v1.0.8

opencode-x v1.0.8

v1.0.8 is a large feature release with new agent workflows, Claude Code agent support, research tooling, checkpointed long-session memory, stronger provider compatibility, improved MCP/LSP behavior, and many TUI quality-of-life fixes.

Highlights

• Claude Code agent support: native support for Claude Code-style agent definitions, gated for safe rollout.
• Workflow engine: QuickJS-powered workflow system for scripted multi-agent automation.
• Research mode: new research tool and /research command for deeper investigation flows.
• Long-session checkpoints: checkpoint writer helps rebuild useful context after compaction.
• Smarter providers: OpenAI response chaining, cache retention, truncation support, stricter schema handling, and generic reasoning_options support.
• Better MCP/LSP: MCP roots/capabilities, local server cwd, structured tool output, and multi-root LSP diagnostics.
• TUI upgrades: transcript search, notification queue, OSC-8 links, workflow status, and editor chord fixes.

New features

Claude Code agents

• Added native Claude Code agent definition support.
• Integrated agent loading into opencode-x agent registry.
• Added feature gate for controlled Claude plugin agent rollout.

Workflows

• Added QuickJS sandboxed workflow engine for deterministic automation.
• Added workflow discovery from .opencode/workflows/ and ~/.config/opencode/workflows/.
• Added sample workflows: implement.js, review-and-fix.js, test-debug-loop.js, test.js.
• Added /workflow TUI command with create/start/list/stop flows.
• Added workflow status UI with running/waiting/agent-started states.
• Added crash-recovery journal support, concurrency limits, depth limits, and graceful shutdown.

Research

• Added research tool and /research command.
• Added command/docs support for research-oriented workflows.

Long-session checkpoints

• Added checkpoint system for long session context rebuild.
• Checkpoint writer summarizes session state into structured files before compaction.
• Added checkpoint boundary tracking in database.
• Added checkpoint writer agent prompt.
• Added rebuild context injection after compaction.
• Improved checkpoint reliability by writing files programmatically instead of relying on agent file tools.

Dream/distill self-improvement

• Added dream/distill prompt system.
• Added configurable dream/distill context limits.
• Added automatic dream/distill trigger support.
• Improved context extraction to include all text messages within configured limits.
• Raised distill quality bar to skip trivial patterns.

Provider capabilities

• Added OpenAI response chaining through previousResponseId.
• Added OpenAI Responses API options: promptCacheRetention, truncation.
• Added generic provider support through reasoning_options.
• Added schema compatibility support for stricter providers.
• Improved reasoning/tool/schema behavior across future model registry updates.

MCP and LSP

• Added MCP client capabilities declaration, roots support, local server cwd, and structured tool output handling.
• Added MCP tool filtering documentation and examples.
• Added multi-root LSP workspace support for cross-project diagnostics.
• LSP clients can now add workspace folders dynamically when agents touch external project files.
• Diagnostics are filtered to agent-opened files to avoid flooding.
• Improved TypeScript, Deno, Go, Rust, Python, and C/C++ language server handling.

TUI

• Added transcript search.
• Added notification queue.
• Added OSC-8 terminal hyperlinks.
• Added workflow status component.
• Fixed external editor chord handling.
• Improved no-argument slash commands.

Improvements

• Enabled more aggressive default performance/isolation behavior, including proactive pruning, context collapse, higher tool result budget, and Exa default behavior where configured.
• Reduced post-compaction token overhead by avoiding extra LLM turns, removing duplicate global memory injection, compressing todowrite description, and consolidating agent instructions.
• Improved TUI spinner pacing, toast countdown behavior, and autocomplete submit behavior.
• Removed old /goto command in favor of newer search/navigation behavior.
• Updated AI SDK versions.
• Updated OpenTUI to 0.4.1.
• Added design docs for workflows, response chaining, worktree isolation, provider support, LSP, and diagnostics.

