package lifecycle import ( "context" settingsmodels "github.com/kandev/kandev/internal/agent/settings/models" ) // SkillDeployer materialises an agent profile's skills + custom prompt into // the workspace before the agent process starts. ADR 0005 Wave A introduced // this seam so the runtime owns "make the agent's skills available"; Wave E // moved the concrete implementation (manifest builder + per-executor // strategies) into runtime/lifecycle/skill so every launch — kanban or // office — flows through the same path. // // Implementations are responsible for short-circuiting when the profile has // nothing to deploy. The runtime does NOT pre-filter beyond a basic // empty-profile check, so each implementation can decide what "empty" // means for its delivery strategy. type SkillDeployer interface { DeploySkills(ctx context.Context, req SkillDeployRequest) (SkillDeployResult, error) } // SkillDeployRequest carries everything a SkillDeployer needs to materialise // per-agent skills + prompt before launch. The Profile is the resolved row // from the agent_profiles table; WorkspacePath is the host filesystem path // the agent will run inside; ExecutorType is the executor backend // (local_pc / local_docker / sprites) so the deployer can pick a strategy. type SkillDeployRequest struct { Profile *settingsmodels.AgentProfile WorkspacePath string ExecutorType string WorkspaceID string SessionID string } // SkillDeployResult carries the side-effects a successful deploy produced // that the runtime needs to wire onto the in-flight LaunchRequest: // - Metadata: per-executor metadata keys (e.g. kandev_runtime_dir for // Docker bind-mounts, skill_manifest_json for Sprites uploads). // - InstructionsDir: the on-host or sprite-side path where instruction // files were materialised. Office's prompt builder reads from this // path; kanban launches that don't consume instructions can ignore it. type SkillDeployResult struct { Metadata map[string]any InstructionsDir string } // AgentProfileReader is the minimal surface SkillDeployer implementations and // the runtime need to fetch a full profile row by id. It mirrors the // agent settings store's GetAgentProfile method without forcing a full // import of the store package on every call site. type AgentProfileReader interface { GetAgentProfile(ctx context.Context, id string) (*settingsmodels.AgentProfile, error) } // noopSkillDeployer is the default. When skill_ids and desired_skills are // empty (or no deployer has been wired), this is what runs — a fast-path // no-op. type noopSkillDeployer struct{} // DeploySkills satisfies SkillDeployer with a zero-effort implementation. func (noopSkillDeployer) DeploySkills(_ context.Context, _ SkillDeployRequest) (SkillDeployResult, error) { return SkillDeployResult{}, nil } // NoopSkillDeployer returns the default deployer used when no concrete // strategy has been wired. Kanban-only setups without office wiring get // this; office plugs its own deployer in via Manager.SetSkillDeployer. func NoopSkillDeployer() SkillDeployer { return noopSkillDeployer{} }