[nik13]

Summary

Keep and repair @traceai/cohere; do not delete it. The old integration missed generated v2 clients and all chatStream calls, and treated SDK camelCase fields as wire-format snake_case.

• Patch the actual generated v2 prototype reached through CohereClient.v2, covering both client.v2 and the methods bound by CohereClientV2. Keep CohereInstrumentation / manuallyInstrument working.
• Trace streaming until terminal usage, exhaustion, failure, iterator return or abort. Aggregate text and tool-call deltas without changing events, controllers, errors, or HttpResponsePromise.withRawResponse().
• Fix chat history, tool calls, token usage and rerank score extraction. Preserve EMBEDDING spans and use RERANKER for reranking instead of treating everything as an LLM call.
• Respect suppression, parent context, concurrent calls, multiple SDK copies, idempotent patching and disable/enable, including already-bound v2 methods.
• Apply privacy controls to raw payloads, Cohere-specific fields, tool content, error details and embedding vectors. Do not invent model names.
• Bound the SDK peer range to 7.x/8.x; pin offline fixtures to 7.0.0 and 8.3.0. Fix the package's native ESM entry and replace the README's broken exporter example with maintained package setup.

Why this route

Cohere 8.3.0 binds v2 methods to a separate generated client, rather than the v1 prototype: SDK source. The public client.v2 accessor exposes that same generated implementation: SDK source. Prototype discovery constructs an inert client with no I/O or environment credential lookup. Instrumentation must be configured before constructing customer clients.

The existing package API is retained instead of requiring customers to copy helper code or migrate to a different client wrapper. Endpoint scope remains chat, chatStream, embed, and rerank.

Verification

Before the implementation change, the real-SDK reproducer failed all three cases: v2 emitted zero spans; v1 history extraction omitted earlier messages; the global patched flag also prevented a second SDK copy from being instrumented.

After the change:

• Frozen filtered dependency install passed; unrelated lockfile resolutions are preserved.
• CJS, ESM and ESNext TypeScript builds passed.
• 71 tests passed (35 new): real SDKs 7.0.0 and 8.3.0 against localhost, v1/v2 streaming, embeddings/rerank typing, request-option forwarding, HTTP/parser/iterator errors, early return, abort, suppression, parenting, concurrency and enable/disable.
• Fresh-process CommonJS automatic instrumentation and native ESM manual instrumentation passed.
• Actual OTLP/HTTP export assertions passed for span kinds, usage, hidden content/errors and vector redaction.
• Package tarball built successfully.

All local build/test/package commands ran under the supplied outbound-network-denied sandbox, with localhost permitted. No vendor/model API calls or real credentials were used. The dedicated CI job repeats the tests inside a network namespace with only loopback enabled.

pnpm install --frozen-lockfile --ignore-scripts --filter '@traceai/cohere...'
pnpm --filter '@traceai/cohere...' --workspace-concurrency=1 run build
pnpm --filter '@traceai/cohere' exec jest --runInBand
pnpm --filter '@traceai/cohere' pack --pack-destination /tmp/cohere-package

E2E coverage

not applicable (no product e2e/ harness) · flows hit: n/a · marker: n/a · verdict: not applicable

The real-SDK localhost and OTLP wire tests are the integration regression gate. No live Cohere verification is claimed. SDK 7.0.0 has no chatStream; stream coverage uses 8.3.0.

Release and docs

Release-please owns package versioning; no versions or changelog entries were hand-bumped. Package README changes ship with this fix. No separate docs-site PR is opened against an unreleased package change. This PR does not merge, tag or publish a release.

🤖 Generated with Claude Code