Skip to content

feat(elevenlabs-agents): ElevenLabs Agents conversation forwarder (TH-8231) - #256

Merged
nik13 merged 6 commits into
devfrom
feat/th-8231-elevenlabs-agents
Oct 26, 2026
Merged

nik13 merged 6 commits into
devfrom
feat/th-8231-elevenlabs-agents

Conversation

@nik13

@nik13 nik13 commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor

Status 2026-10-21: ready for review. Live-verified: GET backfill (2026-10-17) and signed webhook delivery from a real ElevenLabs workspace to a Cloud Run forwarder into Future AGI cloud (2026-10-19). CI now runs on this branch and is green (Plan, fi_instrumentation, package (elevenlabs-agents)); 118 package tests pass locally on Python 3.11. Still not observable live (documented as limits, not blockers): ElevenLabs retries and duplicate deliveries, a span with status ERROR, and a long call's body size. The "BLOCKED-ON" and "no CI checks" notes further down are historical. Release: publish traceAI-elevenlabs-agents 0.1.0 to PyPI after merge; docs PR future-agi/docs#939 installs from PyPI and merges after that.

Summary

New package traceAI-elevenlabs-agents (python/frameworks/elevenlabs-agents), the customer-run forwarder from the approved TH-8231 architecture (D-V01). ElevenLabs runs the agent, so there is nothing to instrument. After a call, ElevenLabs POSTs an OTLP JSON trace (post_call_transcription_otel) to a webhook the customer hosts. This package is that webhook:

  • It verifies ElevenLabs-Signature: t={unix},v0={hmac} (HMAC-SHA256 over {timestamp}.{raw_body}, 30 min window) with the standard library. A delivery whose signature fails is never parsed. The body is read (up to 20 MiB; a larger declared length gets 413 unread) because the HMAC covers the raw bytes.
  • It maps the conversation onto the Observe Voice-tab names. The parentless elevenlabs.conversation root gets fi.span.kind=CONVERSATION, gen_ai.system=elevenlabs, session.id, call.duration and call.total_turns, plus the transcript (conversation.transcript, fi.conversation.transcript, flattened rows). Agent turns with elevenlabs.producing_llm become LLM spans with the model and gen_ai.usage.* tokens from elevenlabs.llm_usage.*. elevenlabs.tool.{name} spans become TOOL spans. Turns and tools get input.value / output.value. Trace, span and parent ids, names, timestamps and status codes are kept as sent, and unknown span names pass through.
  • It exports uncompressed OTLP/HTTP JSON to {FI_BASE_URL}/tracer/v1/traces with X-Api-Key/X-Secret-Key, and answers 2xx only after the collector accepted the batch.
  • It keeps a per-conversation idempotency record (SQLite, in the customer's process), because ElevenLabs keeps the trace id but re-mints span ids on every export.
  • traceai-elevenlabs-agents backfill --conversation-id ... does one GET /v1/convai/conversations/{id}?format=opentelemetry through the same mapper.
  • Privacy follows the traceAI TraceConfig convention, with no package-specific switch. Content is captured by default. FI_HIDE_INPUTS / FI_HIDE_OUTPUTS (or TraceConfig(...) passed as ForwarderConfig.trace_config) remove it. In a conversation the turns are the messages, so the *_MESSAGES / *_TEXT variants remove a side the same way, and either side removes the combined transcript. FI_PII_REDACTION works as it does everywhere else in traceAI. This matches sibling TH-8242 (openai-realtime).
  • Never exported in any mode (privacy.py):
    • resource attributes outside an allowlist (no host.*, process.* or os.*);
    • span attributes under environment and identity prefixes (host., process., user., enduser., code. …), stack-trace keys and credential-named keys;
    • stack traces (error text is cut to one line);
    • emails, local paths and private host names in metadata, status messages and exception messages. These are replaced with <EMAIL_ADDRESS> / <FILE_PATH> / <HOSTNAME>, and an email-shaped user.id is not written;
    • signed media URLs, and any payload project_name / fi.* resource attribute.

