feat(ollama-cloud): add reasoning_effort support (xhigh->max)

[kshitijk4poor]

Summary

Hermes can now send reasoning_effort to Ollama Cloud's OpenAI-compatible /v1/chat/completions endpoint — but only for models that actually support thinking — unlocking DeepSeek V4's "Max thinking" tier via reasoning_effort: "max". Previously the bare ollama-cloud profile silently dropped all reasoning config.

Salvage of #29221 by @s010mn onto current main (cherry-picked, authorship preserved), plus one follow-up commit from me with three correctness fixes — all verified live against ollama.com.

Changes

• plugins/model-providers/ollama-cloud/__init__.py (@s010mn + my follow-up): OllamaCloudProfile overrides build_api_kwargs_extras to emit top-level reasoning_effort. xhigh/max → max; low/medium/high pass through.
• follow-up (me), all live-verified on deepseek-v4-pro / gemma3 / qwen3-coder:
  1. Capability-gated. Only models whose native /api/show capabilities list contains thinking get reasoning_effort. Resolved once per (model, base_url) in run_agent._supports_reasoning_extra_body via a cached probe (hermes_cli.models.ollama_model_supports_thinking), threaded into the profile as supports_reasoning — no live HTTP in the per-request path. Mirrors the existing LM Studio capability pattern. The original ignored the flag and emitted for every model.
  2. Disable actually disables. Ollama Cloud defaults to thinking ON and ignores extra_body.thinking:{disabled} (verified). The only off switch is top-level reasoning_effort: "none". The original returned ({}, {}) on disable → thinking stayed on. Now emits {"reasoning_effort": "none"}.
  3. Omit unrecognized effort. The original forwarded unknown strings verbatim, including minimal (a real Hermes level). Ollama Cloud hard-400s unrecognized values (accepted set: {low, medium, high, max, none}). Now omitted → model default.

Why the core touches

run_agent.py + hermes_cli/models.py gain the capability probe (where reasoning-capability resolution already lives, e.g. LM Studio). The plugin profile only consumes the resolved supports_reasoning flag — it doesn't reach into core.

Validation (live against ollama.com)

model / input	result
deepseek-v4-pro /api/show	capabilities: [...,"thinking"] → gated ON
gemma3:27b, qwen3-coder:480b	no thinking cap → no reasoning_effort
thinking + xhigh	top-level reasoning_effort="max"
thinking + enabled:false	reasoning_effort="none" (real off switch)
extra_body.thinking:{disabled}	ignored (thinking stayed ON)
reasoning_effort:"ultra"	HTTP 400 → justifies omitting unknowns

• 24/24 profile tests pass; 194 provider/transport tests unaffected; ruff clean.
• E2E verified through real imports + the real transport.

Credit

Salvaged from #29221 by @s010mn. Cherry-picked with authorship preserved; one follow-up commit on top.

[alt-glitch]

This was generated by AI during triage.

Related: salvage of #29221 (@s010mn), part of the ollama-cloud reasoning_effort cluster alongside OPEN #30262 (superset: also adds DeepSeek thinking-mode) and the broader max-effort work in #49355 / #49644. Depends on the AUTHOR_MAP chore #51481 (merge first so check-attribution passes). Maintainer to pick the canonical ollama-cloud implementation from this cluster.

[teknium1]

Merged via PR #65723 — your follow-up commit was cherry-picked onto current main with your authorship preserved in git history (5d9a72b7c). The base commit by @s010mn had already landed separately (221cd6024), so this salvage carried your three correctness fixes: capability gating, the real reasoning_effort:"none" off switch, and unknown-effort omission.

The live verification against ollama.com (especially discovering that extra_body.thinking:{disabled} is silently ignored) is exactly the kind of E2E rigor we want. Thanks!