgram
* fix release freeswitch media properly
* if a leg is opus, modify b leg offer opus first
* if a leg is opus, modify b leg offer opus first
* wip
* wip
* fix review comments
* fix review comments
* fix review comments
Add support for sending 'amd' property in createCall REST API and also added support for using any of the speech vendors for STT
---------
Co-authored-by: Dave Horton <daveh@beachdognet.com>
