LoginSign Up
← Back to all sparks
O

opencode

AI-ASSISTANTS
Velocity5.0

Open-source AI coding agent built for the terminal

opencode.ai ↗

opencode tracks the frontier model pace through weekly provider-layer maintenance.

multi-providerprovider-reliabilitymodel-catalogacp-sessionsbedrockazure-auth
◆Current state
opencode is a multi-provider AI coding environment shipping weekly patch releases that keep pace with the frontier model calendar — GPT-6/Astra, Grok 4.7, DeepSeek V4.1 Flash, and Bedrock reasoning variants have all landed recently. Provider-layer reliability is the dominant work: Azure CLI auth via Entra ID, Cloudflare AI Gateway routing for third-party models, Together AI streaming, and Bedrock reasoning replay have all received fixes. Session handling for the ACP (Agent Control Protocol) layer — preserving model, effort, and reasoning boundaries across forks and resumes — is getting incremental attention as agent-mode usage grows.
◆Where it's heading
The product is broadening horizontally across providers rather than pushing new capabilities. Each release adds a new model or fixes a provider edge case; the agentic surface (ACP sessions, tool call timing, apply_patch) is being made more reliable but not fundamentally extended. Azure, Bedrock, and Cloudflare are the three enterprise provider paths getting the most active work, suggesting an enterprise-readiness push rather than a consumer feature sprint.
◆Prediction
Continued model catalog additions and ACP session stability improvements are the near-term pattern. A directional move would require something like native agent orchestration, a new tool-use surface, or a pricing change — none visible in the current entries.

◆Recent moves

  1. 17h ago

    Add Grok 4.7 and DeepSeek V4.1 Flash; fix Bedrock and Together AI

    v1.18.32 fixes Bedrock image attachment handling and Together AI streaming, while community contributions add Grok 4.7 and DeepSeek V4.1 Flash to the model catalog. The model additions extend coverage but don't change the product's direction — they're table-stakes freshness for a multi-provider tool.

    View source ↗
  2. 7d ago

    Restore ACP session options across load/resume/fork

    v1.18.31 restores ACP session state — model selection, effort, mode, and reasoning chunk boundaries — when loading, resuming, or forking sessions, a reliability fix that matters for developers running long-context agentic tasks that need to pick up where they left off. Adaptive thinking summarization for GitHub Copilot models is a smaller but real improvement to reasoning quality.

    View source ↗
  3. 13d ago

    GPT-6 Astra system prompt; reasoning variants for GitLab and Claude

    v1.18.30 adds the Astra system prompt for GPT-6 models and reasoning effort variants for GitLab GPT and Claude — extending the product's compatibility surface to recently released frontier models. The Bedrock DeepSeek ID preservation and provider SDK bumps are reliability work needed to keep existing integrations stable.

    View source ↗
  4. 17d ago

    Fix GPT-6 model filtering for OpenAI subscription accounts

    v1.18.29 patches two bugs in GPT-6 OAuth model filtering that caused GPT-6 and gpt-6-astra to not appear for OpenAI subscription users. This is maintenance follow-through on the GPT-6 support added in the prior release.

    View source ↗
  5. 18d ago

    Session tracking header for GitHub Copilot; desktop auth fix

    v1.18.28 adds a session ID header to GitHub Copilot requests for better cross-session tracking and fixes desktop authentication flow. These are integration-layer adjustments with no user-facing capability change.

    View source ↗
  6. 19d ago

    Default 5-minute provider timeouts; scope Anthropic thinking to Claude 5.1+

    v1.18.27 defaults provider header and SSE chunk timeouts to five minutes, which meaningfully reduces spurious failures when slow models take time to start streaming. The Anthropic thinking block binding fix — scoping it to Claude 5.1+ and making it configurable — prevents older deployments from breaking on an opt-in feature they don't need.

    View source ↗