LoginSign Up
← Back to all sparks
O

opencode

AI-ASSISTANTS
Velocity5.0

Open-source AI coding agent built for the terminal

opencode.ai ↗

Provider compatibility is where opencode spends its releases now, not features.

coding-agentprovider-compatibilitysession-compactionlocalizationdesktop-app
◆Current state
opencode ships a patch release every day or two, and the work splits cleanly in two: core changes that keep an expanding roster of model providers behaving correctly, and desktop polish covering localisation, right-to-left layout and session handling. The recent releases fix Kimi system prompt selection for Moonshot, reasoning-effort handling for xAI, sampling defaults for DeepSeek V4 Flash, and Meta prompt routing for Muse models. Session compaction was reworked to keep recent turns whole and produce summaries that smaller models can actually use.
◆Where it's heading
The centre of gravity has moved from building the agent to making it survive contact with a dozen incompatible provider APIs. Each release absorbs another provider's quirks — reasoning field names, PDF vision support, device-code login, retry semantics — which is the cost of positioning as provider-neutral. The parallel investment in locale coverage and right-to-left support points at a deliberate push beyond English-speaking users, with community contributors carrying much of it.
◆Prediction
Expect the patch cadence to hold, with more provider-specific compatibility fixes as new models land and further desktop localisation. A minor-version bump would likely be needed for anything beyond this maintenance pattern, and nothing in these entries signals one.

◆Recent moves

  1. 6d ago

    Fix Kimi prompt selection and xAI xhigh reasoning effort

    Two narrow provider fixes: Kimi system prompt selection for official Moonshot and Kimi endpoints, and xhigh reasoning effort for xAI models. Exactly the per-provider correctness work that dominates this feed.

    View source ↗
  2. 6d ago

    Compaction keeps recent turns whole; retries get capped with jitter

    The broadest release in this window: session compaction now preserves complete recent turns and writes summaries smaller models can follow, automatic retries are capped with jitter to stop retry storms, and four providers get corrected routing or sampling defaults. Reliability work rather than new capability, but it touches how every long session behaves.

    View source ↗
  3. 9d ago

    Config parser ignores unknown fields; macOS app survives window close

    Config parsing stops failing on unknown top-level fields, projects opened from Home register properly, and the macOS app keeps running after its last window closes. Small robustness fixes in the same maintenance rhythm.

    View source ↗
  4. 12d ago

    Message chronology fixes, session JSON export, wider locale coverage

    Message ordering moves from ID-based to real chronology, which fixes revert and fork behaviour on imported sessions, and repeated compaction stops dropping orphaned tool results. Session transcript export as JSON lands alongside broader locale coverage.

    View source ↗
  5. 13d ago

    xAI device-code login and retryable provider errors for headless runs

    xAI login collapses to a single device-code flow that works headless, and provider errors mid-stream are preserved so compatible providers can retry rather than fail outright. Both changes target remote and automated environments specifically.

    View source ↗
  6. 15d ago

    Early right-to-left layout support and locale-aware plurals on desktop

    Early right-to-left layout support arrives across tabs, drawers and session views, with markdown parsing moved off the main thread to keep rendering responsive. The internationalisation push that later releases continue starts here.

    View source ↗