fix(opencode): preserve /v1 in base_url when switching from anthropic_messages to chat_completions models

[quannon]

Summary

When using the opencode-go provider with a config default model that uses anthropic_messages mode (e.g. minimax-m2.7) and switching to a chat_completions model (e.g. deepseek-v4-flash) via the /model slash command, the base_url loses /v1, resulting in an HTTP 404 error.

Fixes: #16878

Root Cause

In runtime_provider.py, the non-pool "API-key providers" block (~lines 1213-1237), when the credential pool is exhausted:

1. configured_mode = _parse_api_mode(model_cfg.get("api_mode")) reads "anthropic_messages" from the stale config
2. This stale mode is applied before the opencode-go/zen api_mode detection can run
3. The opencode block was an elif, so it never ran when configured_mode was truthy

Fix Applied

Promoted opencode-go/zen detection to an explicit if block that runs before the configured_mode override, so it always correctly detects api_mode based on the target model.

Same change applied to _resolve_runtime_from_pool_entry for consistency.

Verification

# default=minimax-m2.7 → switch to deepseek-v4-flash
resolve_runtime_provider(requested="opencode-go", target_model="deepseek-v4-flash")
# → api_mode="chat_completions", base_url="https://opencode.ai/zen/go/v1" ✓

# reverse: default=deepseek-v4-flash → switch to minimax-m2.7
resolve_runtime_provider(requested="opencode-go", target_model="minimax-m2.7")
# → api_mode="anthropic_messages", base_url="https://opencode.ai/zen/go" ✓

Tests

12/12 passing: python -m pytest tests/hermes_cli/test_model_switch_opencode_anthropic.py -v

[teknium1]

Thanks for the fix, @quannon! The bug you identified is real and your analysis of the root cause is exactly right.

Unfortunately, an independent fix for the same issue landed on main about 7 hours before this PR was opened — likely authored in parallel:

• Commit: b52ceccfa — fix(opencode): re-derive api_mode per target model on /model switch
• Shipped in: v2026.4.30
• File: hermes_cli/runtime_provider.py — both _resolve_runtime_from_pool_entry (line 263) and the api-key-providers path (line 1295) were updated with the same structural change you proposed: promoting the opencode-zen/go detection to an explicit if block above the configured_mode override.

The fix also explicitly cites #16878. Since the same change is already live in the latest release, this PR is superseded.

This review was performed by the automated hermes-sweeper.