Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 293 Agent status: 2 with issues Updated Oct 7, 2026
AgentGid

CodeWhale changelog: what's new each month

Every stable CodeWhale release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

September 2026

3 releases: 0.9.12 → 0.10.0

September brought new model integrations, enhanced plugin management, and the Computer Use capability with desktop control features. The month focused on expanding provider support and security controls for organizational use.

Highlights

  • Computer Use plugin now available with 38 tools for automated desktop interactions.
  • Alibaba Model Studio added as a new data-driven provider.
  • Organization-level plugin allowlists for managed policy control.
  • Concentrate gateway introduced as a first-class BYOK Responses option with live models.
  • Desktop and computer-control tools restricted from child process inheritance for security.

New

  • Computer Use plugin with 38 bundled tools for desktop automation.
  • Alibaba Model Studio provider integration.
  • Concentrate BYOK Responses gateway with live model support.
  • Pet terminal mode accessible via /pet command.
  • Computer Use 0.3.1 as notarized Mac application.
  • Organization-level plugin allowlist policies for managed environments.

Improved

  • Computer Use download page on website now displays state correctly.
  • Session-scoped patch approval now scoped to individual files.
  • Desktop control and computer-use tools no longer inherit to child processes.

Fixed

  • Computer Use download page state resolution.
  • Plugin inheritance behavior for security-sensitive tools.

August 2026

8 releases: 0.9.4 → 0.9.11

August brought expanded model provider support, improved terminal display, and better session management to CodeWhale. The month focused on adding first-class integrations with new AI providers and fixing edge cases in existing features.

Highlights

  • Mistral AI and Codestral models are now first-class providers alongside existing options.
  • Grok 4.6 became the default xAI model, and OrcaRouter added as a named provider.
  • Web search now works without an API key using Firecrawl Cloud by default.
  • Terminal display improved with bounded, wrapping hover popovers for slash-command completion in narrow terminals.
  • Session history now supports append-only tree navigation with /tree, /branch, and /fork commands.

New

  • Model auto-selection via model="auto" for prompt-based tier selection.
  • Runtime API controls for persistent goals and bounded memory inspection.
  • Append-only session-tree history with branching and forking support.
  • Unified Fleet roster for built-in dispatch postures with pinned indicator.
  • Headless Bash capability to transfer explicitly requested persistent Unix services.
  • Reasoning-effort normalization that reads published model capabilities.
  • Portable, secret-free config export and import with reviewable plan.
  • DeepSeek-V4 and deepseek-v4-flash-vision-exp discovery and selection.

Improved

  • Web builds on main now emit actionable manual-deploy reminders.
  • Prompt-cache prefix is now pinned for the session.
  • Tool loop no longer causes issues with prompt-cache management.
  • Context-window honesty displays at every surface for accurate token accounting.
  • Output-ceiling honesty for Anthropic-family models now reports accurate limits.
  • Plugin compatibility is now managed per-component for better granularity.
  • Sudo and setuid helpers work again for wheel-group administrators.
  • Provider output limits now take precedence over conservative defaults.

Fixed

  • Abort-class process deaths no longer poison the terminal.
  • Bash tool no longer wedges when complete-output exceeds limits.
  • Slash-command completion row truncation in narrow terminals.
  • MCP Registry catalog no longer ships unnecessary full data.
  • Terminal launch improved by using codewhale.bat instead of raw .exe.

July 2026

5 releases: 0.8.67 → 0.9.3

July brought workflow automation, credential security improvements, and expanded model support to CodeWhale. The month saw steady progress from version 0.8.67 through 0.9.3 with new commands and better permission controls.

Highlights

  • New /workflow command lets you create and manage automation workflows directly in CodeWhale.
  • Saved workflows now become slash commands, making reusable automation easily accessible.
  • Added /preview-request (with /dryrun alias) to preview requests before execution.
  • Native support for DeepSeek v4-flash model via DeepSeek's Responses API.
  • Improved credential security with redaction of configured, environment, and file-backed credentials.

New

  • /workflow command (with /workflows and /wf aliases) for managing workflows.
  • Constitution-first setup wizard providing a unified /setup shell.
  • Bare /goal command to declare goals from conversation without an active goal.
  • /preview-request command to preview requests before running them.
  • /permissions command to list active user permission rules and their sources.
  • Native DeepSeek v4-flash model support.
  • Per-turn maxtoolcalls enforcement at the engine admission gate.
  • Pipe-only codewhale auth print-api-key handoff for explicit provider selection.

Improved

  • Saved workflows now appear as slash commands under .workflow.js files.
  • Model selection in /model now properly operates fleet workers.
  • Automatic compaction thresholds now anchored to the route's spendable input.
  • TUI submit failures now kept inside the shell via custom-provider route.
  • Improved Kimi K3 model-picker disambiguation in the UI.
  • Enhanced model-facing agent tool role schema with closed type property.
  • /compact command now accepts a focus parameter for manual compaction.
  • Better handling of slash commands, hotbar actions, and CLI entrypoints across Lane/Fleet.

Fixed

  • Legacy model auto mode no longer incorrectly elects a network classifier.
  • Fixed credential redaction for environment and file-backed credentials.
  • Corrected OpenAI cost estimation to fail closed instead of reporting inaccurate base-rate estimates.
  • Fixed model-facing agent tool role schema type property enforcement.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. CodeWhale: pricing, features and alternatives · All changelogs