Skip to content

Tale v0.5.90

Choose a tag to compare

View all tags
@github-actions github-actions released this 14 Oct 01:23
v0.5.90
aca81bd
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

  • tale deploy upgrades a workspace from v0.5.77 across the automation writer protocol cutover. From v0.5.79 to v0.5.89 it refused every such workspace with "An existing protocol-1 database must upgrade through managed deployment apply…", and managed apply cannot take a tale deploy workspace, so these workspaces could not leave v0.5.77. Before it changes anything, tale deploy now checks the running backend's release, the database's migration history and the automation runs of every organization, and stops with the next step to take if one of them is not ready. Right before the new version starts, it stops the v0.5.77 backend while it holds a lock against new automation runs and has confirmed that none is unfinished, so no run is put on hold. A deploy interrupted after that point finishes when it is run again, tale deploy --dry-run reports what it would do, and tale rollback still refuses to go back to v0.5.77. The Upgrades guide has the steps.
  • tale update --version <x.y.z> updates again. Before, the CLI read it as a request for its own version, printed that and changed nothing; tale update -v <x.y.z> works on every release.
  • Choose the skills a chat reply uses. Type / at the start of a message, or choose Use a skill in the composer's + menu, to use one of your skills for that message; or check up to four under Skills in the model picker, and every reply of that conversation reads them before it answers. The assistant still reads a skill on its own when one fits the request (#4963).
  • Chat says why it could not make a file. A failed step names the cause under its row, in the person's language, such as All light sandboxes are in use or The program ended with an error, and screen readers read it with the row. When the file tools cannot be set up, a notice names the cause instead of the assistant suggesting a task. A PDF preview that cannot load says so instead of loading forever, a reply tried again keeps its file, and a reply on an Anthropic model that has used all its tool rounds finishes instead of failing (#4963).
  • Connect a coding agent from Tale itself. Connect a coding agent on the Connected agents page opens a guided setup: the commands for Claude Code and Codex at this deployment's address (Codex adds Tale and signs in with one command), the address for any other agent, a line that says when the agent has signed in, also when it signs in again, and a first prompt to try. The same setup opens from Edit with your coding agent in the automation editor, from Fix with your coding agent on a failed run with a prompt that names the run, from Create automation › With your coding agent, from the last step of the setup wizard, and at the top of Settings › API › MCP (#4958, #4959, #4960, #4962, #4965, #4966, #4967, #4968). Where an operator turned agent sign-in off, these places point to the MCP settings and the guide instead.
  • When the sign-in page that an application or a coding agent sends a person to cannot read the app's details, Cancel still works and Try again reads them again, instead of leaving buttons that do nothing (#4951, #4955). On a phone, the Connected agents list shows each agent's access and last use under its name. The user guides in English, German and French describe signing in from an agent, with screenshots of the consent page and the list (#4948, #4957).
  • A waiting task review no longer stays with a reviewer who was removed or lost edit access (#4396).
  • A custom API-key expiry stays on the chosen date across a daylight-saving change, dates the API cannot represent are disabled, and a dialog left open overnight checks the lifetime again when the key is created (#4582).
  • A Codex run that stopped on its provider's rate limit waits at least 60 seconds before it tries again, in tasks and in automation agent steps, instead of restarting at once and adding to the pressure (#4952).
  • An open row menu on the Sandboxes settings page no longer closes when the page updates (#4961).

Upgrade notes

  • From v0.5.89 to v0.5.90 no database migration or configuration change is needed, and the API contract stays at 3.44.0.
  • A tale deploy workspace that still runs v0.5.77 or earlier crosses the automation writer protocol cutover once, in a maintenance window. Let every automation run in every organization finish or cancel it, and pause schedules and triggers until the upgrade is done; confirm your off-host backup; then run tale update -v 0.5.90, tale deploy --dry-run and tale deploy, without --stop, --services or --override-all. The API is unavailable from the moment the old backend stops until the new version is healthy, while it applies every database migration since v0.5.77; if that outlasts HEALTH_CHECK_TIMEOUT, run tale deploy again. A workspace older than v0.5.77 goes to v0.5.77 first (tale update -v 0.5.77, tale deploy). With your own Compose files or the Kubernetes manifests, nothing holds this barrier for you: follow "Upgrade from v0.5.77 or earlier" in the Upgrades guide. CLIs before v0.5.90 ignore tale update --version; use -v.
  • This corrects the v0.5.89 notes, which left it out: since v0.5.89 (#4926) coding agents can sign in to the MCP endpoint without an API key unless an operator turns that off. TALE_MCP_CLIENT_METADATA_DOCUMENTS covers agents that publish a client metadata document, such as Claude Code and Codex, and needs outbound HTTPS from the backend to the agent's host; TALE_MCP_DYNAMIC_CLIENT_REGISTRATION covers agents that register themselves, such as Cursor and VS Code. Both default to true; set both to false to keep agents on API keys (Let coding agents sign in). People see and disconnect their agents under Settings › API › Connected agents.
  • In chat, Auto now picks a model from the message text alone; an attached document no longer moves a message to a stronger model, while an attached image still needs a model that reads images. Choose a model in the picker when a long document needs the stronger one.
  • Known limitations: built-in skills (docx, pdf, pptx, xlsx, visual-aspect-analyzer) reach an organization only when it is created, so an organization created on an earlier release without them, or one that deleted them, cannot choose them, and the assistant no longer claims they exist there (#4964). REST callers do not yet receive the reason when a chat reply could not make files.

API contract changes

None in this range. The contract stays at 3.44.0: 161 operations.

What's Changed

Full Changelog: v0.5.89...v0.5.90

Contributors

yannickmonney
Assets 8
Loading