[Feature]: Inject custom headers into outgoing LLM requests (OPENCODE_OUTBOUND_HEADERS)

[waiyong]

Is your feature request related to a problem?

When opencode talks to an LLM through a self-hosted gateway (LiteLLM, vLLM), that gateway often needs specific request headers — feature flags, routing hints, or tracing toggles such as Arize Phoenix's x-enable-phoenix-tracing. There is currently no way to set custom headers on opencode's outgoing LLM requests from this plugin.

The plugin already exposes a chat.headers-shaped surface conceptually (it owns the OTel integration), and @opencode-ai/plugin provides a chat.headers hook that fires for each outgoing LLM request. That hook is the natural, protocol-level place to add headers — no request interception or monkey-patching.

Describe the solution you'd like

Add a chat.headers hook with two settings:

• OPENCODE_OUTBOUND_HEADERS — comma-separated key=value headers to inject (parsed by the existing parseAttributePairs).
• OPENCODE_OUTBOUND_ENDPOINTS — comma-separated endpoint URLs. When set, headers are injected only for requests whose resolved URL matches by hostname; when unset, headers go to every provider (backward compatible).

Endpoint scoping matters because chat.headers fires for every LLM request, including third-party providers (Anthropic, OpenAI). Without scoping, an unintended header — or worse, a credential someone puts in the env var — would be sent to all of them. The resolved URL mirrors opencode's own resolution: non-empty provider.options.baseURL if overridden, else model.api.url.

Both settings would also be available via the plugin-tuple options form (outboundHeaders, outboundEndpoints), following the established option > env > default precedence from #65.

Describe alternatives you've considered

• Global fetch interception to add headers — rejected for the same reasons raised in [Feature]: native MCP trace-context propagation (SEP-414) — stdio support, transport-agnostic #62: it is a process-wide workaround that carries transport concerns better handled at the hook layer.
• Doing this in opencode core — the header need is deployment-specific (which gateway, which flags), so a plugin-level toggle is the right granularity.

Additional context

I have an implementation ready (handler extracted to src/handlers/chat-headers.ts, unit tests, README, all four checks passing) and will open a PR referencing this issue. This is scoped to custom headers only; W3C traceparent propagation is a natural follow-up but is a separate, opt-in concern (it would leak trace IDs to third parties by default), so I would like to keep it to a second PR.

One design question for you: endpoint matching is currently by hostname, which tolerates port/path differences. I am happy to switch to origin (scheme + host + port) if you prefer stricter matching.