fix(cli): pass target_model to resolve_runtime_provider in _ensure_runtime_credentials

[joviur]

Summary

_ensure_runtime_credentials() called resolve_runtime_provider() without passing target_model, so the resolver derived api_mode and base_url from model_cfg.default instead of the active model. When the config default uses anthropic_messages (e.g. minimax-m3 on opencode-go) and the user runs hermes chat -m <chat_completions_model>, the stale anthropic_messages mode strips /v1 from the base URL and every non-MiniMax model 404s.

Root cause

hermes_cli/cli_agent_setup_mixin.py:_ensure_runtime_credentials (line 41-45) called resolve_runtime_provider(requested=..., explicit_api_key=..., explicit_base_url=...) without target_model. The resolver then fell back to model_cfg.get("default") to derive api_mode, picking the wrong mode for any model that differs from the config default.

resolve_runtime_provider already accepts a target_model kwarg (added in #16878 / PR #16890 for the /model switch path) — it just wasn't being passed here.

Fix

Pass self.model as target_model so the resolver derives api_mode and base_url for the model actually being used.

         _primary_exc = None
         runtime = None
         try:
+            # Pass the active model so the resolver can derive the correct
+            # api_mode / base_url for it (e.g. opencode-go needs to pick
+            # ``chat_completions`` for non-MiniMax models even when the config
+            # default is a MiniMax model that would imply ``anthropic_messages``
+            # and strip ``/v1`` from the endpoint).
+            _target_model = getattr(self, "model", None) or None
             runtime = resolve_runtime_provider(
                 requested=self.requested_provider,
                 explicit_api_key=self._explicit_api_key,
                 explicit_base_url=self._explicit_base_url,
+                target_model=_target_model,
             )
         except Exception as exc:
             _primary_exc = exc

Why it's safe

• target_model is already an optional kwarg of resolve_runtime_provider (since Bug: /v1 stripped from base_url when switching models via /model with opencode-go provider #16878 / PR fix(opencode): re-derive api_mode per target model on /model switch #16890, merged in v2026.4.30).
• getattr(self, "model", None) or None is defensive: if self.model is missing or empty, it passes None and the resolver uses the default (previous behavior).
• Does not touch resolve_runtime_provider or runtime_provider.py.
• Does not touch the self.api_mode = "chat_completions" initialization in cli.py:3600.
• Backward-compatible: for minimax-m3 as default, the resolver now explicitly receives target_model="minimax-m3" and returns anthropic_messages — same result as before, but deterministic instead of by coincidence.

Verification

• New regression test test_ensure_runtime_credentials_passes_target_model_to_resolver fails on pre-fix code, passes with the fix.
• 24/24 tests in test_cli_provider_resolution.py pass.
• 1118/1119 in tests/cli/ + related tests/hermes_cli/ suites pass (1 pre-existing flaky failure in test_resume_quiet_stderr.py, unrelated — passes in isolation).
• E2E: hermes chat -m qwen3.7-plus -q "ok" goes from HTTP 404 to a valid response; minimax-m3 continues to work.

# Before fix:
hermes chat -m qwen3.7-plus -q "ok"    → HTTP 404
hermes chat -m kimi-k2.7-code -q "ok"  → HTTP 404
hermes chat -m deepseek-v4-pro -q "ok"  → HTTP 404

# After fix:
hermes chat -m qwen3.7-plus -q "ok"    → valid response ✅
hermes chat -m kimi-k2.7-code -q "ok"  → valid response ✅
hermes chat -m deepseek-v4-pro -q "ok"  → valid response ✅
hermes chat -m minimax-m3 -q "ok"      → valid response ✅ (unchanged)

Related

• Fixes [Bug]: hermes chat -m <model> uses stale api_mode from config default → 404 on opencode-go #54147
• Bug: /v1 stripped from base_url when switching models via /model with opencode-go provider #16878 — same bug family, fixed the /model switch path but not the CLI single-query path.
• delegate_task 404 when delegation model differs from main model on opencode-go/opencode-zen #15319 — same root cause in delegate_tool.py (missing target_model), fixed with the same one-line pattern.
• fix(cli): refresh runtime snapshot after _ensure_runtime_credentials (#53234) #53251 — complementary fix for runtime_override staleness in _init_agent (openai-codex); compatible with this change.

[tonydwb]

Code Review Summary

Verdict: LGTM (approve-ready, COMMENT due to token permissions)

Fix: passes the active model (self.model) as target_model to resolve_runtime_provider so the api_mode/base_url are derived for the model actually being used, not the config default.

Changes

• _ensure_runtime_credentials() now passes target_model=self.model
• 1 regression test verifying the resolver receives the correct model

Looks Good

• Clean one-line fix with clear test coverage
• The test correctly verifies both the resolver call and the resulting CLI state
• Edge case: getattr(self, "model", None) handles models without a model attribute

---

Reviewed by Hermes Agent

[teknium1]

Thanks for tracing the stale-model routing path. The primary call on current main is still missing target_model at hermes_cli/cli_agent_setup_mixin.py:41-45, and resolve_runtime_provider() uses target_model or model_cfg.default for OpenCode mode selection (hermes_cli/runtime_provider.py:2052-2053), so the main fix is sound.

Problems

• The fallback path has the same omission: hermes_cli/cli_agent_setup_mixin.py:60 resolves _fb_provider without a target model, before assigning self.model = _fb_model at line 67. It can therefore still derive api_mode from model.default instead of the selected fallback model.

Suggested changes

• Pass target_model=_fb_model in the fallback resolver call and add a fallback regression test. The linked #54365 already carries this small superset.

Automated hermes-sweeper review.

[teknium1]

Please apply the same propagation to the fallback resolver call: current main resolves _fb_provider without target_model before assigning self.model = _fb_model, so a fallback can still derive its route from model.default. Pass target_model=_fb_model there and cover that branch.

[joviur]

Closing this PR voluntarily — the fix was reimplemented upstream and I'm no longer maintaining this branch. Thanks!