Opencode — Three Months from Bun Segfaults to Background Subagents

https://github.com/anomalyco/opencode/commits/dev/

Why This Matters

Three months of Opencode's dev branch — February 15 through May 17, 2026 — covers ~120 tagged releases spanning v1.2.3 → v1.15.4. The project is moving fast enough that a meaningful version of "what's new" requires looking at the whole stretch rather than any single release. Major themes across the window:
  1. Platform reach expanded — Node.js as a first-class runtime alongside Bun, GitLab Agent Platform support, full HTTP proxy, OTLP observability.
  2. Frontier models tracked aggressively — Opus 4.7 defaults, GPT-5.5 OAuth fixes, Mistral Medium 3.5 with reasoning, Azure gpt-5.5 routing.
  3. Session model matured — git-backed review, multistep auth, worktree/orphan-branch recovery, instruction precedence, forked-session compaction integrity.
  4. Background subagents land experimentally — the late-window feature that signals where the agent UX is going.
The pace is the story. Opencode is shipping the kind of week-over-week incremental refinement that's hard to track from outside but adds up to a substantially different product over 90 days.

The Arc, In Five Phases

Phase 1 — Mid-Feb to Mid-March (v1.2.3 → v1.2.27): Stabilization

The window opens with platform stability work — Windows segfault fixes against Bun 1.3.10, an fsmonitor daemon leak that was accumulating 60GB+ of committed memory after test runs, TUI/server config split, and the start of a permission-system refactor toward Effect. Quiet but load-bearing — the foundation work that everything later in the window depends on.

Phase 2 — March 22 (v1.3.0): The First Big Feature Drop

The major release of the window's first half. Five things land at once:
  • GitLab Agent Platform support with automatic discovery of workflow models from your GitLab instance. Workflow models can drive opencode's local tools (file read/write, shell, etc.) over WebSocket. The first time the project committed to a non-GitHub platform integration at this depth.
  • Git-backed Session Review — review uncommitted changes and branch diffs directly within opencode. Git changes become the default review source, with the desktop review tree synchronized to whatever source you select.
  • Multistep Authentication for TUI and Desktop, which is what makes GitHub Copilot Enterprise actually work.
  • Interactive Update Flow — major/minor updates show a confirmation dialog before upgrading, with the option to skip specific versions.
  • Node.js support alongside Bun, with a dedicated entry point and a build script that bundles the server with database migrations.
Also notable: the Anthropic OAuth plugin was removed in this release — a signal about how the project is thinking about auth boundaries.

Phase 3 — April 8 (v1.4.0): SDK Breaking Changes + Subagent Polish

A smaller-but-sharper release. Breaking SDK changes to fix performance issues:
  • Diff metadata no longer carries to/from full file contents — only a patch field with the unified diff. Applies to edit/patch tool metadata and to all snapshot diffs in user messages and session diffs.
  • UserMessage.variant moved under model (i.e., userMessage.model.variant).
Plus OTLP observability export, full HTTP proxy support, PDF drag-and-drop for attachments, opencode run --dangerously-skip-permissions for auto-approving non-denied prompts, and subagent session improvements (clearer titles, navigation, progress states).

Phase 4 — Mid-April to Early-May (v1.14.17 → v1.14.40): Model + LSP Maturation

The version-numbering jump from v1.4.x to v1.14.x catches you out, but the substance is consistent — a stretch of refinement focused on model integrations and LSP behavior:
  • Opus 4.7 defaults to display: summarized for Anthropic Bedrock requests.
  • GPT-5.5 with OpenAI OAuth gets correct context limits, avoiding compaction issues.
  • Mistral Medium 3.5 with reasoning added by community contributor @rubdos.
  • Roslyn LSP added for Razor / .cshtml / C# script files.
  • LSP permission prompts include request details (operation, file, cursor position).
  • Instruction precedence made more predictable — global before project before skill.
  • Forked-session compaction integrity — compacted history stays intact across forks (@spark4862).
  • Sessions filter to current path by default, with a setting to show the whole project.
  • .well-known/opencode configs supported, pointing to a separate remote config file.
The community contributor density across this phase is striking — many of the most useful fixes in the window came from outside the core team.

Phase 5 — May 14–17 (v1.14.50 → v1.15.4): Background Subagents + Effect Event System

The window closes with two structural changes:
  • Experimental background subagents — tasks can keep running while you continue working. This is the feature most worth tracking forward. Background-by-default is the direction the entire agent-UX space is moving in, and Opencode is now there as an experimental flag.
  • Effect-based core event system for more complete event delivery across sessions and integrations. Pairs with the earlier Effect-ification of the permission and question subsystems — the codebase is converging on Effect as its concurrency substrate.
Plus required billing-origin headers for NVIDIA endpoints, Azure gpt-5.5 request fixes, Linux/Windows auto-hidden menu bar in Desktop, restored markdown rendering by default in TUI, and a project-scoped bus event fix so file watchers reach the right instance.

What The Shape Tells You

Three observations across the 90-day window:
  1. Effect is becoming the architectural commitment. PermissionService, QuestionService, the core event system — successive subsystems are getting Effectified. That's the kind of decision that's invisible from outside but constrains everything downstream.
  2. Platform reach is the explicit strategy. Node.js runtime, GitLab platform integration, full HTTP proxy, OTLP observability, multistep auth for enterprise providers. Each of these unlocks a specific class of user (enterprise, on-prem, GitLab-shop) that the previous version couldn't serve. The product is being made deployable in environments that pure-Bun-pure-GitHub couldn't reach.
  3. Background subagents is the leading edge. Most of the window is incremental — bug fixes, model additions, refinements. The experimental background-subagent flag in v1.14.51 is the only thing in the window that hints at a meaningfully different interaction model. Worth watching whether it graduates out of experimental before the next major version.

Our Take

Three months of Opencode is more product change than most projects ship in a year. The thing it doesn't have — yet — is a single anchor feature you can point at as "the reason to switch." The arc instead is the cumulative effect of consistent weekly shipping against a clear roadmap: more platforms, more models, more reliable sessions, more first-class enterprise capabilities, and now the beginnings of background work.
If you evaluated Opencode in February and bounced off, the version you'd be looking at today is meaningfully different — particularly if your blockers were Node.js support, GitLab integration, multistep enterprise auth, or fork/worktree session stability. Worth a fresh look.