Conversation
… session_id OITracer applies a session id set through OpenInference context (setSession) when the AGENT span starts; message processing then overwrote it with the SDK's own session_id on the system/init message and again on the result, for query(), unstable_v2_prompt(), and V2 session turns. js/AGENTS.md lists context propagation of the session id as a must for every instrumentor. Phoenix binds a trace to the first session.id it ingests, so a turn with no tool calls (AGENT span ingested first, carrying the SDK id) landed in a different session from a turn with tool calls (TOOL spans first, carrying the context id). Check the captured context once per span; when it supplies a session id, skip the SDK write. Without a context session the SDK id still fills session.id as before. Fixes Arize-ai#3775 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MXxV7qtX4GyTfo7QWaREnL
Author
|
I have read the CLA Document and I hereby sign the CLA |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #3775
What changed
A
session.idsupplied by the caller through OpenInference context (setSession(context.active(), { sessionId })) is applied to the AGENT span at start byOITracer, and was then overwritten by the SDK's ownsession_idon thesystem/initmessage and again on theresultmessage. This happened in three places:processMessage()forquery(),processSessionMessage()for V2 session turns, andwrapPrompt()viaextractResultSuccessAttributes()/extractResultErrorAttributes().Now each span checks its captured context once (
hasContextSessionId()), and when the caller supplied a session id the SDK write is skipped (withoutSessionId()on the result attributes; the init handler omits it). Without a context session the SDK id still fillssession.idexactly as before, so the existing assertion inv1Query.test.ts(session.id === "sess-123") is unchanged.js/AGENTS.mdlists "propagate context attributes via OITracer (session ID, user ID, metadata, tags)" as a must for every instrumentor. This package already propagates the session id; this PR stops it from being clobbered afterwards, which brings it in line with how metadata, user id and tags are already handled (applied at start, never overwritten by message processing).Why it matters in Phoenix
Phoenix binds a trace to the first
session.idit ingests for that trace and ignores later ones. Spans export when they end, so in a turn with tool calls the TOOL spans (which keep the context session id) land first and the trace goes into the caller's session; in a turn with no tool calls the AGENT span lands first with the SDK id and the trace goes into a session named after the SDK UUID. The same agent's turns end up split across two sessions depending on whether it happened to use a tool. Reproduced against the real@anthropic-ai/claude-agent-sdk@0.3.218with one turn (output in the issue).Tests
Three new tests, one per code path, each wrapping the call in
context.with(setSession(...))and asserting the AGENT span keeps the context session id (and still records the model name from init):query()withsystem/init+resultmessagesunstable_v2_prompt()unstable_v2_createSession()turn viasend()+stream()They use
provider.register()to install the async-hooks context manager for the duration of the test and reset the globals infinally, so no new devDependency. All three fail onmainand pass here; the rest of the suite is unchanged (45/45).Trade-off to flag
When the caller sets a session id, the SDK's
session_idno longer appears on the span at all. It was only ever there by accident in that case, but it is a useful value (it is whatclaude --resumetakes). If you would like it kept under a dedicated attribute when the context wins, say the word and I will add it to this PR; I left it out to keep the change to the reported behaviour.Changeset included (patch).
🤖 Generated with Claude Code
https://claude.ai/code/session_01MXxV7qtX4GyTfo7QWaREnL