Member Junction
    Preparing search index...

    Interface MJAIAgentHarnessEntity_IHarnessCapabilitySettings

    Strongly-typed shape of AIAgentHarness.CapabilitySettings (the MJ: AI Agent Harnesses entity), bound to the column via JSONType metadata so CodeGen emits a typed accessor.

    An agent harness is an external agent runtime with its own reasoning loop and tool sandbox — Claude Code (Agent SDK or headless CLI), Codex CLI, OpenCode, Pi, Cline, Gemini CLI. MJ runs one as the reasoning substrate for an MJ agent while keeping identity, permissions, governed data access, payload contracts, HITL and cost control on the MJ side. These flags declare what each harness's adapter can actually do, so the runtime knows what it must emulate rather than assume.

    That emulation is the reason these are metadata and not a code constant. Harnesses differ in ways that change token cost and audit granularity, not merely convenience: a harness without MJAIAgentHarnessEntity_IHarnessCapabilitySettings.SessionResume needs prior context replayed into a fresh invocation on every turn, which the runtime has to budget for against MaxTokensPerRun.

    All properties are optional — an omitted flag means the capability is not supported, which is the safe default: the runtime falls back to emulation or refuses the feature rather than calling into an adapter that cannot honour it.

    Holding these as JSON (rather than dedicated BIT columns) keeps the registry table simple and lets a new harness capability be added without a schema migration — just extend this interface and re-push. Mirrors the IBridgeProviderFeatures / IRemoteBrowserProviderFeatures model.

    See /plans/external-agent-harness.md.

    interface MJAIAgentHarnessEntity_IHarnessCapabilitySettings {
        McpClient?: boolean;
        MidTurnCancellation?: boolean;
        ModelSelection?: boolean;
        PermissionHooks?: boolean;
        PermissionPolicy?: boolean;
        SessionResume?: boolean;
        StructuredOutput?: boolean;
        UsageReporting?: boolean;
        WorkspaceScoping?: boolean;
    }
    Index

    Properties

    McpClient? MidTurnCancellation? ModelSelection? PermissionHooks? PermissionPolicy? SessionResume? StructuredOutput? UsageReporting? WorkspaceScoping?

    Properties

    McpClient?: boolean

    The harness can be pointed at an MCP server, enabling the read-only intra-turn loopback into MJ data (entity reads, RunView, queries) under a per-run scoped credential. When false the agent can still act — all authority-transferring operations go through the turn protocol regardless — but it cannot read MJ data mid-turn.

    MidTurnCancellation?: boolean

    The harness can be cancelled mid-turn and will stop promptly, so a cancellation token reaching the adapter actually interrupts in-sandbox work instead of only being honoured at the next turn boundary.

    ModelSelection?: boolean

    The harness accepts a model override at launch, so AIAgentHarness.DefaultModel and any per-agent override are actually honoured. When false the harness runs on whatever model its own configuration selects and MJ's model preference is advisory only.

    PermissionHooks?: boolean

    The harness exposes permission hooks the adapter can intercept, so a mutating in-sandbox operation can be paused and surfaced as an MJ: AI Agent Requests HITL prompt. This is about interactive approval mid-turn; see PermissionPolicy for static policy enforcement.

    PermissionPolicy?: boolean

    The adapter translates MJ's HarnessPermissionPolicy into flags the harness actually honours, so a configured posture and allow/deny list take effect. When false the policy is inert — the harness runs on its own defaults regardless of what the agent's metadata says.

    This is deliberately separate from PermissionHooks: a harness can enforce a static policy at launch (Claude Code's --allowedTools, Pi's --tools) while having no interactive hook to pause on. Conflating the two is what let four adapters silently ignore a strict posture while the runtime warned about the wrong thing.

    The runtime warns when a policy is configured and this is false, because an unenforced policy is worse than no policy: the operator believes something is gated.

    SessionResume?: boolean

    The harness can resume a prior session by ID, so turn N+1 continues the same reasoning context rather than starting cold. When false the adapter must emulate continuity by replaying accumulated context into a fresh invocation each turn — correct, but the token cost grows with turn count and must be budgeted against the run's guardrails.

    StructuredOutput?: boolean

    The harness has a native structured-output mode (e.g. Claude Code's --output-format) that can be constrained to emit the Loop next-step JSON envelope at turn end. Without it the runtime leans on BaseAgent's malformed-response retry machinery to coax the envelope out, which costs extra turns on a harness prone to conversational drift.

    UsageReporting?: boolean

    The harness reports token usage and (where available) cost per turn. Required for the per-turn AIPromptRun accounting that feeds MaxCostPerRun / MaxTokensPerRun; without it a run's spend is invisible to MJ's guardrails and only wall-clock and iteration limits can interrupt it.

    WorkspaceScoping?: boolean

    The harness accepts a working directory it will confine file operations to, letting the sandbox provider scope a workspace per run/agent/user. When false the provider must isolate at the process or container boundary instead.

    Settings

    Member Visibility

    On This Page

    Properties