Files
fusion/packages/engine/src/agent-runtime.ts
Fusion bccc39f741 feat(FN-3961): propagate taskEnv into agent subprocess sessions
Propagates `taskEnv` (task-scoped environment variables) through the agent session creation pipeline, including executor sessions, spawned child agents, and workflow sessions. The engine's session factory (`agent-session-helpers.ts`), executor, PI agent creation (`pi.ts`), and step session executor

Fusion-Task-Id: FN-3961
2026-05-10 23:36:10 -07:00

166 lines
6.6 KiB
TypeScript

/**
* Agent runtime adapter abstraction layer.
*
* Provides a typed interface for creating and managing agent sessions
* across different runtime implementations (default pi runtime, plugin-provided runtimes).
*
* ## Interface Contract
*
* All runtimes must implement:
* - `id`: Unique runtime identifier (e.g., "pi", "code-interpreter")
* - `name`: Human-readable name
* - `createSession()`: Create a new agent session
* - `promptWithFallback()`: Prompt with automatic retry/compaction
* - `describeModel()`: Get model description from session
*/
import type { AgentSession, SessionManager, ToolDefinition } from "@mariozechner/pi-coding-agent";
import type { PermanentAgentGatingContext } from "@fusion/core";
import type { SkillSelectionContext } from "./skill-resolver.js";
import type { FallbackModelUsedPayload } from "./pi.js";
import type { AgentActionGateContext } from "./agent-action-gate.js";
import type { SystemPromptLayers } from "./prompt-layers.js";
/**
* Options for creating an agent session.
* Mirrors the options accepted by createFnAgent.
*/
export interface AgentRuntimeContext {
sessionPurpose?: string;
toolMode?: "coding" | "readonly";
customToolNames?: string[];
requestedSkillNames?: string[];
}
export interface AgentRuntimeOptions {
/** Working directory for the agent session */
cwd: string;
/** System prompt for the agent */
systemPrompt: string;
/**
* Optional structured prompt layers for cross-session caching.
* When present, runtimes that support prompt caching use the `stable`
* layer as a cacheable prefix and the `dynamic` layer as the per-session
* suffix. Runtimes that don't support caching ignore this and use
* `systemPrompt` (the collapsed string) instead.
*
* Callers MUST also provide `systemPrompt` as the collapsed equivalent
* for backward compatibility.
*/
systemPromptLayers?: SystemPromptLayers;
/** Tool set to use: "coding" for full tools, "readonly" for read-only access */
tools?: "coding" | "readonly";
/** Additional custom tools to merge with the base toolset */
customTools?: ToolDefinition[];
/** Callback for text output from the agent */
onText?: (delta: string) => void;
/** Callback for thinking/thought output from the agent */
onThinking?: (delta: string) => void;
/** Callback when a tool starts execution */
onToolStart?: (name: string, args?: Record<string, unknown>) => void;
/** Callback when a tool finishes execution */
onToolEnd?: (name: string, isError: boolean, result?: unknown) => void;
/** Default model provider (e.g. "anthropic") */
defaultProvider?: string;
/** Default model ID within the provider (e.g. "claude-sonnet-4-5") */
defaultModelId?: string;
/** Optional fallback model provider for retryable errors */
fallbackProvider?: string;
/** Optional fallback model ID */
fallbackModelId?: string;
/** Default thinking effort level (e.g. "medium", "high") */
defaultThinkingLevel?: string;
/** Optional pre-configured SessionManager for persistence */
sessionManager?: SessionManager;
/** Optional skill selection context */
skillSelection?: SkillSelectionContext;
/** Convenience: skill names to include in the session */
skills?: string[];
/** Runtime-facing context for non-pi runtimes that cannot consume JS ToolDefinition objects directly. */
runtimeContext?: AgentRuntimeContext;
/** Optional task-scoped environment variables for session-local subprocesses. */
taskEnv?: NodeJS.ProcessEnv;
/**
* Last-chance abort hook fired by the runtime *immediately before* the
* underlying LLM session is instantiated — i.e., after all of the runtime's
* own awaited setup work (provider registration, resource loading, etc.).
* Throw from this callback to cancel session creation.
*
* Runtimes SHOULD invoke this hook at their latest synchronous decision
* point so callers can enforce time-sensitive predicates (notably the
* engine pause flag) without a TOCTOU window between an outer check and
* the actual session spawn. Runtimes that ignore this hook degrade
* gracefully — the caller's outer check still fires before
* `runtime.createSession()` is invoked, so the abort window is bounded by
* the runtime's internal setup latency rather than unbounded.
*/
beforeSpawnSession?: () => Promise<void> | void;
/** Callback fired when runtime falls back from primary model to fallback model. */
onFallbackModelUsed?: (payload: FallbackModelUsedPayload) => Promise<void> | void;
/** Optional task context for fallback notifications. */
taskId?: string;
taskTitle?: string;
actionGateContext?: AgentActionGateContext;
/** Permanent-agent action gating context for v1 category classification enforcement. */
permanentAgentGating?: PermanentAgentGatingContext;
}
/**
* Result of creating an agent session.
*/
export interface AgentSessionResult {
/** The created agent session */
session: AgentSession;
/** Path to the persisted session file (undefined for in-memory sessions) */
sessionFile?: string;
}
/**
* Agent runtime adapter interface.
*
* All session runtimes (default pi runtime, plugin-provided runtimes) must
* implement this interface to ensure consistent behavior across engine subsystems.
*
* ## Implementation Notes
*
* - `createSession()` should return a fully initialized session ready for prompting
* - `promptWithFallback()` should handle retry/compaction automatically
* - `describeModel()` should return a human-readable model identifier
*/
export interface AgentRuntime {
/** Unique runtime identifier (e.g., "pi", "code-interpreter", "web-search") */
readonly id: string;
/** Human-readable name for the runtime */
readonly name: string;
/**
* Create a new agent session.
*
* @param options - Session creation options
* @returns Promise resolving to the session result
*/
createSession(options: AgentRuntimeOptions): Promise<AgentSessionResult>;
/**
* Prompt the session with user input.
*
* Implementations should handle:
* - Automatic retry on transient errors
* - Context compaction on context limit errors
* - Model fallback on retryable model selection errors
*
* @param session - The session to prompt
* @param prompt - The prompt text
* @param options - Optional prompt options (e.g., images for vision)
*/
promptWithFallback(session: AgentSession, prompt: string, options?: unknown): Promise<void>;
/**
* Get a human-readable model description from a session.
*
* @param session - The session to describe
* @returns Model description (e.g., "anthropic/claude-sonnet-4-5") or "unknown model"
*/
describeModel(session: AgentSession): string;
}