Runtime dependency: fi-instrumentation-otel (for TraceConfig only). Everything else is standard library. Python >=3.10,<3.14. The elevenlabs SDK is a dev/test oracle only. Ships a poetry.lock (Poetry 2.0.1, like livekit).

Docs page: future-agi/docs PR #939 (src/pages/docs/integrations/traceai/elevenlabs-agents.mdx).

What changed in this revision (review fixes, commit 8025926)

  • Privacy default reversed to the binding 2026-10-16 clarification. Content is now on by default via fi_instrumentation.TraceConfig and the standard FI_HIDE_* env vars. FI_ELEVENLABS_CAPTURE_CONTENT and ForwarderConfig.capture_content are removed (never released).
  • New never-exported layer, applied in both modes. The old opt-in mode passed the payload's resource attributes, status messages and event attributes through unchanged. A reviewer found host.name=alice-laptop…, process.executable.path=/Users/alice/…, a status message with /home/alice/app/agent.py on alice-laptop.local and a dynamic-variable email stored in an opt-in project. None of these is exported now in either mode. The new E2E below proves it on stored rows.
  • The webhook path also passes data.user_id when a delivery carries one.
  • Wording: "rejected before its body is read" is now "never parsed" (the body is read for the HMAC), here and in the docs.
  • The README leads with the git install until the PyPI release, and the test count is corrected.
  • Added poetry.lock.

How to run

cd python/frameworks/elevenlabs-agents
pip install -e . pytest elevenlabs==2.73.0
pytest tests                                    # 112 tests
# end to end against a Future AGI collector, ElevenLabs stood in:
FI_API_KEY=... FI_SECRET_KEY=... FI_BASE_URL=http://localhost:14338 examples/run_e2e.sh my-project ./out

run_e2e.sh runs two phases. The first uses the defaults and exports into <project>. The second sets FI_HIDE_INPUTS, FI_HIDE_OUTPUTS, FI_HIDE_INPUT_MESSAGES and FI_HIDE_OUTPUT_MESSAGES to true and exports into <project>-hidden.

Versions

ElevenLabs surface post_call_transcription_otel webhook + GET ...?format=opentelemetry, OpenTelemetry traces re-read 2026-10-15: contract unchanged vs the 2026-10-03 pin
elevenlabs SDK (oracle) 2.73.0 (latest stable, released 2026-10-15) primary; 2.70.0 (PRD pin) matrix leg. webhooks_custom.construct_event is byte-identical in both
fi-instrumentation-otel 1.1.0 (latest, 2026-08-27) primary; 0.1.11 (lowest allowed by >=0.1.11) matrix leg: 111 pass, and the FI_PII_REDACTION test is skipped because that release has no such switch
Python unit tests 3.10.17, 3.12.10, 3.13.15 (112 pass each); E2E on 3.12.10
Future AGI local th8103-e2e stack (future-agi ea12cc3f44), collector OTLP/HTTP :14338

E2E evidence (real forwarder process, real local Future AGI; ElevenLabs = STANDIN)

Installed exactly as the docs page says, into a fresh python -m venv: pip install "git+https://github.com/future-agi/traceAI.git@feat/th-8231-elevenlabs-agents#subdirectory=python/frameworks/elevenlabs-agents". This resolved to 8025926 and pulled fi-instrumentation-otel 1.1.0 from PyPI. Then traceai-elevenlabs-agents serve ran, configured only with the documented env vars. examples/standin.py built docs-shaped deliveries, signed them as documented, and served the GET endpoint. The real SDK 2.73.0 construct_event accepted each signed delivery first, and the real SDK conversations.get(format="opentelemetry") parsed the GET stand-in body.

All 28 steps (14 per phase) returned the expected status (e2e-report.json ok=true, both forwarder processes exited 0 on SIGTERM):

  • the conversation was forwarded, and the retry got 200 duplicate;
  • the abnormal disconnect was forwarded;
  • wrong, stale (31 min), missing and mutated signatures got 401;
  • post_call_transcription, post_call_audio and call_initiation_failure got 200 ignored;
  • a body with no resourceSpans got 400;
  • backfill of the delivered conversation got 200 duplicate, and backfill of a GET-only conversation got 200 forwarded;
  • with the collector down the forwarder answered 502 and nothing was stored.

