cli: provider stream failures retry silently forever - looks like a hang with no console feedback

Summary

Selecting a model whose upstream endpoint fails (currently opencode-go/deepseek-v4.1-flash, which returns HTTP 503 "Endpoint is unavailable" from the Zen Go router) makes OpenCode retry the stream silently and indefinitely. The CLI produces no output and no error, so a 1.6-second upstream failure looks like an endless hang.

Environment

• opencode version: 2.0.3
• OS: Windows 10 Pro 10.0.19045 (MINGW64_NT-10.0-19045, x86_64)
• Terminal: Windows console, TERM=xterm-256color
• Shell: Git Bash (bash 3.6.10, MINGW64)
• Install/channel: latest, npm package @opencode/cli@2.0.3
• Active plugins: @franlol/opencode-md-table-formatter@latest, @prevalentware/opencode-goal-plugin (the goal plugin currently fails to load with a SchemaError — unrelated, noted for completeness)

Reproduction

1. Run against the shared background service:

   timeout 90 opencode run -m opencode-go/deepseek-v4.1-flash "Reply with exactly: OK"
   

2. Observe: no output, no error, process never exits. After 90s timeout kills it (exit 124) with 0 bytes of output.
3. Direct test bypassing OpenCode confirms the trigger — POST to https://opencode.ai/zen/go/v1/chat/completions with x-opencode-session header:
   • model deepseek-v4.1-flash → HTTP 503 in ~1.6s:

     {"error":{"type":"server_error","message":"Error from provider (Console Go): Upstream request failed: Endpoint is unavailable."}}
     

   • control model deepseek-v4-pro → HTTP 200 in ~2s, normal completion.

Expected Behavior

The client should surface an error such as "provider request failed (503): Endpoint is unavailable" and stop (or retry a bounded number of times with visible feedback), instead of waiting forever with no console messages.

Actual Behavior

• The client submits the message and holds GET /api/event and POST /api/session/<id>/wait open; the server never completes the session.
• ~/.local/share/opencode/log/opencode.log for the hung session shows no stream providerID=... entry, no retry entries, and no provider error — nothing between cli starting and the eventual client-side interrupt.
• Earlier logs (Aug 14) show the same class of failure with a silent retry loop: repeated stream providerID=opencode-go ... + llm runtime selected entries every ~7s for a minute+, never surfaced to the user.

Additional Context

• Reproduces consistently for this model; started right after the model appeared in the catalog (release_date 2026-09-10). Works fine with opencode-go/deepseek-v4-pro and opencode-go/deepseek-v4-flash (both verified end-to-end).
• Workaround: use opencode-go/deepseek-v4-pro or opencode-go/deepseek-v4-flash.
• Side observation: the official deepseek/deepseek-v4-flash provider also stalls (API returns 200 headers but the body never completes) — separate upstream issue.