fix(js): inherit parent's ls_agent_type in Vercel _getLsAgentType helper by ybathula707 · Pull Request #3319 · langchain-ai/langsmith-sdk

Yoshi (ybathula707) · 2026-07-31T20:30:05Z

Stacked on #3318. Merges after that PR lands.

What

Vercel's _getLsAgentType helper onlly checked parent.run_type === "tool" — never consulted parent.extra.metadata.ls_agent_type. Inner LLM spans created by wrapAISDK inside a middleware-tagged traceable() self-stamped root.
Both the outer middleware chain run and the real main model call shared conversation key "root", so guardrail LLM content leaked into the main conversation in the Messages view.

[0] entry_guardrail     ls_agent_type=middleware   ✓ (outer traceable)
[1] openai.responses    ls_agent_type=root         ✗ should be middleware
[8] exit_guardrail      ls_agent_type=middleware   ✓
[9] openai.responses    ls_agent_type=root         ✗ should be middleware

How

Consolidates Vercel's ls_agent_type logic into experimental/vercel/_agent_type.ts atop the shared util ls_agent_type.ts util

5 caller sites updated (wrap.ts × 4, telemetry.ts × 1).

Tagging precedence per span:

Any known non-root parent tag (middleware / subagent / compaction) → inherits explicitly at config-build time.
parent.run_type === "tool" → subagent
Parent tagged root → inherits root
No parent → default root via the util
Otherwise → undefined;

Fixes LSO-3681. Related LSO-3607.

Test Plan

Unit tests: js/src/tests/vercel_agent_type.test.ts 20/20 pass.
Telemetry vitest: src/tests/vercel/telemetry/methods.vitesttest.ts — 36/36 pass
Post-merge: fixture regen — 4 Vercel snapshots

github-actions · 2026-07-31T20:31:03Z

JS perf benchmark

Lower is better. Noisy on shared runners — treat as a signal, not a gate.

Base64-heavy payload

Single large base64 string per message — the shape the worker-offload path is optimized for.
Payload: 2511.2 KB in / 5.2 KB out, 100 runs.

metric	main	this PR	delta
Wall time (ms)	1196.98	1235.09	+3.2%
createRun total (ms)	107.81	95.84	-11.1%
createRun p50 (ms)	0.84	0.83	-0.7%
createRun p95 (ms)	3.20	1.04	-67.4%
createRun p99 (ms)	8.45	8.37	-0.9%
createRun max (ms)	8.45	8.37	-0.9%
updateRun total (ms)	33.84	30.55	-9.7%
updateRun p95 (ms)	0.59	0.60	+1.4%
loop lag total (ms)	753.42	758.57	+0.7%
loop lag p50 (ms)	0.12	0.12	-5.0%
loop lag p95 (ms)	5.10	5.99	+17.5%
loop lag p99 (ms)	59.73	63.80	+6.8%
loop lag max (ms)	81.35	87.15	+7.1%

Structural payload

Many small strings across a wide/nested object graph. Should bypass worker offload and use sync flush.
Payload: 1239.5 KB in / 13.3 KB out, 100 runs.

metric	main	this PR	delta
Wall time (ms)	2354.07	1957.95	-16.8%
createRun total (ms)	796.56	578.47	-27.4%
createRun p50 (ms)	6.49	5.50	-15.3%
createRun p95 (ms)	18.09	12.40	-31.5%
createRun p99 (ms)	23.95	24.50	+2.3%
createRun max (ms)	23.95	24.50	+2.3%
updateRun total (ms)	60.30	30.65	-49.2%
updateRun p95 (ms)	3.43	0.54	-84.3%
loop lag total (ms)	1908.44	1547.22	-18.9%
loop lag p50 (ms)	0.08	0.11	+32.4%
loop lag p95 (ms)	14.82	11.50	-22.4%
loop lag p99 (ms)	148.09	120.58	-18.6%
loop lag max (ms)	229.00	194.31	-15.1%

open-swe

Open SWE Review found 1 potential issue.

