refactor(trace): record output messages through record()

[rgao-coreweave]

0.16.5 added outputMessages to Turn.record(); 0.16.6 added it to SubAgent.record() too. That closes the parity gap, so all five gen_ai.output.messages writers migrate together instead of leaving a split.

// every call site is now one line
recordOutput(call.span, [text]);                       // subagent
recordOutput(span, turns.flatMap(t => t.text), model); // teammate turn
recordOutput(turn.span, text, latestModel);            // turn

A single recordOutput() helper holds the shape conversion and the
already-ended guard, so no call site repeats them. Net effect across the repo is
+2 lines. assistantOutputMessages() has no callers left and is deleted;
assistantMessages() is now module-private. Afterwards ATTR.OUTPUT_MESSAGES
survives only as a constant used by tests, and no src/ code hand-serializes
the attribute.

Wire format is unchanged

Verified against a live in-memory exporter before writing the change:

manual    : [{"role":"assistant","content":"first reply"},{"role":"assistant","content":"second reply with \"quotes\" and \n newline"}]
via record: [{"role":"assistant","content":"first reply"},{"role":"assistant","content":"second reply with \"quotes\" and \n newline"}]
IDENTICAL

Message.content is an optional field the SDK stringifies verbatim, so {role, content} survives as-is, and SubAgent.end() serializes _outputMessages exactly as Turn.end() does. Two independent pins confirm it: the turn-side assertions added in #170, and the pre-existing subagent assertions (25 of them fail if the subagent writes are removed).

Two deliberate details

The Turn calls stay guarded. record() writes at end() and warns when the span has already ended. Making it unconditional would emit weave.Turn.record() called after end() noise into the daemon log on recovery paths that write nothing today. Passing undefined for either field is a no-op inside the SDK, so a previously recorded model is never clobbered.

gen_ai.response.model is left alone. record({ model }) writes gen_ai.request.model, a different attribute. The setAttributes({ [ATTR.RESPONSE_MODEL]: … }) calls in hookHandler.ts are therefore not folded in, since that would silently move the value to a different key.

Verification

npm run check

tsc clean, 203/203 pass.

🤖 Generated with Claude Code

[rgao-coreweave]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• refactor(trace): move tool spans onto 0.16.6 end() options #173
• refactor(trace): record output messages through record() #171 👈 (View in Graphite)
• test(trace): cover gen_ai.output.messages on turn and teammate spans #170
• chore(deps): bump weave to 0.16.6 #169
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[w-b-hivemind]

HiveMind Sessions

4 sessions · 40m · $13

Session	Agent	Duration	Tokens	Cost	Lines
Multi-Agent Code Review of TypeScript Files 4e2b4f5e-e561-4d9e-94c4-3bc98b53f8ba	claude	58s	2.1K	$0.30	+0 -0
Multi-Agent Code Review of Parser and Lifecycle ce6d3866-6899-471e-ba0b-457a190257ea	claude	1m	2.7K	$0.35	+0 -0
Multi-Agent Code Review of TypeScript Files ab11b764-d095-4007-8aeb-bbd935665e35	claude	1m	2.0K	$0.35	+0 -0
Build and Instrument Agent Team with Weave Tracing 4a3aa5b0-d015-4d3a-8048-426e02002931	claude	36m	83.8K	$12	+521 -52
Total		40m	90.6K	$13	+521 -52

View all sessions in HiveMind →

Run claude --resume 4e2b4f5e-e561-4d9e-94c4-3bc98b53f8ba to pickup where you left off.