Opencode Just Became the Open-Source Coding Agent You Can Actually Trust

Originally published at terminalblog.com.

Opencode v1.18.18 dropped two days ago (August 13, 2026) and it's the kind of release that turns a "cool project" into a daily driver. The 160,000-star open-source coding agent spent the last month quietly fixing the paper cuts that made you keep a commercial backup: MCP servers getting stuck in reconnect loops, provider model configs breaking mid-session, desktop app freezes, and session ordering chaos.

If you tried Opencode a few months ago and walked away frustrated — this is the version to come back for.

Since then, the release train kept rolling: twelve more versions (v1.18.19 → v1.18.30) landed between August 20 and September 9, and the theme is consistent — fewer freezes, better retries, and model support that keeps pace with the frontier. The two fixes beginners will feel most are the new five-minute timeouts (no more infinite spinner on a slow model startup) and GPT-6 Astra finally showing up for OpenAI subscription users.

Here's the wave, sorted by what actually matters for daily use.

GPT-6 Astra Arrives for ChatGPT Subscribers

The single most useful change in the wave is a fix most release notes skip over. In v1.18.29, Codex OAuth model filtering learned to recognize integer GPT versions like gpt-6 — which means gpt-6-astra no longer vanishes from the model picker for people signed in with a ChatGPT subscription.

If you've been reading about GPT-6 Astra (Codex's new default, Goose v1.50) and wondering why it never appeared in Opencode, this was the bug. v1.18.30 follows up with the Astra system prompt for GPT-6 models, so the agent actually tunes its behavior for the model family instead of treating it like a generic GPT.

For anyone running Opencode against their OpenAI subscription, this turns "I can't use my newest model" into "my newest model is my default."

The Hangs Are Finally Capped

Long-time Opencode users know the worst failure mode: the agent sends a request, the provider takes its time warming up a big model, and the terminal sits there spinning — with no error and no escape.

v1.18.27 fixes that class of problem with sensible defaults:

Provider header timeouts default to five minutes, so slow model startups fail with a real error instead of hanging forever
Streamed chunk timeouts also default to five minutes, with false supported to disable them if you genuinely need unlimited waits
Hitting a timeout now avoids unhandled errors when canceling half-finished SSE reads

v1.18.20 attacks the other half of the equation: when a provider responds with finish_reason: network_error (or variants like network-error), Opencode now retries the response instead of silently ending the turn. xAI capacity and temporary unavailability stream errors get the same retry treatment. The practical result: a flaky network or an overloaded provider no longer kills your session mid-task.

Cloudflare AI Gateway: Now a First-Class Route

If you route models through Cloudflare's AI Gateway, two releases matter:

v1.18.19 added native OpenAI and Anthropic passthroughs for Cloudflare AI Gateway models
v1.18.23 fixed routing for third-party providers (non-Workers models now work through the gateway's REST API) and converts dotted Anthropic model IDs like claude-haiku-4.5 to the dashed slug Anthropic expects

Taken together: a Cloudflare-gated setup that used to break on Anthropic model names or silently drop non-Workers providers now behaves like a direct connection.

Subagent Failures Become Resumable

For teams running multi-agent workflows, v1.18.20 closes a silent-failure gap: failed subagent tool calls now surface with a resumable task_id instead of returning an empty result. Permission requests triggered by subagents during opencode run are also answered properly now — previously they could stall a headless run.

ChatGPT Subscription Parity Keeps Improving

Alongside the Astra fix, v1.18.19 matched Codex rate limits more closely to ChatGPT subscription limits (so heavy OpenAI usage behaves the same whether you're in Codex or Opencode), and v1.18.28 sends the session ID as GitHub Copilot's interaction header for better request tracking across a session. v1.18.26 makes Claude 5 sessions tolerate stale thinking blocks instead of failing after prompt or tool changes, and lets Bedrock GPT-5.6 models accept none reasoning effort.

The Rest of the Wave (Quick Hits)

v1.18.30 provider hygiene: Bedrock DeepSeek model IDs (including ARN-based IDs) preserved correctly; Azure and OpenAI provider SDKs updated with compatibility fixes; reasoning effort variants for supported GitLab GPT and Claude models
Azure CLI sign-in now asks for the resource name directly instead of querying Azure management APIs (v1.18.26); Entra ID sign-in through the Azure CLI also landed in v1.18.24
Desktop: session renames save reliably from the title editor and tab context menu (v1.18.26); device auth uses the desktop client ID; archived sessions disappear from the Home list immediately (v1.18.24)
/connect now correctly shows which providers are authenticated (v1.18.19)
Malformed model pricing no longer breaks usage cost calculation; OpenAI websocket requests exceeding message-size limits fall back gracefully (v1.18.19)

The Headline: MCP Finally Works Reliably

The biggest win in v1.18.18 isn't a flashy feature. It's that MCP SSE connections no longer get stuck in reconnect loops after server error responses.

