kandev/docs/decisions/0048-plugin-host-utility-agent-invoke.md

0048 — Plugins invoke a plugin-selected utility agent

• Status: accepted
• Date: 2026-07-21
• Area: backend, frontend, protocol
• Related: 0043 — Plugin host data API, 0002 — Host utility agentctl for sessionless ACP flows (the inference tier this reuses), 0018 — Runtime settings overrides

Context

A plugin that wants to do an LLM step (summarize a conversation, classify an issue) has no sanctioned way to run a completion: it would need to ship and manage its own provider API key. Kandev already runs LLM completions on the operator's configured agents; a plugin should be able to borrow one.

Kandev also already has the right primitive. ADR 0002's host-utility tier (internal/agent/hostutility.Manager.ExecutePrompt) runs a one-shot, non-interactive, sessionless completion against a warm agentctl instance — exactly what title generation, commit messages, and the Slack assistant use. There is no need for a new agent loop, and the interactive agent runtime (internal/agent/runtime, internal/office) is the wrong tool: it is streaming, stateful, and requires an executor + workspace + worktree.

Decision

Add a capability-gated Host.InvokeUtilityAgent, backed by each plugin's configuration and the existing host-utility tier.

1. Plugin config: utility_agent. Plugins declaring agent_invoke declare a config_schema field named utility_agent, with type: string and format: utility-agent. Settings > Plugins renders it as a picker of the existing built-in and custom utility agents. The picker displays the name but stores the selected agent's stable ID, scoped to the plugin; no global user setting or agent profile is involved.

2. New capability agent_invoke. A boolean Capabilities.AgentInvoke, enforced exactly like state/secrets: Host.InvokeUtilityAgent returns gRPC PermissionDenied (capability 'agent_invoke' not declared) when the manifest doesn't declare it.

3. New RPC + SDK method. InvokeUtilityAgent(InvokeUtilityAgentRequest{prompt}) returns (InvokeUtilityAgentResponse{text}) on service Host; SDK Host.InvokeUtilityAgent(ctx, prompt string) (string, error). The request message is the forward-compatible extension point (a future system_prompt or max_tokens is an added proto field, no SDK signature change).

4. Reuse the host-utility tier (ADR 0002). The kandev-side handler: gate agent_invoke → read the calling plugin's utility_agent config → resolve its ID through the established utility-agent configuration → call a narrow utilityRunner.ExecutePrompt(ctx, agentType, model, "", prompt) and return the text. utilityRunner is a thin pluginsHostUtilityAdapter over hostutility.Manager wired in backendapp, so internal/plugins never imports the agent runtime (the same cycle-avoidance as the Slack assistant's adapter and ADR 0043's data sources). No task, session, workspace, or worktree is involved.

5. Typed "not configured" failure. If no utility agent is selected — or the selected utility agent has since been deleted or disabled — the handler returns gRPC FailedPrecondition (no utility agent configured / configured utility agent "<id>" not found), never a silent empty completion. A stale selection is treated as "unconfigured", not an internal error.

Consequences

• A plugin declares capabilities.agent_invoke: true and calls host.InvokeUtilityAgent(ctx, prompt) — no API key, no provider wiring. This is what unblocks the "My Daily Standup" plugin's summarization step.
• The operator stays in control: nothing runs until each plugin selects a configured utility agent, including its cost and availability characteristics.
• We reused the sessionless inference path instead of building an agent loop; the only net-new machinery is a plugin config picker and one gated RPC handler.
• The InvokeUtilityAgentRequest/Response proto is a public contract, extended additively.

Alternatives considered

• Let plugins bring their own API key. Rejected: every plugin re-implements key management and secret storage, and the operator loses cost control. The whole point is to delegate to kandev's already-configured agent.
• Run the completion through internal/agent/runtime / office. Rejected: both are streaming, stateful, and require an executor + workspace; neither offers a synchronous prompt→text call. ADR 0002's host-utility tier already does exactly this, sessionlessly.
• Reuse default_utility_agent_id. Rejected: that default serves Kandev's internal utility calls and cannot express a plugin-specific selection.