Fresh projects, so the counts are exact:

Project Trace What
th8103-rest-elevenlabs-agents-default (traceAI defaults) 7267a3b7b89394dff15a165aeeadc9bb conversation, tool call, 2 LLM turns
3801318caf7dced47385fb750c9228e0 abnormal disconnect: root ERROR, tool ERROR
cce6435864c6464a79b6b545f1856329 GET backfill only (carries user_id)
th8103-rest-elevenlabs-agents-hidden (four FI_HIDE_* = true) ed3982282a67dff708e74b3b714efeb8, 26adb61e097946a016d7b4a9b76ac24c, 637b0bac13ce47e9709c8aac1a05cae0 same three

fi_verify PASS on both (3 traces, 20 spans each):

project 'th8103-rest-elevenlabs-agents-default': type=['observe'] traces=3 spans=20
trace 7267a3b7-b893-94df-f15a-165aeeadc9bb (8 spans)
- [conversation] elevenlabs.conversation in=70B out=37B (42000ms)
  - [chain] elevenlabs.recv.agent_response in=0B out=64B (3000ms)
  - [chain] elevenlabs.recv.user_transcript in=70B out=0B (4000ms)
  - [llm] elevenlabs.recv.agent_response model=gpt-4o-mini tok=412/23 in=70B out=67B (11000ms)
    - [tool] elevenlabs.tool.get_exchange_rate in=73B out=64B (1000ms)
  - [chain] elevenlabs.recv.user_transcript in=55B out=0B (3000ms)
  - [llm] elevenlabs.recv.agent_response model=gpt-4o-mini tok=450/4 in=55B out=37B (2000ms)
  - [unknown] elevenlabs.event.something_new in=0B out=0B (0ms)
trace 3801318c-af7d-ced4-7385-fb750c9228e0 (4 spans)
- [conversation] elevenlabs.conversation in=62B out=49B status=ERROR:call ended: websocket closed abnormally zqx-th8231-c9aaec1e6… (18000ms)
  - [chain] elevenlabs.recv.user_transcript in=62B out=0B (3000ms)
  - [llm] elevenlabs.recv.agent_response model=gpt-4o-mini tok=120/9 in=62B out=49B (5000ms)
    - [tool] elevenlabs.tool.book_table in=68B out=0B status=ERROR:upstream timeout zqx-th8231-c9aaec1e627d-9921 (10000ms)
PASS

project 'th8103-rest-elevenlabs-agents-hidden': type=['observe'] traces=3 spans=20
- [conversation] elevenlabs.conversation in=0B out=0B (42000ms) … [llm] model=gpt-4o-mini tok=412/23 in=0B out=0B …
- [conversation] elevenlabs.conversation in=0B out=0B status=ERROR: (18000ms) … [tool] elevenlabs.tool.book_table status=ERROR:
PASS

The stored root ERROR message in the default project is call ended: websocket closed abnormally <marker> <FILE_PATH> on <HOSTNAME>. The stand-in sent /home/alice/app/agent.py on alice-laptop.local.

Voice tab. The real GET /tracer/trace/list_voice_calls/ lists each conversation exactly once in both projects: 3 calls each, durations 42 s / 42 s / 18 s, turn_count 5 / 5 / 2. The retry and the backfill of a delivered conversation added no call and no span. Screenshots:

  • the Voice call detail (default) renders the 5-turn transcript;
  • the same view (hidden) shows the call with Turns 5 and Duration 42 s, with no transcript;
  • the trace view (default, error trace) shows the scrubbed one-line error.

Published checker. docs/public/fi_verify_voice.py check was run on the rows Future AGI stored. It does not go fully green for a webhook forwarder:

  • default project, every trace:
    • V2, V3, V4, V5, V6, V8, V9, V10, V11 and V12 PASS. V8 (transcript) and V11 (LLM prompt and completion) now pass without any opt-in.
    • V1 FAIL: it needs an SDK exporter receipt, which a webhook forwarder does not have.
    • V7 FAIL: call.talk_ratio is omitted because the payload has no per-side audio durations.
  • hidden project: V8 and V11 also FAIL, by design, because FI_HIDE_* removes the transcript and the turn IO.

