fix(provider): recover stalled model streams

[dcadenas]

Issue for this PR

Closes #37580

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

A model endpoint can return response headers and then stop producing body data. OpenCode currently waits forever in that state, which can leave foreground Task parents stranded behind a stalled child.

This adds a provider-neutral stream-idle watchdog with a safe default. It starts with a 15-minute deadline, learns from successful streams per provider and transport, and keeps the adaptive deadline between 15 and 30 minutes. An explicit chunkTimeout remains a fixed override, while false, zero, or a negative value disables the watchdog.

The watchdog recognizes either an SSE response header or stream: true in the request. This matters because the direct ChatGPT Codex endpoint was observed returning a streaming response without Content-Type. Non-streaming requests remain unchanged. When a response omits Content-Type, the watchdog activates only when the JSON request explicitly sets stream: true, covering the observed response without applying adaptive timing to unrelated fetches.

A timeout becomes a typed retryable provider error and uses the existing retry and status flow. This does not change Task orchestration, durable session records, retry policy, or UI behavior.

Related work

• fix: time out non-SSE streams #39516 expands the existing opt-in chunkTimeout to every response body. This change instead provides a zero-config adaptive default and limits wrapping to requests identified as streaming, including responses without Content-Type.
• fix(opencode): make long-lived provider streams robust to silent SSE terminations #39970 detects incomplete EOF responses and gaps between parsed model events, including stalls hidden by SSE keepalives. Those changes are complementary. This change specifically handles transport-level body silence and does not require users to configure chunkTimeout.

How did you verify your code works?

• bun test --timeout 30000 test/provider/header-timeout.test.ts test/provider/stream-liveness.test.ts test/session/processor-effect.test.ts test/session/retry.test.ts — 72 passed
• bun typecheck in packages/core
• bun typecheck in packages/opencode
• Repository pre-push typecheck — 30/30 tasks passed
• An isolated foreground Task E2E completed normally, then completed again after a forced post-header child stall triggered the existing retry flow.
• The missing-Content-Type case is covered directly without proxy header normalization.

Screenshots / recordings

Not applicable; there are no UI changes.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, I found two potentially related PRs:

1. PR fix(opencode): make long-lived provider streams robust to silent SSE terminations #39970 - fix(opencode): make long-lived provider streams robust to silent SSE terminations

   • fix(opencode): make long-lived provider streams robust to silent SSE terminations #39970
   • Related: Addresses similar streaming timeout/termination issues

2. PR fix: time out non-SSE streams #39516 - fix: time out non-SSE streams

   • fix: time out non-SSE streams #39516
   • Related: Directly addresses stream timeouts for non-SSE providers

These PRs appear to be addressing overlapping concerns about streaming robustness and timeouts in provider responses. You may want to review them to ensure PR #40010 doesn't duplicate existing fixes or to coordinate any related changes.