[strange0g]

What

azure/kimi-k2.7-code reasons by default and returns its chain of thought in reasoning_content, but the catalog declared no effort control, so OpenCode exposed no effort variants for it.

Fix

providers/azure/models/kimi-k2.7-code.toml:

reasoning = { supported = true, toggle = true, effort = ["low", "medium", "high", "max"] }

The [api.chat-completions] table already sets compatibility = { reasoning_field = "reasoning_content" }.

Evidence (live, 2026-10-25, on a Foundry resource)

POST /models/chat/completions?api-version=2024-05-01-preview:

• no parameter → reasoning runs and reasoning_content is populated;
• reasoning_effort=none|low|medium|high|max → accepted; none gives reasoning_tokens: 0;
• the endpoint's own 400 names the enum and rejects minimal/xhigh (unlike DeepSeek V4 on the same surface);
• the model returns tool_calls.

Scope note

I checked the other Azure Kimi entries too. kimi-k2.5, kimi-k2.6, and the Kimi K3 deployments are not deployed on the resource I tested (every call returns "The API deployment for this resource does not exist"), so I left them unchanged — there is no live evidence for their controls here.

Validation

• bun validate — passes.
• bun run v1:check — reports the expected data change (listed for approval, per AGENTS.md): api reasoning_options: [] -> [toggle, effort low/medium/high/max] for azure/kimi-k2.7-code and its azure-cognitive-services alias.

[github-actions]

Action items

• [high] [violation] providers/azure/models/kimi-k2.7-code.toml:40 - Check: AGENTS.md Reasoning controls §3 and audit-reasoning-options Step 3 toggle rules — off via effort=none must be effort with none and no toggle. Why: The PR's own evidence says off is reasoning_effort=none on a single field, not a separate on/off field, matching same-surface peer providers/azure/models/deepseek-v4-pro.toml:16 which uses effort = ["none", "low", "high", "max"] with no toggle and notes off is effort=none rather than a toggle. The proposed toggle = true with effort = ["low", "medium", "high", "max"] invents a separate toggle and hides the documented none level. Action: Change to reasoning = { supported = true, effort = ["none", "low", "medium", "high", "max"] } with no toggle, and keep a leading top-of-file # Effort: reasoning_effort = none|low|medium|high|max comment noting off is effort=none.