Open in Web • View Open SWE trace

… integration on_trace_start and the subagent stamp path both unconditionally overwrote user-supplied ls_agent_type. RunConfig(trace_metadata={ls_agent_type: "middleware"}) was silently clobbered back to "root" on trace open, causing middleware guardrail runs to leak into the main conversation in the Messages view. Refactors metadata assembly at both write sites: - on_trace_start / onTraceStart: reads trace.metadata (was ignored), then setdefault on ls_agent_type; ls_integration and ls_integration_version stay force-set as SDK identity. - on_trace_end / onTraceEnd: also force-set ls_integration to close the mirror spoof vector (user's trace.metadata could still override on this later lifecycle write). - Subagent stamp (AgentSpanData under FunctionSpanData): preserves user narrowing intent (middleware / compaction) but still applies structural detection when the current tag is missing, "root" (inherited default), or already "subagent". Fixes LSO-3608. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Under the reframed model (Arian): default root is stamped ONLY on the trace-root run when the user didn't supply. Nested traces (OAI Agents SDK invoked inside a user's outer traceable) don't get a default stamp — the enclosing traceable's tag propagates via create_child inheritance. User opt-out via ls_agent_type: null / None deletes the key from the run's metadata at both trace-start and trace-end (mirror-lifecycle write site). Subagent structural detection unchanged: still preserves user narrowing tags (middleware / compaction), still overrides root/absent -> subagent so the messages/v2 view can group agent-as-tool spans correctly. This is the documented trade: user-set root at trace level gets overridden at subagent-detected spans because we can't distinguish user-set root from default-stamped root at this write site. Also drops the on_trace_end ls_integration force-set from this PR — scope creep from the original LSO-3608 fix. Can be filed separately if SDK-identity spoofing at trace-end lifecycle is a real concern. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…edence Two bot-flagged issues in openai-agents-sdk integration: 1. Nested null opt-out re-inherited parent tag. Deleting the key from merged_metadata / mergedMetadata removed it from the update object only — create_child still copied the parent's ls_agent_type into the child's extra.metadata. Fix: keep None/null in merged metadata so the child's extra.metadata explicitly carries the null value, overriding create_child's inheritance. 2. Trace-end merge reversed user precedence. onTraceStart put trace.metadata after this._metadata (trace wins), but onTraceEnd reversed the order (processor wins). Fix: match onTraceStart order at trace-end so per-invocation user policy wins consistently. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Both `on_trace_start` / `onTraceStart` now delegate the default-root stamp to the shared util landed in #3317. Preserves the LSO-3608 precedence guarantee (user-supplied ls_agent_type wins via key-presence check). Subagent-detection sites stay hardcoded — they use a 2-element subset of narrowing tags (excluding subagent), clearer inline than importing.

…nts JS test coverage Adds a shared non-root tag set to the util (matches the vocabulary that OAI-Agents subagent detection and Vercel narrowing inherit already need). OAI-Agents subagent-detection sites now import it instead of hardcoding a 2-value tuple. Python site also tightened via metadata setdefault. Adds 6 JS precedence tests that were previously Python-only: nested-parent no-stamp, trace_end merge precedence, processor-vs-trace merge order, subagent-overrides-root, subagent-preserves-narrowing, subagent-idempotent.

…econciled

Consolidates the duplicated helper across wrap.ts and telemetry.ts into a shared experimental/vercel/_agent_type.ts. Adds narrowing-parent-tag inheritance (middleware / subagent / compaction) ahead of the existing run_type === "tool" heuristic. Fixes guardrail LLM spans self-stamping "root" and leaking into main conversation via the messages/v2 parser's root-privilege short-circuit. Fixes LSO-3681. Related LSO-3607. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

resolveLsAgentType now returns LsAgentType | undefined. Undefined means "don't stamp" — caller should skip the field so the run has no ls_agent_type on its own metadata (Vercel-safe explicit inheritance covers the propagation gap). Precedence: 1. Any known user-supplied parent tag (root/middleware/subagent/compaction) inherits explicitly (Vercel inner spans don't reliably use traceable's outer_metadata mechanism). 2. parent.run_type === "tool" -> subagent (Vercel-specific convention). 3. undefined when nested with no user signal. 4. "root" only at true top-level (no parent runtree). Adds lsAgentTypeMetadata() spread helper so all 5 call sites cleanly skip stamping when the resolver returns undefined — avoids `ls_agent_type: undefined` landing as a metadata value. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ritance Existing telemetry vitest expects a chain nested inside a tool span to be tagged subagent even when the tool span itself carries ls_agent_type=root (default-stamped or user-supplied). The previous ordering inherited root first and never reached the run_type=tool check. New precedence: 1. Narrowing tag (middleware/subagent/compaction) always wins. 2. parent.run_type === "tool" -> subagent (structural override of root). 3. Parent tagged root inherits explicitly. 4. No parent runtree -> root. 5. Otherwise undefined. Renames the constant to NARROWING_TAGS to match the semantic (root is not in the set of tags that beat structural detection). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