If you've used MCP with any agent, you know the pain: the server hiccups, the client spins forever trying to reconnect, and your workflow dies. Opencode now:

Reconnects MCP servers after expired SDK sessions (including concurrent requests)
Honors configured MCP OAuth callback ports in mcp debug
Restored compatibility with legacy MCP SDK clients
Improved compatibility with newer MCP servers and OAuth flows

This alone makes Opencode viable for teams running custom MCP tooling — the kind of reliability that used to be a commercial-agent moat.

Provider Model Routing: No More Surprise Failures

OpenCode now respects provider-defined reasoning options instead of falling back to the wrong reasoning controls. This fixes a whole class of "why is my model acting dumb?" issues:

GitHub Copilot model routing now honors each model's advertised chat or responses endpoint
Azure Cognitive Services endpoint support restored for Azure-hosted models
OpenAI Responses phase handling fixed so it doesn't break conversations
Cerebras reasoning replay fixed — earlier assistant reasoning now sent back in provider-supported field
Mistral prompt caching stabilized with correct prompt cache keys per SDK
MiniMax M3 thinking variant selection fixed
xAI prompt cache routing improved with PDF file support in Responses models
Meta model handling improved for reasoning variants and provider requests

The practical impact: you pick a model, it works the way the provider intended. No more debugging why reasoning effort seems ignored or why a model that should support PDFs keeps erroring.

Desktop App: The Freezes Are Gone

The desktop app got a massive stability pass. If you've ever had the app hang on a large pasted image, freeze when closing a project, or show blank session titles — those are fixed:

Composer lag reduced when drafts include large pasted images or attachments
Right-to-left layout issues fixed across tabs, drawers, resizing, titlebar interactions
Session tabs no longer clip when resized smaller
Stale session tab state in title bar fixed
File tree tab clipping prevented when resized
Debug gutter labels and values aligned
Home page scrolling fixed — sticky headers and session list behave correctly
Startup readiness fixed — WSL server loading included before app reports ready
Custom agent selector visibility fixed when custom agents are available
Home page command palette search can now find and open sessions
Tab keyboard navigation follows visible tab order
External links open in system browser instead of inside the app
Markdown parsing moved off main thread — app stays responsive while rendering messages
Linux packages now install AppStream metadata

The V2 desktop sidecar is now opt-in, backed by the bundled CLI service, with collapsible model provider sections in V2 settings.

Session Management: Chaos → Order

Session handling got a fundamental overhaul:

Chronological message ordering now stays correct even when imported or legacy message IDs are out of order
Revert and fork actions use real message chronology instead of message ID ordering
Truncation cleanup removes stale files by file timestamp more reliably
Repeated compaction keeps earlier tool-call history in summaries instead of dropping orphaned results
Blob-based attachments now load correctly in the web UI
Session timelines respect chronological message boundaries more reliably
Stored messages stay ordered by creation time
Session lists sort by persisted activity time more reliably
Sessions without titles fall back to generated names instead of appearing blank
Project search matches any known recent project instead of only the first five

This means your history is actually trustworthy — forks, reverts, and imports behave the way you'd expect.

TUI & Core: The Daily Quality-of-Life Fixes

GitHub pull request reviews now include the PR number and URL in context
Up Arrow shortcut closes the subagent picker when first item is selected
Automatic session retries capped with jitter to reduce retry storms
PDF attachments enabled for GitHub Copilot models advertising PDF vision support
DeepSeek V4 Flash sampling defaults applied correctly on supported providers
Muse family models routed to correct Meta system prompt
xAI login simplified to single device-code flow (works better in headless/remote)
Structured mid-stream provider errors preserved so compatible providers can retry failed responses
More transient provider/network errors retried instead of failing immediately
Cache writes counted in ACP usage totals
Remote workspace 5xx response bodies logged in host logs for proxy debugging
Host directory path no longer sent to remote workspaces — prompts resolve from remote project root
Queued ACP session updates waited for before ending a turn
Unknown top-level config fields ignored instead of failing config parsing
Projects opened from Home registered so available to rest of app
Right-click in Home opens project menu
Fallback to local directories when project picker server doesn't support search
macOS app stays running after last window closes, reopens window on activation
Chinese terminology updated to widely recognized translations
Server's current default model used instead of stale local config

The Upstream Sync: Opencode v1.17.13 → v1.18.13

OpenCode (the SST/Anomaly version) adopts upstream improvements from the opencode-ai fork, including:

