fix(langchain): forward provider-reported cost to costDetails

[mittalpk]

Summary

Fixes #828.

Some OpenAI-compatible providers (e.g. OpenRouter) report a total cost directly on the response instead of leaving it to be derived from Langfuse's own model price table. LangChain surfaces this as usage.cost inside response_metadata, but CallbackHandler.handleLLMEnd never read it, so generations were ingested with empty costDetails and Langfuse fell back to a price-table lookup that fails for any model it doesn't recognize (most OpenRouter model slugs).

• Adds extractCostDetails(), mirroring the existing extractModelNameFromMetadata() pattern, wired into handleLLMEnd's costDetails attribute.
• Adds two integration tests: cost forwarding via a minimal fake chat model reporting response_metadata.usage.cost, and confirms costDetails is omitted (not a wrong value) when no provider cost is present.

A workspace issue this surfaced

packages/langchain had no devDependency pin on @langchain/core, so it resolved to a different, older copy (1.1.8) than the rest of the workspace (1.1.24) — two physically distinct AIMessage classes, so the existing instanceof AIMessage checks in CallbackHandler silently failed whenever the package was exercised through its own local dependency resolution rather than a consuming app's single hoisted install. This is likely why no integration test previously exercised handleLLMEnd's AIMessage-dependent extraction logic at all (only a tool-observation test existed).

Fixed with a devDependency pin (^1.1.24, matching the rest of the workspace) plus resolve.dedupe: ["@langchain/core"] on the integration/e2e vitest projects. This doesn't affect real consumer apps, which only ever install one copy of @langchain/core.

Separate one-line fix

.husky/pre-commit uses bash-only &> redirection inside a #!/bin/sh script. Under dash (default /bin/sh on Debian/Ubuntu), &> is misparsed as "background this command, then redirect a separate empty command," so the pnpm-on-PATH check always reported it missing regardless of reality, blocking pnpm precommit:check from ever running on a dash-based system. Split into its own commit since it's unrelated to the cost-forwarding fix.

Verification

pnpm build
pnpm test:integration   # 304 passed
pnpm test:unit          # 99 passed
pnpm lint               # clean
pnpm typecheck          # clean
pnpm format:check       # clean

AI-assisted: yes (Claude Code), disclosed per CONTRIBUTING.md. I reviewed and understand every line of the change, and did the debugging (including tracking down the dual-module-instance root cause) interactively rather than accepting a first-pass diff.

Greptile Summary

This PR forwards provider-reported LangChain costs into Langfuse generation data. The main changes are:

• Extract provider cost from response metadata.
• Add integration coverage for present and absent costs.
• Align and deduplicate the workspace LangChain dependency.
• Make the pre-commit package check POSIX-compatible.

Confidence Score: 4/5

Provider costs can still be dropped when consumers load separate LangChain module instances.

• The new extraction requires the provider message to share the handler's exact class constructors.
• Vitest deduplication fixes the workspace tests but does not control consumer dependency resolution.
• The remaining changes are contained and have focused integration coverage.

packages/langchain/src/CallbackHandler.ts and tests/integration/langchain.integration.test.ts

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
packages/langchain/src/CallbackHandler.ts:940-943
**Module Identity Drops Valid Costs**

When an application resolves the handler and provider against separate `@langchain/core` module instances, a valid provider-created message fails these `instanceof` checks. Its reported cost is then silently omitted, and the local Vitest dedupe cannot prevent this in consumer installations. Check the message shape instead of constructor identity.

```suggestion
        typeof generation["message"]?.response_metadata !== "object" ||
        generation["message"].response_metadata === null
```

### Issue 2 of 2
tests/integration/langchain.integration.test.ts:17
**Declaration Splits Import Block**

This class is declared before the existing helper imports later in the module, so those imports are no longer at the top as required by the repository rule. Move the class below the complete import block.

_{Reviews (1): Last reviewed commit: "fix: use POSIX-compatible redirect in pr..." | Re-trigger Greptile}

Greptile also left 1 inline comment on this PR.

Context used:

• Rule used - Move imports to the top of the module instead of p... (source)

Learned From
langfuse/langfuse-python#1387

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[vercel]

@mittalpk is attempting to deploy a commit to the langfuse Team on Vercel.

A member of the Team first needs to authorize it.

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Module Identity Drops Valid Costs

When an application resolves the handler and provider against separate @langchain/core module instances, a valid provider-created message fails these instanceof checks. Its reported cost is then silently omitted, and the local Vitest dedupe cannot prevent this in consumer installations. Check the message shape instead of constructor identity.

Suggested change

	!(
	generation["message"] instanceof AIMessage ||
	generation["message"] instanceof AIMessageChunk
	)
	typeof generation["message"]?.response_metadata !== "object" ||
	generation["message"].response_metadata === null

Prompt To Fix With AI

This is a comment left during a code review.
Path: packages/langchain/src/CallbackHandler.ts
Line: 940-943

Comment:
**Module Identity Drops Valid Costs**

When an application resolves the handler and provider against separate `@langchain/core` module instances, a valid provider-created message fails these `instanceof` checks. Its reported cost is then silently omitted, and the local Vitest dedupe cannot prevent this in consumer installations. Check the message shape instead of constructor identity.

```suggestion
        typeof generation["message"]?.response_metadata !== "object" ||
        generation["message"].response_metadata === null
```

How can I resolve this? If you propose a fix, please make it concise.

[mittalpk]

Good catch — fixed in e1138c8. Switched to checking that response_metadata is a non-null object rather than checking instanceof AIMessage/AIMessageChunk, so it no longer depends on message identity across module instances.