@bluecopa/harness 0.1.0-snapshot.13 → 0.1.0-snapshot.130

package/README.md

@@ -2,9 +2,17 @@
 
 Provider-agnostic TypeScript agent framework with Claude-code-compatible tool semantics.
 
-The harness provides the core loop that drives an AI agent: send messages to an LLM, execute the tool calls it returns, feed results back, and repeat until the LLM produces a final text response.
+Published on npm as **`@bluecopa/harness`**.
 
-## Quickstart
+Two execution modes: a simple single-agent loop (`createAgent` + `VercelAgentLoop`) and a process-based orchestrator (`ArcLoop`) that dispatches parallel processes with context management, memory, and resilience.
+
+## Install
+
+```bash
+pnpm add @bluecopa/harness
+```
+
+## Development
 
 ```bash
 pnpm install
@@ -13,9 +21,11 @@ pnpm test
 
 ## Architecture
 
+### Single-Agent Loop
+
 ```
 ┌──────────────┐     ┌──────────────┐     ┌──────────────────┐
-│  createAgent │────▶│  AgentLoop   │────▶│  LLM (Claude)    │
+│  createAgent │────►│  AgentLoop   │────►│  LLM (Claude)    │
 │  (turn loop) │     │  (nextAction)│     │                  │
 └──────┬───────┘     └──────────────┘     └──────────────────┘
        │                                           │
@@ -27,20 +37,82 @@ pnpm test
 └──────────────┘
 ```
 
-1. `createAgent` drives a deterministic step loop
-2. Each step calls `loop.nextAction(messages)` to get the LLM's decision
-3. If it's a tool call, the harness executes it via `ToolProvider` and appends the result
-4. If it's a final action, the loop ends and returns the result
+### ArcLoop Orchestrator
 
-## Using with the sandbox
+```
+Orchestrator (ArcLoop — Opus 4.6 by default)
+  │  tools: Thread, Check, Cancel, Remember, ReadEpisode
+  │
+  │  Turn 1 (parallel):
+  ├──► Process 0 ("read auth", model=fast)        ─┐
+  ├──► Process 1 ("read routes", model=fast)       ─┼──► Episodes
+  ├──► Process 2 ("read tests", model=fast)        ─┘
+  │
+  │  Turn 2 (dispatch dependent work):
+  ├──► Thread("fix bug", context=[ep0,ep1,ep2])    ──► Episode
+  │
+  │  Turn 3 (parallel):
+  ├──► Thread("run tests", context=[ep3])          ─┐
+  ├──► Thread("update docs", context=[ep3])        ─┘
+  │
+  └──► Final text response
+```
 
-The most common setup connects the harness to a running sandbox service via `ControlPlaneE2BExecutor`:
+Full architecture doc: [`docs/arc.md`](../docs/arc.md)
 
-```ts
-import { createAgent } from './src/agent/create-agent';
-import { E2BToolProvider } from './src/providers/e2b-tool-provider';
+---
+
+## ToolProvider
+
+The contract for tool execution. All agent modes use this interface.
+
+```typescript
+interface ToolProvider {
+  bash(command: string, options?: BashOptions): Promise<ToolResult>;
+  readFile(path: string, options?: ReadOptions): Promise<ToolResult>;
+  writeFile(path: string, content: string): Promise<ToolResult>;
+  editFile(path: string, oldText: string, newText: string): Promise<ToolResult>;
+  glob(pattern: string, options?: GlobOptions): Promise<ToolResult>;
+  grep(pattern: string, path?: string, options?: GrepOptions): Promise<ToolResult>;
+  webFetch?(options: WebFetchOptions): Promise<ToolResult>;
+  webSearch?(query: string): Promise<ToolResult>;
+  capabilities(): ToolProviderCapabilities;
+}
+
+interface ToolResult {
+  success: boolean;
+  output: string;
+  error?: string;
+}
+```
+
+Built-in implementations:
+
+| Provider | Description |
+|----------|-------------|
+| `LocalToolProvider` | Runs tools on the local filesystem |
+| `E2BToolProvider` | Routes tools to a sandbox VM via `ControlPlaneE2BExecutor` |
+| `CompositeToolProvider` | Combines multiple providers (e.g. local filesystem + sandbox bash) |
+
+## SandboxProvider
+
+Higher-level sandbox operations beyond basic tool calls:
+
+```typescript
+interface SandboxProvider {
+  exec(command: string, options?: SandboxExecOptions): Promise<SandboxExecResult>;
+  readSandboxFile(path: string): Promise<SandboxFileBlob>;
+  writeSandboxFile(path: string, content: SandboxFileBlob): Promise<void>;
+}
+```
+
+Used by `SkillManager` for executing skill scripts in isolated VMs.
+
+## Connecting to a Sandbox
+
+```typescript
 import { ControlPlaneE2BExecutor } from './src/providers/control-plane-e2b-executor';
-import { VercelAgentLoop } from './src/loop/vercel-agent-loop';
+import { E2BToolProvider } from './src/providers/e2b-tool-provider';
 
 // Connect to sandbox service
 const executor = new ControlPlaneE2BExecutor({
@@ -50,187 +122,172 @@ const executor = new ControlPlaneE2BExecutor({
 });
 await executor.initialize();  // creates a Firecracker VM
 
-// Build and run the agent
-const agent = createAgent({
-  toolProvider: new E2BToolProvider(executor),
-  loop: new VercelAgentLoop(),  // needs ANTHROPIC_API_KEY
-});
+const toolProvider = new E2BToolProvider(executor);
 
-const result = await agent.run('create a bar chart of sales data');
-console.log(result.output);   // LLM's final response
-console.log(result.steps);    // number of tool steps
+// ... use with createAgent or ArcLoop
 
-await executor.destroy();     // tears down the VM
+await executor.destroy();  // tears down the VM
 ```
 
-For a complete working example, see [`examples/chat-assistant/src/chat.ts`](../examples/chat-assistant/src/chat.ts).
-
-### From environment variables
+From environment variables: `ControlPlaneE2BExecutor.fromEnv()` reads `SAMYX_BASE_URL` and `SAMYX_API_KEY`.
 
-`ControlPlaneE2BExecutor.fromEnv()` reads `SAMYX_BASE_URL` and `SAMYX_API_KEY` automatically:
+---
 
-```ts
-const executor = ControlPlaneE2BExecutor.fromEnv();
-```
-
-## Using locally (no sandbox)
+## Single-Agent Mode (`createAgent`)
 
-For development without a sandbox service, use `LocalToolProvider` which runs tools on the local machine:
+For simple tasks that don't need orchestration:
 
-```ts
+```typescript
 import { createAgent } from './src/agent/create-agent';
 import { LocalToolProvider } from './src/providers/local-tool-provider';
 
 const agent = createAgent({
   toolProvider: new LocalToolProvider(process.cwd()),
-  loop: new VercelAgentLoop(),
+  loop: new VercelAgentLoop(),  // needs ANTHROPIC_API_KEY
 });
 
 const result = await agent.run('list all TypeScript files');
+console.log(result.output);
 ```
 
-## Key modules
-
-### Agent creation (`src/agent/create-agent.ts`)
+### Configuration
 
-`createAgent(options)` returns an agent with a `.run(prompt, options?)` method. Options:
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `toolProvider` | `ToolProvider` | required | Executes tool calls |
+| `loop` | `AgentLoop` | `VercelAgentLoop` | LLM decision loop |
+| `sandboxProvider` | `SandboxProvider` | — | Higher-level sandbox operations |
+| `maxSteps` | `number` | 30 | Max tool steps per run |
+| `telemetry` | `HarnessTelemetry` | — | OpenTelemetry-style tracing |
+| `skillIndexPath` | `string` | — | Path to skill index JSON for routing |
 
-| Option | Type | Description |
-|--------|------|-------------|
-| `toolProvider` | `ToolProvider` | Required. Executes tool calls |
-| `loop` | `AgentLoop` | LLM decision loop (default: `VercelAgentLoop`) |
-| `sandboxProvider` | `SandboxProvider` | Optional. Higher-level sandbox ops (file download, exec with env) |
-| `maxSteps` | `number` | Max tool steps per run (default: 30) |
-| `telemetry` | `HarnessTelemetry` | Optional. OpenTelemetry-style tracing |
-| `skillIndexPath` | `string` | Optional. Path to skill index JSON |
+### VercelAgentLoop
 
-### Agent loop (`src/loop/vercel-agent-loop.ts`)
+Calls Claude via the Vercel AI SDK. Supports parallel tool calls and configurable system prompt.
 
-`VercelAgentLoop` calls Claude via the Vercel AI SDK. It supports:
-- Parallel tool calls (returns `ToolBatchAction` when the LLM requests multiple tools at once)
-- Configurable system prompt
-- Model selection via `HARNESS_MODEL` env var (default: `claude-sonnet-4-5`)
-
-```ts
+```typescript
 const loop = new VercelAgentLoop({
   systemPrompt: 'You are a helpful coding assistant.',
+  model: 'claude-sonnet-4-5',  // or HARNESS_MODEL env var
 });
 ```
 
-### Tool provider (`src/interfaces/tool-provider.ts`)
+### LCMToolLoop
 
-The contract for tool execution:
+Wraps another loop to add Lossless Context Management and optional REPL orchestration:
 
-```ts
-interface ToolProvider {
-  bash(command: string, options?: BashOptions): Promise<ToolResult>;
-  readFile(path: string, options?: ReadOptions): Promise<ToolResult>;
-  writeFile(path: string, content: string): Promise<ToolResult>;
-  editFile(path: string, oldText: string, newText: string): Promise<ToolResult>;
-  glob(pattern: string, options?: GlobOptions): Promise<ToolResult>;
-  grep(pattern: string, path?: string, options?: GrepOptions): Promise<ToolResult>;
-  webFetch?(options: WebFetchOptions): Promise<ToolResult>;
-  webSearch?(query: string): Promise<ToolResult>;
-  capabilities(): ToolProviderCapabilities;
-}
+```typescript
+import { LCMToolLoop } from './src/loop/lcm-tool-loop';
+import { VercelAgentLoop } from './src/loop/vercel-agent-loop';
 
-interface ToolResult {
-  success: boolean;
-  output: string;
-  error?: string;
-}
+const loop = new LCMToolLoop({
+  innerLoop: new VercelAgentLoop(),
+  toolProvider: mySandboxProvider,
+  enableRepl: true,           // default: true
+  bridgeDir: '/var/run/bridge',
+  onActivity: (entry) => console.log(entry),
+  onLlmRequest: async (prompt) => callLLM(prompt),
+  onWebFetchRequest: async (url) => fetch(url),
+});
 ```
 
-Built-in implementations:
+**Standard mode**: Lossless context trimming — the LLM always sees a coherent, budget-fitting view of the full conversation.
 
-| Provider | Description |
-|----------|-------------|
-| `LocalToolProvider` | Runs tools on the local filesystem |
-| `E2BToolProvider` | Routes tools to an E2B-compatible executor over HTTP |
-| `CompositeToolProvider` | Combines multiple providers (e.g. sandbox + web) |
+**REPL mode**: When the LLM returns a Bash action with the REPL marker, the loop writes a Python script into the sandbox, injects the bridge module, runs the script, and polls for sub-requests (LLM, web_fetch, ask_user) that the harness fulfills.
 
-### Action types (`src/agent/types.ts`)
+---
 
-The LLM returns one of these action types each turn:
+## ArcLoop (Orchestrator Mode)
 
-```ts
-// Single tool call
-interface ToolCallAction {
-  type: 'tool';
-  name: 'Bash' | 'Read' | 'Write' | 'Edit' | 'Glob' | 'Grep' | ...;
-  args: Record<string, unknown>;
-}
+For complex tasks that benefit from parallel processes, context management, and memory:
 
-// Multiple independent tool calls (executed in parallel)
-interface ToolBatchAction {
-  type: 'tool_batch';
-  calls: ToolCallAction[];
-}
-
-// Final text response (ends the loop)
-interface FinalAction {
-  type: 'final';
-  content: string;
-}
-```
-
-### LCM tool loop (`src/loop/lcm-tool-loop.ts`)
-
-`LCMToolLoop` wraps another loop to add LCM-based tool routing, REPL script execution, and bridge-based tool dispatch. Used in the chat-assistant example.
-
-### Sandbox provider (`src/interfaces/sandbox-provider.ts`)
+```typescript
+import { createArcAgent } from './src/arc/create-arc-agent';
 
-Higher-level sandbox operations beyond basic tool calls:
+const agent = await createArcAgent({
+  toolProvider: myToolProvider,
+  episodeStore: myEpisodeStore,       // required
+  sessionMemoStore: mySessionMemoStore, // required
+  longTermStore: myLongTermStore,       // required
+  taskId: 'task-1',
+  sessionId: 'session-1',
+});
 
-```ts
-interface SandboxProvider {
-  exec(command: string, options?: SandboxExecOptions): Promise<SandboxExecResult>;
-  readSandboxFile(path: string): Promise<SandboxFileBlob>;
-  writeSandboxFile(path: string, content: SandboxFileBlob): Promise<void>;
+// Streaming
+for await (const event of agent.stream(messages, signal)) {
+  if (event.type === 'text_delta') process.stdout.write(event.text);
+  if (event.type === 'process_dispatched') console.log(`  → ${event.action}`);
+  if (event.type === 'done') console.log(`Done in ${event.stats.durationMs}ms`);
 }
-```
-
-### Observability (`src/observability/otel.ts`)
-
-`HarnessTelemetry` provides OpenTelemetry-style spans and metrics for agent runs.
 
-### Arc: Orchestrator + Thread Architecture (`src/arc/`)
+// Non-streaming
+const result = await agent.run(messages, signal);
+```
 
-`ArcLoop` is an `AgentLoop` implementation where an orchestrator LLM dispatches bounded threads via a single `Thread` tool. Threads produce episodes (summary + full trace). The orchestrator only sees summaries, keeping its context small.
+### ArcLoopConfig
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `model` | `string` | `'claude-opus-4-6'` | Orchestrator model (ID or tier name) |
+| `modelMap` | `Record<ModelTier, string>` | haiku/sonnet/opus | Maps fast/medium/strong to model IDs |
+| `apiKey` | `string` | — | Anthropic API key |
+| `systemPrompt` | `string` | built-in | Custom orchestrator system prompt |
+| `maxTurns` | `number` | 30 | Max orchestrator turns |
+| `processTimeout` | `number` | 120_000 | Per-process timeout (ms) |
+| `processMaxSteps` | `number` | 20 | Per-process max tool steps |
+| `contextWindowSize` | `number` | 200_000 | Context window in tokens |
+| `outputReserve` | `number` | 20_000 | Tokens reserved for output |
+| `autoMemory` | `boolean` | true | Auto-detect patterns from episodes |
+| `episodeStore` | `EpisodeStore` | required | Stores episode summaries + traces |
+| `sessionMemoStore` | `SessionMemoStore` | required | Stores session memos |
+| `longTermStore` | `LongTermStore` | required | Stores long-term memories |
+| `taskId` | `string` | required | Task identifier |
+| `sessionId` | `string` | required | Session identifier |
+| `toolProvider` | `ToolProvider` | required | Tool execution |
+| `processTools` | `Record<string, AnyTool>` | builtinTools | Tools available inside processes |
+| `extraOrchestratorTools` | `Record<string, AnyTool>` | — | Custom orchestrator tools |
+| `onOrchestratorTool` | `function` | — | Handler for custom orchestrator tools |
+| `resilience` | `ResiliencePolicy` | — | Composable resilience pipeline |
+| `traceWriter` | `function` | — | Callback for trace event emission |
+
+### Resilience
+
+```typescript
+import { resilience } from './src/arc/resilience';
+
+const pipeline = resilience()
+  .retry({ maxRetries: 2, baseDelay: 1000 })
+  .timeout({ durationMs: 30_000 })
+  .circuitBreaker({ failureThreshold: 5 })
+  .build();
+
+const agent = await createArcAgent({
+  // ...config
+  resilience: pipeline,
+});
+```
 
-```ts
-import { createArcAgent } from './src/arc/create-arc-agent';
-import { InMemoryEpisodeStore } from './src/arc/stores/episode-store';
-import { InMemorySessionMemoStore } from './src/arc/stores/session-memo-store';
-import { InMemoryLongTermStore } from './src/arc/stores/long-term-store';
+### Trace Emission
 
-const agent = createArcAgent({
-  toolProvider: new LocalToolProvider(process.cwd()),
-  episodeStore: new InMemoryEpisodeStore(),
-  sessionMemoStore: new InMemorySessionMemoStore(),
-  longTermStore: new InMemoryLongTermStore(),
-  taskId: 'task-1',
-  sessionId: 'session-1',
+```typescript
+const traces: TraceEvent[] = [];
+const agent = await createArcAgent({
+  // ...config
+  traceWriter: (event) => traces.push(event),
 });
-
-const result = await agent.run('Fix the authentication bug');
 ```
 
-Key features:
-- **Parallel threads**: orchestrator calls Thread N times in one turn → all run concurrently
-- **Four-tier memory**: thread context → episodes → session memos → long-term
-- **Per-thread models**: Haiku for reads, Sonnet for implementation
-- **Template compression**: zero-LLM-call episode summaries
-- **Async consolidation**: non-blocking background distillation
+Traces can be validated against the formal model: `cd verify && cargo run -- trace file.ndjson`
 
-Full architecture doc: [`docs/arc.md`](../docs/arc.md)
+---
 
-## Package layout
+## Package Layout
 
 ```
 src/
 ├── agent/          # createAgent, step executor, types
-├── arc/            # ArcLoop orchestrator, threads, memory hierarchy
+├── arc/            # ArcLoop orchestrator, processes, memory, resilience
+│   ├── resilience/ # Retry, circuit breaker, timeout, bulkhead, fallback
 │   ├── stores/     # RxDB + in-memory store implementations
 │   └── object-store/ # Pluggable cloud sync (fs, memory)
 ├── interfaces/     # ToolProvider, SandboxProvider, AgentLoop contracts
@@ -240,17 +297,20 @@ src/
 ├── hooks/          # Pre/post tool call hooks
 ├── permissions/    # Tool permission checks
 ├── sessions/       # Session persistence
-├── subagents/      # Subagent spawning and task tools
+├── subagents/      # Subagent spawning
 ├── skills/         # Skill index, routing, and management
 ├── optimization/   # Benchmark runner
 └── observability/  # OpenTelemetry integration
+
+verify/             # Rust formal verification (Stateright model checker)
+testing/            # Adversarial scenario replay harness
+tests/              # Vitest test suite
 ```
 
 ## Documentation
 
-- **Arc architecture**: [`docs/arc.md`](../docs/arc.md)
-- Provider guide: `docs/guides/providers.md`
-- Skills guide: `docs/guides/skills.md`
-- Observability guide: `docs/guides/observability.md`
-- Release process: `../docs/RELEASE.md`
-- Full example: [`../examples/chat-assistant/src/chat.ts`](../examples/chat-assistant/src/chat.ts)
+- [Arc architecture](../docs/arc.md) — process model, context window, memory, resilience, verification
+- [Testing](../docs/testing.md) — test layers, running tests, writing new tests
+- [Sandbox setup](../docs/PUBLIC_SANDBOX.md) — deploying the sandbox service
+- [Release process](../docs/RELEASE.md) — versioning and publishing
+- [Example](../examples/chat-assistant/src/chat.ts) — complete working chat assistant

package/dist/arc/app-adapter.d.ts

@@ -0,0 +1,109 @@
+import { A as ArcEvent, S as StepUsage } from '../types-B3VY0xnE.js';
+import 'zod';
+import './profile-graph.js';
+import 'ai';
+
+type ArcAppRunStats = {
+    tools: number;
+    commands: number;
+    reads: number;
+    searches: number;
+    fileChanges: number;
+    failures: number;
+    inputTokens: number;
+    outputTokens: number;
+};
+type ArcAppProcessMeta = {
+    label?: string;
+    profile?: string;
+    model?: string;
+    dispatchedAt: number;
+    status: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+    step?: number;
+    usage?: StepUsage;
+};
+type ArcAppBackgroundTurn = {
+    id: string;
+    label?: string;
+    profile?: string;
+    model?: string;
+    dispatchedAt: number;
+};
+type ArcAppState = {
+    activeTurnId: string | null;
+    processes: Record<string, ArcAppProcessMeta>;
+    backgroundTurns: Record<string, ArcAppBackgroundTurn>;
+    runStats: ArcAppRunStats;
+};
+type ArcAppObservedActivity = {
+    phase: 'start' | 'end';
+    processId?: string;
+    toolName: string;
+    args?: Record<string, unknown>;
+    ok?: boolean;
+    preview?: string;
+    command?: string;
+    path?: string;
+};
+type ArcAdapterInputEvent = ArcEvent;
+type ArcAppEvent = {
+    type: 'state_changed';
+    state: ArcAppState;
+} | {
+    type: 'stats_changed';
+    stats: ArcAppRunStats;
+} | {
+    type: 'activity_observed';
+    activity: ArcAppObservedActivity;
+} | {
+    type: 'assistant_progress';
+    id?: string;
+    text: string;
+} | {
+    type: 'process_progress';
+    id: string;
+    meta: ArcAppProcessMeta;
+} | {
+    type: 'process_summary';
+    id: string;
+    meta?: ArcAppProcessMeta;
+    summary: string;
+    display: string;
+} | {
+    type: 'process_failed';
+    id?: string;
+    error: string;
+} | {
+    type: 'assistant_output';
+    text: string;
+};
+type Listener = (event: ArcAppEvent) => void;
+declare function formatArcAppProcessSummary(summary: string): string;
+declare function createEmptyArcAppRunStats(): ArcAppRunStats;
+declare function formatArcAppRunStats(stats: ArcAppRunStats): string;
+declare function buildArcAppRunSummary(stats: ArcAppRunStats, modelId: string): string | null;
+declare function createInitialArcAppState(): ArcAppState;
+declare class ArcAppAdapter {
+    private state;
+    private listeners;
+    private bashCommands;
+    private processTextBuffers;
+    constructor(initialState?: ArcAppState);
+    subscribe(listener: Listener): () => void;
+    getState(): ArcAppState;
+    reset(): void;
+    beginTurn(turnId: string): void;
+    finishTurn(turnId: string): void;
+    backgroundTurn(turnId: string, meta: {
+        label?: string;
+        profile?: string;
+        model?: string;
+    }): boolean;
+    handleArcEvent(event: ArcAdapterInputEvent): void;
+    private emit;
+    private handleProcessActivity;
+    private handleProcessTextDelta;
+    private flushProcessTextBuffer;
+}
+
+export { type ArcAdapterInputEvent, ArcAppAdapter, type ArcAppBackgroundTurn, type ArcAppEvent, type ArcAppObservedActivity, type ArcAppProcessMeta, type ArcAppRunStats, type ArcAppState, buildArcAppRunSummary, createEmptyArcAppRunStats, createInitialArcAppState, formatArcAppProcessSummary, formatArcAppRunStats };