Privacy proof (as the 2026-10-16 clarification defines it)

Each phase wrote its own unique marker into every content field: user/agent text, tool params/results, reasoning, RAG query, analysis summary, a dynamic-variable email, the GET user_id, an exception stack-trace event and the ERROR status message. The markers were zqx-th8231-c9aaec1e627d-9921 (default) and zqx-th8231-e9712c5fafdd-9921 (hidden). The stand-in also sent adversarial extras: host.name=alice-laptop-<marker>.local, process.executable.path=/Users/alice/<marker>/agent, /home/alice/app/agent.py on alice-laptop.local in a status message, a signed X-Goog-Signature media URL, and a payload project_name.

After the run, every row and every column of every ClickHouse table holding rows for the project was dumped and scanned. Every Postgres tracer_* row for it was scanned too. For spans that covers resource_attrs, span_events, attributes_extra, attrs_string/number/bool, input, output, status_message, metadata, model and the rest. The other ClickHouse tables were traces, trace_sessions, spans_per_session, end_users and the rollups. Totals: 53.7 KB (default) and 43.7 KB (hidden).

Needle default project hidden project
this phase's marker 127 (transcript, IO, status messages, analysis, reasoning) 0
the other phase's marker 0 0
content phrases ("What is the US dollar", "Book me a table", "One US dollar", "usd eur rate", "upstream timeout", "websocket closed", the account number) present 0 each
alice, /Users/, /home/, alice-laptop, example.com, host.name, process.executable, exception.stacktrace, Traceback, File ", X-Goog-Signature, Expires=, someone-elses-project, stand-in webhook secret / xi key prefixes, FI API key value, FI secret key value 0 each 0 each
<EMAIL_ADDRESS> / <FILE_PATH> / <HOSTNAME> 2 / 1 / 1 (the scrubbed dynamic-variable email and status message) 0

Forwarder logs (four processes): 0 markers, 0 key values, 0 v0= signatures, 0 alice, 0 Traceback. Lines carry the conversation id and the outcome only.

FI_HIDE_INPUTS=true FI_HIDE_OUTPUTS=true alone hides the same things as the four switches used here (unit-tested: test_trace_config_sides). The docs page shows the shorter form.

Limitations / unsupported

  • No live ElevenLabs delivery yet (STANDIN). Fixtures follow the documented payload and the attribute names the Langfuse cookbook reads from real exports (elevenlabs.tool.params/result, elevenlabs.producing_llm, elevenlabs.llm_usage.<model>.<kind>, elevenlabs.turn.index). They are not captured from a workspace.
  • call.talk_ratio and per-turn audio durations are not written (they are not in the payload). The Voice detail Duration, latency and talk-% tiles stay empty, while the duration and turns badges are filled.
  • Not handled: monitoring WebSocket (Enterprise, D-V12), post_call_audio/recordings (D-V03; signed URLs dropped), TTS-only API, RAG chunk text.
  • Platform behaviour seen:
    • the collector stores span events as name+time only, so exception messages on events are dropped even with defaults;
    • the Voice list shows every forwarded call as completed, while the trace status shows ERROR.
  • The scrubbing of emails, paths and host names applies to metadata channels, not to what the caller and agent said. An email a caller speaks stays in the transcript unless FI_PII_REDACTION=true or that side is hidden (see Decisions).
  • Unauthenticated callers can make the process buffer up to 20 MiB per request before the HMAC check. The docs tell operators to rate- and size-limit at the ingress.
  • The Voice-tab assertion in tests/voice_tab_check.py is a flagged local stand-in. The TH-8339 shared assert_voice_tab helper does not exist on dev or in the open harness PR test(harness): shared OTLP harness for TH-8103 contract tests #203 (coordinator decision for TH-8231).
  • This branch has no CI checks (statusCheckRollup is empty), so the 112 tests ran only locally (3.10 / 3.12 / 3.13 and the two dependency legs above).
  • Root README catalog row not added (BRIEF: catalog rows are handled separately).
  • Edge: if the process dies after the collector accepted a batch but before it recorded done, a retry within 5 minutes gets 503, and after 5 minutes the conversation is re-exported (duplicate spans).
  • The earlier projects th8103-rest-elevenlabs-agents / -optin still hold runs from the previous privacy model. The evidence above comes from the fresh -default / -hidden projects only.