Fixes

• Fixed provider request timeout poisoning where one header timeout could abort later requests.
• Fixed provider cache/scope behavior so native-only parameters are not sent to proxy endpoints.
• Fixed Anthropic cache scope gating and OpenAI store behavior for proxies.
• Fixed OpenAI MCP tool schema sanitization, Gemini type array handling, and unsupported schema keyword stripping.
• Fixed parallel tool call config handling.
• Fixed tool anchor memory growth with FIFO cap.
• Fixed MCP/LSP sidebar status refresh races and MCP initialization race.
• Fixed LSP duplicate spawning for external roots.
• Fixed MCP OAuth error escaping, idle server cleanup, and progress timeout resets.
• Fixed hidden session content-filter errors and double file injection.
• Fixed background subagents briefly appearing cancelled before runner start.
• Fixed workflow sandbox stability, concurrent workflow handling, and shutdown behavior.
• Fixed checkpoint global memory reading.
• Fixed Linux mock.module leakage, dream/distill tests, and memory-test hangs.
• Fixed published package startup by stripping type: module from packed package metadata.
• Fixed missing QuickJS dependencies in package metadata.
• Fixed CI branch filters and timeouts.

Developer / SDK

• Added session/message plumbing for checkpointing and response chaining.
• Regenerated SDK types for new message/error shapes.
• Added MCP client mock updates for roots support.
• Added config fields for workflows, dream/distill, provider reasoning options, and runtime feature controls.
• Added migrations for checkpoint boundary tracking.
• Split orchestration executor into foreground, background, push-to-background, and middleware layers for isolation/concurrency/guards.
• Added wire diagnostics and tool-call JSON repair support.
• Added package/prepack fixes for published binary compatibility.
• Updated dependency lockfile and SDK/provider packages.

Install

npm install -g @sdeonvacation/opencode-x@1.0.8

Run:

opencode-x

GitHub release includes 12 platform binaries.

v1.0.7

Highlights

🚀 Swarm Mode — Batch-Parallel Subagent Execution

• Launch multiple subagents in parallel with a single template + item list
• Same operation across N files/modules/tests runs concurrently (up to 20)
• Supports background: true for fire-and-forget batch work
• Results returned as structured XML with per-item status and timing

🌳 Worktree Isolation v2

• New /worktree slash command for subagent git isolation
• Each subagent operates in its own git worktree branch
• Prevents conflicts between parallel agents modifying the same files
• Global LSP instance shared across all agent types for performance

Other Changes

• Restored task_status tool for polling background subagent state
• Background subagents enabled by default
• Renamed binary from opencode to opencode-x
• Fixed model variant preservation on push-to-background auto-resume
• Fixed permission pattern matching in non-VCS projects
• Prevented orphaned Java LSP processes on unclean exit
• Cherry-picked upstream fixes: WebSocket transport for OpenAI, prompt corruption fix, auto-compaction respect, reasoning summary gating, shell cancel race fix, destructive edit match prevention
• CI improvements and test fixes

v1.0.3

What's Changed

Features

• Stream idle timeout: Default 60s SSE inactivity timeout with auto-retry — prevents subagents from hanging indefinitely when provider streams go silent. Configurable via stream_idle_timeout in opencode.json.
• Claude Code hook compat: Claude Code payload spec + tool input rewrite support
• Skill frontmatter: Claude-code frontmatter compatibility for skills

Fixes

• Auto-upgrade: Download binary directly from GitHub releases (no external install script dependency)
• Parallel tool calls: Unblock parallel tool call emission for task tools
• Background tasks: Track per-session, show completion toast immediately, allow re-pushing to background after auto-resume
• Stream reliability: Retry transient Anthropic stream errors; cleanup orphan tool parts on retry; mark prior orphan pending parts on new tool-input
• Provider: Adaptive reasoning for opus 4.7+ via version regex; bump anthropic sdk to 3.0.71 for opus-4.7 thinking
• Session: Include task_id in background-detach tool result; drop empty-text reasoning parts with signature
• Hook: Case-insensitive tool matcher
• LSP/Plugin: Tolerate partial config + restore test suite
• TUI: Add 'hook' to SlashCommand source type
• Tool docs: Cap edit/write content to avoid stream truncation

Infrastructure

• Raw binary release assets (no zip/tar wrapping)
• CI: GitHub-hosted runners for test + typecheck

Full Changelog: v1.0.2...v1.0.3

OpenCode X v1.0.2

Bug fixes and improvements

OpenCode X v1.0.1

Bug fixes and improvements

OpenCode X v1.0.0

Initial release