Fixed

• A parent turn that dispatched coordinated background subagents no longer runs the completion critic (or any lower "are we done?" gate) while that batch is still running — it waits for the batch to finish and reconcile instead of being judged prematurely, and resumes when results are deliverable (#679).
• A blank or whitespace-only retry_of on a task call is now treated as omitted, so a generated tool call carrying an empty retry id no longer trips coordinator retry validation for a nonexistent task. Real ids are trimmed and still validated strictly (#679).

Binaries for Linux (gnu/musl), macOS (x64/arm64), and Windows attach below once the release build finishes. Also on crates.io: cargo install dirge-agent.

Changed

• Shift+Enter inserts a newline in the input box instead of submitting, so you can write multi-line prompts. The newline gesture is a rebindable insert_newline command (defaults Shift+Enter, Alt+Enter, Ctrl+J) rather than a hardcoded key — remap it from keybindings. Alt+Enter and Ctrl+J work in every terminal; Shift+Enter needs the enhanced keyboard protocol (below). Plain Enter still submits, and Home/End continue to move within the current line.

Added

• keyboard_enhancement config (default on). Enables the terminal's enhanced keyboard (kitty) protocol on terminals that support it — kitty, Ghostty, WezTerm, foot, rio — so distinct chords like Shift+Enter actually reach the input editor. A no-op on terminals that don't advertise support (use Alt+Enter or Ctrl+J there); set false to disable.

Binaries for Linux (gnu/musl), macOS (x64/arm64), and Windows attach below once the release build finishes. Also on crates.io: cargo install dirge-agent.

Changed

• Unified finalization judge. The post-turn code reviewer is folded into the completeness critic as one judge call: it checks the task is done and reviews the run's diff for defects, then re-enters the loop with one consolidated follow-up the agent acts on. Previously the default advisory mode ran detached and surfaced findings — even high-severity ones — as a display-only notice the model never saw, so a review that flagged a real bug right after the agent said "done" changed nothing. Now any finding re-enters (high/critical must-fix, medium/low optional); advisory re-enters once, blocking persists until the diff is clean, off reviews completeness only.

Fixed

• /model <id> routes free-form ids to the right provider. It used to swap the live client only on an exact match against a provider's pinned model; any other id (a version bump, a typo, glm-4.6) was renamed on the active client and hit the wrong endpoint. Ids are now matched by family (glm-*, deepseek*, claude-*, gemini-*, gpt-*) and routed to a configured provider of that kind; an id whose family has no provider is refused with a warning. The banner also reads the live provider/model instead of re-resolving from CLI/config.
• Plugin prepare-next-run model swaps can hop providers. The swap builds and installs the target provider's client in place so the follow-up turn runs there, and no longer clobbers session.provider on a same-client rename.

Binaries for Linux (gnu/musl), macOS (x64/arm64), and Windows are attached below once the release build finishes. Also on crates.io: cargo install dirge-agent.