Decisions made (vs PRD/architecture)

  • Privacy follows the 2026-10-16 coordinator clarification. That also matches the PRD's "no second privacy mode" and sibling TH-8242. The earlier FI_ELEVENLABS_CAPTURE_CONTENT opt-in is removed.

  • Which side each field is on. User turns, tool params, RAG query, dynamic variables and user.id are inputs. Agent turns, tool results, reasoning, analysis, data collection and error text are outputs. Unplaceable attributes and the combined transcript need both sides. *_MESSAGES / *_TEXT hide a side like FI_HIDE_INPUTS / FI_HIDE_OUTPUTS, the same reading TH-8242 used.

  • "Never exported" scope. The clarification's list ("local paths, usernames, hostnames, emails, raw stack traces") is applied to environment and identity channels, in both modes:

    • the resource allowlist;
    • identity key prefixes;
    • stack traces everywhere;
    • email/path/host scrubbing of metadata and error text.

    Conversation text is not scrubbed, so that it stays useful as the transcript. Its email redaction is the existing traceAI FI_PII_REDACTION. This is the same line every traceAI instrumentor draws for input.value.

  • TraceConfig.mask is applied to every exported content attribute, as FITracer does (this is what gives FI_PII_REDACTION). Structural ids (session.id, kinds, model, tool name, conversation/agent id) are exempt so grouping never changes.

  • Idempotency is a per-conversation record (the Rick r2 review note), not trace-id/argMax collapse. ElevenLabs re-mints span ids per export (Langfuse cookbook), so GET backfill converges by being acknowledged as duplicate, and --force re-exports.

  • Both transcript spellings are the same JSON list of {role, content, time}. The Voice drawer read path calls json.loads on fi.conversation.transcript and needs a list.

  • Child kinds (LLM/TOOL/CHAIN) are added for hierarchy, and only the root is CONVERSATION. LLM price is not mapped into the cost column; elevenlabs.cost_fiat stays an attribute.

  • Timestamps more than 30 min in the future are rejected too (the SDK only checks the past).

  • The 20 MiB body cap is kept rather than lowered. A 413 is never retried by ElevenLabs, and no real payload size is known yet. Re-check this after the live call.

BLOCKED-ON

(Webhook live test done 2026-10-19; see below.) Remaining: ElevenLabs retries/duplicate delivery, ERROR status, long-call body size. This is the D-V07 publish bar; no account, spend or tunnel was used. Prerequisites for that call:

  • an ElevenLabs workspace whose post-call webhook has the Transcript event, OpenTelemetry transcript payloads and retries on;
  • a public HTTPS URL for the forwarder (no tunnel from this machine, per the BRIEF);
  • an ElevenLabs API key with CONVAI_READ, for the backfill check.

After the call, verify:

  • the real attribute names (elevenlabs.tool.params/result, producing_llm, the llm_usage format, turn.index);
  • what the ERROR status message contains;
  • time_in_call_secs;
  • whether a user_id appears;
  • real retry and duplicate behaviour;
  • the body size of a long call.

Then replace the docs-shaped fixtures with one captured body.

Nav (docs PR): Integrations → traceAI → Voice & Realtime, title ElevenLabs Agents, after Pipecat.

Live vendor verification (2026-10-17)

Real ElevenLabs API (key from the maintainer's workspace), local Future AGI. A throwaway agent (gpt-4o-mini, webhook tools get_zen OK / get_missing 404, an unanswered client tool) ran three short text-only conversations through the elevenlabs SDK 2.73.0; the agent and tools were deleted afterwards. traceai-elevenlabs-agents backfill (installed with the page's git command; 0.1.0, fi-instrumentation-otel 1.1.0, Python 3.12) exported each via GET ?format=opentelemetry. Spend: about $0.02 per conversation reported by ElevenLabs.