Code mode MCP adapter for running confined orchestration scripts against connected MCP tools
execute tool hidden unless code mode is enabled
Model-specific system prompt for Meta Muse Spark
Azure AI support for GPT-5.6
Paginated MCP tool catalogs no longer lose tool metadata or output schema validation
Low reasoning effort preserved for OpenRouter small-model variants
GitHub Copilot model routing honors each model's advertised endpoint
Session lists match equivalent instance directories reliably
Cerebras reasoning replay fixed
Z.ai context-window overflow errors better classified
Unavailable config directories handled gracefully
Reasoning effort variants exposed for Grok models
xAI prompt cache routing improved
Meta model handling improved
Zero billing batch size crashes prevented for GitHub Copilot
OpenAI pro reasoning mode supported
Response storage disabled by default for xAI Responses
OAuth support added for Luna Responses Lite
Codex context limits used for GPT-5.6 over OAuth
Obsolete Codex workaround removed that could interfere with OpenAI Luna Responses Lite
TUI spinner registration fixed — loading indicators keep rendering
CLI environment variables forwarded to TUI worker

Subagent Permission Fix (Big for Teams)

Subagent permission errors that referenced phantom deny rules are fixed. A read-only or delegating agent's readOnlyBash allowlist is no longer projected onto a writable subagent as a bash ceiling — so a delegated subagent can run its own allowed commands (e.g., git status). Edit, notebook, and MCP denials are still inherited as hard ceilings.

This matters if you use delegated agents for code review, testing, or documentation — they no longer get blocked by parent-agent restrictions that shouldn't apply to them.

Clickable File References in Agent Responses

Code spans in agent responses that match real files in your workspace become clickable links that open the file at the referenced line. Non-existent paths stay as plain code. Includes fallback workspace search and "File not found" warning when clicking dead links.

This is a small UX thing that compounds: when the agent says "look at src/auth/login.ts:42", you click it. No more copy-paste-navigate.

Model Reasoning Variants: See What You're Getting

The model's default reasoning variant now shows and selects in chat and Agent Manager. No more guessing which reasoning tier you're actually on.

AWS & GCP Auth: Structured Credentials Supported

Structured AWS access keys and Google Cloud service-account JSON now supported when connecting Bedrock and Vertex AI in VS Code. No more manual env var juggling.

Agent Manager PR Actions

Resolve/unresolve review threads, jump to comments section, and scroll-to-top for PR diff view — the Agent Manager is becoming a genuine code review companion, not just a code writer.

WarpGrep Removed (Good Riddance)

The built-in experimental Morph WarpGrep codebase search tool and its retired configuration flag are gone. The upstream opencode search is better anyway.

Workspace Restoration Preserved

Workspace restoration outcomes preserved when reverting fresh VS Code sessions — your layout, open files, and state survive the revert.

The September wave (v1.18.19–v1.18.30): trust, timeouts, and GPT‑6 Astra

Since v1.18.18, Opencode shipped 11 more releases in four weeks (through v1.18.30, September 9). No wall of features — mostly the same story told louder: connections that don't die, models that show up, and providers that work the way they advertise. The version to install today is v1.18.30.

GPT‑6 Astra, finally visible

v1.18.29 fixed the single most frustrating bug of the month: gpt-6-astra was not showing up for OpenAI subscription users. If you had a ChatGPT plan and kept seeing an older model in the picker, that was it. A related fix lets Codex OAuth model filtering recognize integer versions like gpt-6 — so subscription users get the same model list as API-key users.
v1.18.30 adds the Astra system prompt for GPT‑6 models — the same class of harness-level prompt tuning that makes a new model behave like a coding agent instead of a chat model.
v1.18.30 also adds reasoning-effort variants for supported GitLab GPT and Claude models, so the "what reasoning tier am I actually on?" transparency extends to the GitLab-hosted crowd too.

Timeouts that stop lying to you (v1.18.27)

The ugliest failure mode in terminal agents is a stall that never tells you it stalled. v1.18.27 fixed exactly that:

Provider header requests now default to a five-minute timeout so slow model startups fail loudly instead of hanging forever.
Streamed chunk reads got the same five-minute default, with false supported to disable it for providers that legitimately stream slowly.
SSE reads that time out are now cancelled cleanly — no more unhandled errors when you Ctrl+C out of a wedged request.

And in v1.18.26, Claude 5 sessions now tolerate stale thinking blocks instead of failing after you edit the prompt or swap tools — a fix for anyone who hit "conversation broke after I changed the request."

Anthropic thinking-block control (v1.18.26/27)

thinking.blockBinding — how tightly the harness binds its own thinking-block behavior to Anthropic's provider default — can now be opted out via config (PR #46820) when you need the provider default preserved. The binding is also limited to Claude 5.1+ models so older deployments stop getting requests rejected.

