Skip to content

Tale v0.5.100

Choose a tag to compare

View all tags
@github-actions github-actions released this 06 Nov 00:55
v0.5.100
9be27c4

Highlights

  • A task run on Claude Code that waits for model capacity now goes on waiting while the provider still refuses the model. Since v0.5.99, when the provider refuses a run because its model is at capacity, every task run of the organization on that model waits, and when the wait ends one run starts first; the others followed once it had worked for two minutes. Claude Code itself retries a refused call for about three minutes before it gives up, so that first run passed its two minutes while the provider was still refusing it. The waiting runs then started in waves into the same refusal, the first run's own refusal began a new wait of one minute, and the twelve-hour limit never ended it. Now every call that the provider refuses and Claude Code retries keeps the others waiting: they follow once the first run has worked for two minutes past its last refused call, or as soon as any run on that model finishes. While the model stays at capacity, the wait grows from one minute to 2, 4, 8 and then 15 minutes, and the runs read Waiting for model capacity and take no worker. Twelve hours after the first refusal, the model is no longer waited for, and its runs fail with The AI model stayed at capacity for twelve hours. An automation's agent step refused this way is unchanged: it retries a minute later, as one of its three retries (#5350). See Handle waiting and failed runs.
  • Qwen Code no longer runs a model served through Anthropic's Messages API: a Claude model, or any model of a provider whose API format is Anthropic Messages API. Its tools that take no arguments lose their schema on that API, so every turn of such a run failed with 400 tools.2.custom.input_schema: Field required. With Qwen Code as the Agent runtime, the model pickers no longer offer such a model: for a project agent on the project's Agents tab, in an automation's agent node, and for the standard agent under Settings › Governance › Models. Tale refuses to save a project agent with such a pair, and a Qwen Code run that names such a model now fails before its sandbox starts, with a reason that names the model and Qwen Code. Under Settings › AI providers, Agent runtimes no longer counts such models for Qwen Code: where your providers serve no other model, it reads No model yet, and its note names the Anthropic Messages API. The MCP endpoint's list_models no longer names qwen-code among such a model's harnesses. Nothing changes for Qwen Code's other models or for the other agent runtimes. The restriction stays until Qwen Code or Tale's model gateway ships a fix (#5353). See Choose an agent runtime.
  • The image tool now offers the image model only the pictures it takes, in chat and in a standard agent run without a worker. An image model that draws from the prompt alone is offered no picture, and the assistant is told to say that it can only draw a new one when you ask it to change a picture. Tale counts an image model that an AI gateway lists without describing it, such as gemini-3.1-flash-image, as such a model. A model that takes pictures is offered as many as its catalog allows, at most four, and a model that only edits must be given one in every request; with none to start from, the assistant is told to say that an administrator would choose another image model. Before, every image model was offered up to four pictures, and the assistant learnt a model's limit only when Tale refused the request, in an image step that can read Couldn't run as asked. Such a refusal is the likeliest cause of the occasional failed image steps in chat reported in #5368, which stays open until its cause is confirmed. An agent that works in a worker still offers up to four pictures and learns a model's limit from Tale's refusal (#4875). See Allow image generation.

Upgrade notes

  • This release adds no migration, to the application database or to the knowledge database, and the API contract stays at 3.52.0.
  • A project agent, an automation's agent node or the standard agent that runs on Qwen Code with a Claude model, or with a model of a provider whose API format is Anthropic Messages API, failed every turn before the upgrade and now fails at its start. Choose another Agent runtime for it, such as Claude Code, or a model that Qwen Code can run. On the project's Agents tab, such an agent's model field reads Choose the model it calls, and a save that keeps the old model is refused with Pick a model for this agent. An automation's agent node says that no connected provider offers the saved model for its agent runtime. For the standard agent, an Admin chooses another runtime, another model or Automatic under Settings › Governance › Models; until then it starts for no one, and its tasks say that the chosen model isn't one the person can use. With Qwen Code and Automatic, the standard agent now picks among the models Qwen Code can run.
  • No configuration, limit, credential or sign-in setting needs to change beyond these: the other fixes take effect with the new images.

API contract changes

None in this range. The contract stays at 3.52.0: 163 operations.

What's Changed

Full Changelog: v0.5.99...v0.5.100

Contributors

tale-code-bot
Assets 8
Loading