fix(kimi): disable thinking and fail loudly on empty content

[klaus993]

Follow-up to #14. That PR cleared the 404, but Kimi reviews still never reached the PR.

What the test showed

Triggered /kimi on ethrex#7207 -> run 33441887902:

• Resolved ai-review-kimi.yml@refs/tags/v1 (88321d9), input model: kimi-k2.7-code
• Call Kimi API for review -> success, no 404
• Post review comment -> success
• Post error comment -> skipped
• Comments on the PR: none

So the model id is valid and the account is fine. The good news buried in there: kimi-k2.5 really was retired, and there is no billing or permission problem.

Why nothing posted

Per the chat API docs:

thinking.type -- Thinking switch (kimi-k2.7-code is always enabled and cannot be disabled)

and max_tokens (deprecated alias of max_completion_tokens) is "the maximum number of tokens to generate for the chat completion" -- reasoning tokens count against it.

With max_tokens: 4096, k2.7-code spent the budget on reasoning_content and returned an empty content. The workflow reads only .choices[0].message.content, and the posting step wraps its call in if (review && review.trim() !== '') -- so it succeeded doing nothing. Green run, silent PR.

Changes

1. Default to kimi-k2.6 with a new thinking input, defaulting to 'disabled'.

Restores the non-thinking behaviour kimi-k2.5 had, for which 4096 output tokens is a proven budget.

The trap worth naming: kimi-k2.6 also defaults to {"type": "enabled"}, so swapping the model string alone would hit the identical wall. The difference is that k2.6 accepts "disabled"; k2.7-code rejects it.

The field is an input rather than hardcoded because it is not portable across models -- k2.7-code only accepts "enabled", and kimi-k3 always reasons and uses reasoning_effort instead. Pass thinking: '' to omit the field entirely:

```
jq '... } + (if $thinking == "" then {} else { thinking: { type: $thinking } } end)'
```

2. Treat empty content as an error, not a silent success.

Reports finish_reason and whether reasoning_content was present, so the next occurrence explains itself. Same silent-failure class #13 just fixed for Codex.

Verification

jq payload construction exercised for thinking = disabled / enabled / '' (field correctly omitted on the last). Empty-content branch tested against three response shapes: a real review (posts), empty content with reasoning present (errors, finish_reason=length), and whitespace-only content (errors). YAML parses.

Not yet verified end-to-end against the live API -- that needs the merge and a tag move. I would re-run /kimi on ethrex#7207 straight after.

Note

v1 currently points at 88321d9, which has the broken k2.7-code default. It needs moving again once this lands.