`traceable({ tracingEnabled: false })` stores a ContextPlaceholder (not a RunTree) in ALS. The previous loose type guard treated any non-null object as a runtree, so a nested LangSmithTelemetry({ tracingEnabled: true }) creating a new root run took the "nested-untagged" branch and returned undefined — the new root lost its default ls_agent_type=root stamp. Replace the ad-hoc check with `isRunTree` from run_trees.js. The placeholder now falls back to "root", matching the semantic that this IS the trace root for the newly-enabled telemetry. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

_agent_type.ts drops local LsAgentType/NARROWING_TAGS in favor of the shared util (LsAgentType, NON_ROOT_LS_AGENT_TYPES, resolveDefaultLsAgentType, lsAgentTypeMetadata). No-parent → root now delegates to the util; the Vercel-specific composition (narrowing inherit, run_type=tool, root inherit) stays local because AI SDK inner spans bypass traceable's outer_metadata propagation. Exports renamed to avoid collision with util's lsAgentTypeMetadata(tag): - resolveLsAgentType → resolveVercelLsAgentType - lsAgentTypeMetadata(parent) → vercelLsAgentTypeMetadata(parent) 5 caller sites updated (wrap.ts × 4, telemetry.ts × 1). Test suite renamed imports + added null-on-parent case for precedence parity with the OAI-Agents suite.

open-swe

Open SWE Review found 1 potential issue.

Open in Web • View Open SWE trace

…e-inherit

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Yoshi (ybathula707) marked this pull request as ready for review August 3, 2026 19:39

open-swe Bot reviewed Aug 3, 2026

View reviewed changes

Comment thread js/src/experimental/vercel/_agent_type.ts

Yoshi (ybathula707) and others added 16 commits August 3, 2026 16:15

style: trim verbose comments to match codebase single-line style

57c3833

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

style(js): apply oxfmt on Gap B test file

00d5b13

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

style: trim verbose comments to match codebase single-line style

4bfc83e

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

docs: reframe metadata merge comment around the three sources being r…

4293113

…econciled

docs: reframe on_trace_end merge comment around why we re-merge

76d6672

style: trim verbose comments to match codebase single-line style

1f190b9

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Yoshi (ybathula707) force-pushed the yashaswibathula/lso-3681-vercel-ls-agent-type-inherit branch from 1685ddb to f754038 Compare August 4, 2026 01:23

Yoshi (ybathula707) changed the base branch from main to yashaswibathula/lso-3608-oai-agents-ls-agent-type August 4, 2026 01:23

docs: reframe _agent_type.ts comment around config-build timing

d863759

open-swe Bot reviewed Aug 4, 2026

View reviewed changes

Comment thread js/src/experimental/vercel/_agent_type.ts Outdated

Jacob Lee (jacoblee93) approved these changes Aug 4, 2026

View reviewed changes

Base automatically changed from yashaswibathula/lso-3608-oai-agents-ls-agent-type to main August 4, 2026 23:02

Yoshi (ybathula707) and others added 2 commits August 4, 2026 16:05

Merge branch 'main' into yashaswibathula/lso-3681-vercel-ls-agent-typ…

d115159

…e-inherit

style(js): oxfmt trailing whitespace in _agent_type.ts

d95bb7c

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Yoshi (ybathula707) merged commit 9b047e3 into main Aug 5, 2026
37 checks passed

Yoshi (ybathula707) deleted the yashaswibathula/lso-3681-vercel-ls-agent-type-inherit branch August 5, 2026 00:20

Navigation Menu

Uh oh!

fix(js): inherit parent's ls_agent_type in Vercel _getLsAgentType helper - #3319

fix(js): inherit parent's ls_agent_type in Vercel _getLsAgentType helper#3319
Yoshi (ybathula707) merged 19 commits into
mainfrom
yashaswibathula/lso-3681-vercel-ls-agent-type-inherit

Yoshi (ybathula707) commented Jul 31, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jul 31, 2026 •

edited

Loading

Uh oh!

open-swe Bot left a comment

Uh oh!

Uh oh!

open-swe Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

Yoshi (ybathula707) commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

How

Test Plan

Uh oh!

github-actions Bot commented Jul 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

JS perf benchmark

Base64-heavy payload

Structural payload

Uh oh!

open-swe Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

open-swe Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Yoshi (ybathula707) commented Jul 31, 2026 •

edited

Loading

github-actions Bot commented Jul 31, 2026 •

edited

Loading