Azure without Bun, and Entra ID sign-in (v1.18.24–25)

v1.18.25 fixed Azure authentication so az CLI sign-in works without requiring Bun — a real paper cut for Windows and minimal-Node setups.
v1.18.24 lets Azure providers sign in with Microsoft Entra ID through the Azure CLI instead of forcing an API key, and stops Bedrock reasoning responses from getting cached into unreplayable empty messages (the kind of bug that silently ate your follow-up turns).

Cloudflare AI Gateway finally routes (v1.18.23)

Non-Workers models now work through the Cloudflare AI Gateway's REST API (PR #44281).
Anthropic models through the gateway convert dotted IDs like claude-haiku-4.5 to the dashed slug Anthropic expects — previously these just failed.

Little things that compound

v1.18.28 sends your session ID as GitHub Copilot's interaction header — request tracking across a session actually works now, which matters if you use Copilot as a provider and want consistent usage/rate-limit behavior.
v1.18.22 fixed OpenCode device-login links when servers return relative verification URLs or use a base path, and stopped sending textVerbosity to OpenAI-compatible providers that don't support it.
v1.18.26 lets Azure CLI sign-in ask for the resource name directly instead of querying Azure management APIs (fewer moving parts, fewer permission errors).
v1.18.21 keeps responses going when a model reports an unknown finish reason instead of stopping early, and routes Vertex AI eu/us multi-region Gemini requests through the right endpoints.

If you're on v1.18.18 or anything older: update. The gap from 1.18.18 → 1.18.32 is almost entirely "things no longer fail silently."

September 14–21 Patch Wave (v1.18.31–32)

Two more patch releases dropped after v1.18.30:

v1.18.31 (Sep 14)

ACP session state restored — model, effort, mode, and reasoning chunk boundaries now load correctly when resuming or forking sessions
Remote config auth errors now show during startup instead of silently failing
GitHub Copilot adaptive thinking — requests summarized adaptive thinking for better reasoning visibility

v1.18.32 (Sep 21)

Bedrock image attachments fixed — now only hoisted for Claude, Nova, and Llama 4 models (prevents incorrect attachment handling)
Together AI streaming usage reporting corrected

How to Update

# If installed via install script
curl -fsSL https://opencode.ai/install | bash

# Or via npm
npm update -g @opencode/opencode

# Verify
opencode --version
# Should show 1.18.32 (latest in the wave covered here)

The Bottom Line

v1.18.18 is the first Opencode release that feels like a polished product instead of a promising project.

The MCP reliability alone makes it viable for production workflows. The provider routing fixes mean you stop fighting the model. The desktop stability means you can keep the app open all day. The session ordering fixes mean your history is actually useful.

If you're on the fence about open-source vs. commercial agents — this release narrows the gap significantly. Opencode has always had the better architecture (skills, slash commands, explicit lifecycle). Now it has the reliability to match.

- Claude Code: Skills vs Subagents vs MCP — The 2026 Decision Guide

Published August 15, 2026; updated September 22, 2026 — covers Opencode v1.18.18 through v1.18.32 (August 13 – September 21, 2026). Full changelog: opencode.ai/changelog

FAQ

Q1: What is Opencode?
Opencode is a 160K-star open-source terminal coding agent from SST/Anomaly. It uses a skill-driven architecture (explicit lifecycle, slash commands, skills) and supports multiple model providers with provider-neutral routing.

Q2: What's the biggest fix in v1.18.18?
MCP reliability — SSE connections no longer get stuck in reconnect loops after server errors. This makes Opencode viable for teams running custom MCP tooling.

Q3: Does the desktop app work reliably now?
Yes. v1.18.18 fixes composer lag with large images, session tab clipping, startup readiness (WSL loading), markdown parsing off main thread, and more. The V2 desktop sidecar is opt-in and backed by the bundled CLI service.

Q4: What changed with session management?
Chronological message ordering now stays correct even with imported/legacy messages. Revert and fork actions use real chronology. Session lists sort by persisted activity time. Blank session titles now get generated names.

Q5: What are the provider routing fixes?
OpenCode now respects provider-defined reasoning options: GitHub Copilot model routing honors advertised endpoints, Azure Cognitive Services restored, OpenAI Responses phase handling fixed, Cerebras reasoning replay fixed, and more. Models work the way providers intended.

Q6: How do I update to v1.18.30?
curl -fsSL https://opencode.ai/install | bash or npm update -g @opencode/opencode. Verify with opencode --version.

Q7: Is Opencode ready for daily work?
Yes. v1.18.18 is the first release that feels like a polished product. MCP reliability, provider routing, desktop stability, and session ordering all work — narrowing the gap with commercial agents significantly.