  • Traces: ea5d20b167de14804539f834f26d0630, 38962487f46b48524a4dd423e98839bb, 0269cc4314b821f1047fc94ab7ec7f2d (project th8103-rest-elevenlabs-agents-live); the first two also in ...-live-hidden with the four FI_HIDE_* on.
  • fi_verify PASS both projects (7 spans per trace: conversation -> chain/llm/tool, gpt-4o-mini 717/11, 745/9 ... tokens).
  • Privacy scan of every stored column: content present by default, absent with FI_HIDE_*; email, user name, home path, host name absent in both.
  • Findings: real attribute names matched the fixtures; trace id stable and span ids fresh across exports (confirmed); user_id null for SDK conversations; ElevenLabs reports status OK even for failing tools (404, unanswered client tool), so no live ERROR span was produced.
  • Docs docs_build.sh: BUILD_EXIT=0, AUDIT_EXIT=0. traceAI tests: 114 passed.

Still STANDIN / alpha: webhook delivery (HMAC header, post_call_transcription_otel, retries, long-call body size) is untested live: no public URL (no tunnel). ERROR mapping covered by fixtures only. time_in_call_secs not observed.

Live vendor verification (2026-10-19): webhook delivery

A real ElevenLabs workspace delivered signed post_call_transcription_otel webhooks to this forwarder (installed from the branch with the docs page's git command, deployed to Google Cloud Run, keys in Secret Manager) and the forwarder exported to Future AGI cloud. The workspace webhook (HMAC) was assigned as the post-call webhook of one throwaway text-only agent (Transcript event, OpenTelemetry format). Three SDK 2.73.0 conversations, about $0.06 in total; the agent, webhook, Cloud Run service, secrets and registry were deleted afterwards.

Case Body Forwarder Future AGI cloud trace
tool call, clean end, user_id set 12.3 KB 200 forwarded, 7 spans a242bad17b598c708b926176be38b7ea
tool call, socket dropped (Client disconnected: 1006) 12.2 KB 200 forwarded, 7 spans b97670d7e823ea2f0219ad85e9da56a5
silent, ended by the 2-minute limit 5.5 KB 200 forwarded, 2 spans e48b4246a1ca076c184eb4a825bf2bd8

Read back with GET /tracer/trace/<uuid>/ on api.futureagi.com: conversation > llm / chain / tool, model gpt-4o-mini, all OK.

Proven live: HMAC header and signature check; body shape {type, event_timestamp, data: {conversation_id, agent_id, branch_id, version_id, environment, otlp_traces}}; real attribute names (same as the 2026-10-17 GET); public-HTTPS delivery to a Cloud Run service; bodies of 5 to 12 KB.

Findings and fix (this push): the live body has no data.user_id; the id is the root attribute elevenlabs.user_id. Webhook deliveries therefore never got user.id. map.py now falls back to that attribute (input side, removed by FI_HIDE_INPUTS, still scrubbed for emails). ElevenLabs reports status OK for every span in all three cases, including the dropped socket (elevenlabs.termination_reason carries the reason) and the time-limit call. time_in_call_secs is not in the body (root span length is used). Cloud Run reserves /healthz (its frontend returns 404), so the docs now say to use another path for a health check there.

Fixtures: tests/fixtures/live_webhook_post_call_transcription_otel.json and ..._timeout.json are sanitized real bodies (ids replaced). Tests: 118 pass (4 new: hierarchy/model/tokens/user from the live body, no user id on the timeout body, OK status and 121 s, hide switches).

Local replay + privacy: the real bodies were replayed through the forwarder into the local stack. fi_verify PASS for th8103-el-live (defaults) and th8103-el-live-hidden (FI_HIDE_INPUTS, FI_HIDE_OUTPUTS): 3 traces, 16 spans. Scan of every stored ClickHouse and Postgres row: with defaults the message and tool text appear; with the switches on, 0 hits for the user text, the tool result, the user id, the local username/paths and the account e-mail (74.8 KB scanned).

