v0.9.0

This release adds GLM as a supported backend and ships a set of correctness and platform fixes.

🚢 Highlights

GLM backend

You can now run tasks on GLM (Z.ai) as both the implementation executor and the review supervisor. glm runs the Claude Code binary against GLM's Anthropic-compatible endpoint, so it works anywhere a Claude backend does — executor.cli, environment.yaml executor.default_cli / supervisor.default_cli, and the daemon --supervisor flag.

• Set your Z.ai token as glm_api_key in daemon.yaml (created with 0600 permissions). The token is used only in the executor/supervisor child process; it never appears on the command line or in run evidence.
• If glm is selected without a token, Galley fails fast with a clear configuration error instead of an opaque failure mid-run.

This enables a low-cost implementation model paired with a frontier review supervisor.

Fixed

• Tasks queued with a .yml extension are now executed instead of being silently ignored.
• task list and task show now resolve the running daemon's root the same way as queue / requeue.
• Supervisor needs_revision / hard_stop verdicts are no longer discarded over a finding's blocks_acceptance flag.
• scope.forbidden_paths can no longer be bypassed by letter case on case-insensitive filesystems, and preflight ignores .git when detecting creator changes.
• Windows: daemon stop/verification and interrupted-task recovery now work, and child process trees are terminated on timeout/cancel.

Packaging

• The Claude and Codex Galley plugins are now versioned 0.1.20, with task-authoring, setup, and profile guidance and bundled schemas updated for the glm backend.