Skip to content

v1.0.8

Latest
Latest

Choose a tag to compare

View all tags
@Waishnav Waishnav released this 25 Aug 22:09
· 47 commits to main since this release
v1.0.8
fe712e2

Highlights

This release rebuilds subagent execution around an on-demand local daemon. Agent sessions now survive CLI and MCP server restarts, reuse native provider runtimes, and expose clearer lifecycle and failure information.

Added

  • Run subagents through devspace-agentd, started automatically by agent commands. devspace serve is no longer required for local agent execution (#183).
  • Add native runtime support for Codex app-server, Claude Code, OpenCode, Pi, Cursor ACP, Copilot ACP, and Grok Build ACP (#183, #219).
  • Add devspace agents targets for listing usable providers and profiles.
  • Add devspace agents continue <id> for continuing an existing agent session.
  • Add compact --json output to agent commands for model and script integration.
  • Add daemon diagnostics through devspace agents daemon status, stop, and logs.
  • Add provider-level configuration for enabling agents and setting default models and effort levels.
  • Let devspace init configure ChatGPT, Coding Agents, or both while preserving existing configuration.

Improved

  • Keep provider runtimes warm and reusable while maintaining one durable DevSpace agent record across turns.
  • Allow different agents to run concurrently while limiting each agent to one active turn.
  • Use the current Git project as the workspace for direct CLI usage, with the current directory as the fallback outside Git.
  • Hide disabled or unavailable providers from model-facing catalogs.
  • Return structured agent failure codes with retryability information.
  • Standardize provider reasoning configuration on effort across profiles, configuration, persisted sessions, and CLI arguments (#212).
  • Rename the bundled subagent-delegation skill to subagents and focus it on the structured CLI workflow.
  • Improve sandbox and workspace authority enforcement across Pi, Claude, ACP, Codex, and OpenCode runtimes.

Fixed

  • Preserve agent state and provider session IDs across daemon and MCP server restarts.
  • Prevent concurrent clients from starting competing daemons for the same state directory.
  • Recover safely from stale daemon locks, sockets, interrupted startup, and runtime shutdown races.
  • Terminate provider process trees during shutdown and invalidate unhealthy pooled runtimes.
  • Preserve provider-specific errors and configuration diagnostics across daemon IPC.
  • Wait for OpenCode sessions to finish and traverse paginated message history.
  • Preserve Grok configuration diagnostics and validate supported model and effort values.
  • Keep provider availability checks from mutating configuration or triggering unwanted provider work.
  • Preserve existing setup choices when devspace init runs again.

Upgrade notes

  • Agent profile frontmatter and CLI usage now use effort instead of thinking. DevSpace migrates persisted agent session values, but custom profile files and scripts using thinking or --thinking must be updated.
  • The legacy boolean "subagents": true configuration remains supported. New configurations use explicit per-provider settings.
  • Codex now uses the installed codex app-server command instead of the bundled Codex SDK. Ensure codex is available on PATH, or set CODEX_COMMAND.
  • Existing logical agent IDs remain the identifiers used by show and continue.

Full changelog: v1.0.7...v1.0.8

Assets 2
Loading