Session error path drops cause and stack from ProviderInitError, hiding root cause #24430

Description

When Provider.resolveSDK throws InitError({providerID}, {cause: e}) (packages/opencode/src/provider/provider.ts:1522), the session pipeline catches it and re-emits via new NamedError.Unknown({message: err.message}).toObject(). err.cause and stack are discarded.

Consumers (e.g. cli/cmd/github.ts:986) only see:

Agent error: { name: "UnknownError", data: { message: "ProviderInitError" } }

This makes CI failures undebuggable. The real failure (NPM install error, missing peer dep, network failure, etc.) lives on err.cause and is silently dropped.

Expected: preserve the original named error (name: "ProviderInitError") and pass cause chain through to the consumer, e.g.

new NamedError.Unknown({ message: err.message }, { cause: err.cause })

and serialize cause chain in toObject(). Or stop re-wrapping ProviderInitError as UnknownError at the session boundary.

OpenCode version

1.14.25 (also reproduces on 1.14.22)

Steps to reproduce

1. Trigger any failure inside resolveSDK — easiest path: cloudflare-ai-gateway/openai/gpt-5.4 on a fresh CI runner (see loadOptions(dir) in Npm.add reify path fails on CI runners → silent ProviderInitError for dynamic providers #24431)
2. Observe CI log shows only UnknownError: ProviderInitError — the underlying InstallFailedError and its cause are gone.

Operating System

Linux (GitHub Actions ubuntu-latest), macOS 25.4

Impact

Every cf-ai-gateway repo that hit the 1.14.x install regression saw an unactionable error. Multiple Cloudflare repos (workers-sdk, workerd, kumo, containers, vinext) pinned older opencode versions without an upstream issue because the symptom was not diagnosable.