Still stand-in: ElevenLabs retries and duplicate deliveries (never triggered; the forwarder always answered 200), a span with status ERROR (ElevenLabs sent OK every time), a long call's body size, time_in_call_secs. "Enable retries" could not be checked through the API.

🤖 Generated with Claude Code

nik13 and others added 5 commits October 15, 2026 19:44
…-8231)

Customer-run webhook forwarder for ElevenLabs Agents post-call OpenTelemetry
traces (post_call_transcription_otel) and a one-shot GET ?format=opentelemetry
backfill. Verifies ElevenLabs-Signature (HMAC-SHA256, stdlib only), maps the
conversation onto the Observe Voice-tab attributes (fi.span.kind=CONVERSATION
root, call.duration, call.total_turns, LLM model/tokens, tool spans), exports
uncompressed OTLP/HTTP JSON to {FI_BASE_URL}/tracer/v1/traces, and keeps a
per-conversation idempotency record because ElevenLabs re-mints span ids on
every export. Content capture is off by default (FI_ELEVENLABS_CAPTURE_CONTENT).

Tests: 75 (hand-computed HMAC + elevenlabs SDK construct_event 2.73.0/2.70.0
oracle, mapping, privacy, server, backfill, CLI, the shipped E2E example).
The Voice-tab assertion is a flagged local stand-in until the TH-8339 shared
helper exists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d allowlist (TH-8231)

Content capture now follows the TraceConfig convention every traceAI
instrumentor uses, with no package-specific second switch: content is on by
default and FI_HIDE_INPUTS / FI_HIDE_OUTPUTS (or TraceConfig(...) passed as
ForwarderConfig.trace_config) turn it off. In a conversation the turns are the
messages, so FI_HIDE_INPUT_MESSAGES / FI_HIDE_INPUT_TEXT drop the user side the
same way FI_HIDE_INPUTS does and the OUTPUT variants the agent side; either
side drops the combined transcript. FI_PII_REDACTION works as elsewhere.
FI_ELEVENLABS_CAPTURE_CONTENT is removed (never released). New runtime
dependency: fi-instrumentation-otel, for TraceConfig only.

Never exported in any mode (new privacy.py): resource attributes outside an
allowlist (no host.*, process.*, os.*), span attributes under environment and
identity prefixes and credential-named keys, stack traces (error text is cut to
one line), and emails, local paths and private host names in metadata, status
and exception messages; an email-shaped user.id is not written. Before this,
opt-in mode passed the payload's resource attributes, status messages and event
attributes through unchanged.

Also: the webhook path passes data.user_id when present, docs/strings say a
forged delivery is "not parsed" (the body is read, up to 20 MiB, so the HMAC can
be checked), the README leads with the git install until the PyPI release, the
test count is right, and the package ships a poetry.lock (Poetry 2.0.1, like
livekit).

Tests: 112 (new test_privacy.py; mapping, forwarder, CLI, backfill and the
shipped E2E example rewritten for defaults-on / FI_HIDE_* off), on Python 3.10,
3.12, 3.13, elevenlabs SDK 2.73.0 and 2.70.0, fi-instrumentation-otel 1.1.0 and
0.1.11 (the FI_PII_REDACTION test skips there).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ure (TH-8231)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ive webhook bodies; live webhook fixtures (TH-8231)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The README is the package's PyPI description, so it now shows the command that works
once 0.1.0 is released, instead of an install from this feature branch, which stops
working when the branch is deleted after merge.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nik13
nik13 marked this pull request as ready for review October 21, 2026 22:44
0.1.0 has never been published, so the live user.id fix and the live checks ship in it.
Drops a sentence that still called webhook delivery stand-in only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nik13
nik13 merged commit eb0bdd5 into dev Oct 26, 2026
4 checks passed
@nik13
nik13 deleted the feat/th-8231-elevenlabs-agents branch October 26, 2026 09:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant