intent-runtime

    Compatibility and support

    This documentation describes Intent Runtime V2. See Installation for setup. Result and baseline schemaVersion is 2.0-draft.6, a format identifier independent of the npm package version.

    Runtime

    ItemSupported scope
    Node.js≥20.0.0, ESM
    PlatformsWindows, macOS, Linux; cross-platform regression on Node 20.0.0/24.x
    schema-dslIntegrated at 3.0.4; see the rule matrix
    CommonJS / browsersNot promised as direct runtime entry points for this release

    Integration boundaries

    EntryStatus and scope
    Native, stdio, SDK-independent adapterDirect application calls, standalone stdio, embedded MCP; shared recognition core
    MCP protocolsSDK verification for 2025-06-18, 2025-11-25, 2026-07-28
    Current Codex CLI modelHost mode reuses the model; an actual Linux/CLI 0.162.1/gpt-6.1-sol integration record negotiated 2025-06-18
    Codex DesktopConfiguration provided; actual client not independently verified
    Model APIsOpenAI/xAI Responses and explicit Chat Completions; automated protocol/reply/error tests; actual provider configuration needs its own verification
    JavaScript / TypeScriptESM applications supported; CLI loads .js/.mjs, with TypeScript configuration compiled first
    Streamable HTTPStandalone CLI address/port/path/public-origin; stateful 2025 protocols have SDK/session-isolation checks and public HTTPS Host/Origin rejection regression; embedded adapter also available; arbitrary remote auth/proxy deployments are not thereby verified
    Other clientsCore is composable; no dedicated setup adapter or client compatibility guarantee

    Separate supported targets from tested versions. Recheck your real model, client, SDK, and deployment after changes. Structural checks do not guarantee correct understanding of every natural-language input. A particular corpus pass rate is not a product-wide accuracy claim.

    Dated verification records

    See the current CI workflow. These dated records do not prove every later commit or your model configuration.

    DateEnvironment/methodEvidence and boundary
    2026-10-1571c774d; Linux/Windows/macOS × Node 20.0.0/24.x; quality/docs, 8 jobsExact CI passed; historical engineering/package regression, not future commits or model semantics
    2026-10-15Actual MCP SDK 2.3.1, three declared protocolsExample checker uses controlled candidates and real core/Validator; protocol/field-chain evidence, not actual model acceptance
    2026-10-14Linux, Codex CLI 0.162.1, gpt-6.1-sol, high reasoningCurrent-model prepare/accept and semantic samples ran; this specific configuration does not guarantee all model inputs

    Judge recognition under your actual model configuration and include model/client versions in reports. Controlled executors verify structure and call flow, not model accuracy.

    Complete scenarios use fixed candidates processed by the real core and real Validator where business fields exist. They cover requirements, hierarchy, multiple turns, facts, ambiguity, and field states. They are contract examples, not provider accuracy reports. No-Schema examples do not add artificial business validation.

    Updates and runtime boundaries

    Update with npm install @devcodex/intent-runtime. See the changelog and Codex maintenance.

    Installing the package does not automatically process every client message. Default timeoutMs=0 adds no overall recognition deadline; connection closure, service restarts, and client tool deadlines can still affect calls. Short mechanism checks do not guarantee continuous connections lasting hours or days.

    Get help

    For Codex, run doctor using the maintenance guide. For recognition failures inspect code/stage through Errors. Include package/Node/client versions, OS, reproducible input, and error summaries when opening GitHub Issues. Review material for credentials or sensitive content before publishing it.