v0.12.0

Minor Changes

• 5584716 Thanks @cevr! - Counsel and okra image now use the newest GPT Sol model that the Codex account can use, instead of a pinned model, so a new Sol release needs no okra release. models.dev gives the release order (cached in ~/.okra/models.json for a day), and the Codex CLI's model list filters it. When Codex rejects a model for a ChatGPT account, the command retries with the next older candidate, down to gpt-6-sol. The image default changes from gpt-5.5 to the newest usable GPT Sol.

• 46ad091 Thanks @cevr! - Counsel now writes manifest.json into each run directory, with a new model field. It names the model that answered: the fallback model after a Codex model rejection, and the concrete Claude model behind the opus or fable alias. The manifest also carries the run status and, for a failed run, the failure reason. --dry-run shows the requested model in its invocation preview.

• a8b9248 Thanks @cevr! - Remove the --image-model flag from okra image. The Codex backend always replaces the image tool's model with its own (gpt-image-2-codex), so the flag had no effect there, and the paid API route already rejected it. To use Flare or Sunburst, pass --model gpt-image-2.5-flare or --model gpt-image-2.5-sunburst, which uses the OpenAI Images API.

Patch Changes

• c666534 Thanks @cevr! - Counsel now fails when the other agent does not answer, even when that agent exits 0. A Codex usage limit (turn.failed) and a Claude API error (is_error result, such as a 529 overload or a usage limit) now give exit code 1, print the reason on stderr, and record it as failure in the run manifest. Before, counsel exited 0 with an empty codex.md, or wrote the Claude error text into claude.md as if it were the answer.

• f15e342 Thanks @cevr! - okra image on the Codex route now asks for the --size in the prompt, because the Codex backend ignores the image tool's size setting. Portrait and landscape sizes now come back exact. When the result still differs (a square comes back as 1254x1254), okra prints a note on stderr and points to an OpenAI image model for an exact size.