A3S Code

A3S Code v3.4.0 is a harness-driven runtime for coding agents. The runtime keeps the agent loop small and observable while the harness owns context assembly, tool execution, permission policy, delegation, verification evidence, and run replay.

v3.4.0 Highlights

The current execution shape is:

Agent / AgentSession
  -> context assembly
  -> optional planning
  -> automatic subagent delegation when configured
  -> tool selection or programmatic tool calling
  -> permission and confirmation policy
  -> execution
  -> trace, artifacts, and verification evidence
  -> compaction

Install

npm install @a3s-lab/code
pip install a3s-code
cargo add a3s-code-core

Configure

A3S Code uses ACL config files. Prefer .acl files and labeled blocks; JSON and legacy HCL config files are not supported.

default_model = "openai/MiniMax-M2.7-highspeed"
max_parallel_tasks = 8
auto_parallel = false

providers "openai" {
  apiKey = env("A3S_OPENAI_API_KEY")
  baseUrl = env("A3S_OPENAI_BASE_URL")

  models "MiniMax-M2.7-highspeed" {
    name = "MiniMax M2.7 Highspeed"
    tool_call = true
  }
}

agent_dirs = ["./.a3s/agents"]

auto_delegation {
  enabled        = true
  auto_parallel  = false
  min_confidence = 0.72
  max_tasks      = 4
}

storage_backend = "file"

The top-level auto_parallel setting is a global kill switch for automatic parallel fan-out. It does not remove parallel_task or the SDK session.tasks(...) helper.

Use

import { Agent } from '@a3s-lab/code';

const agent = await Agent.create('agent.acl');
const session = agent.session('/my-project', {
  builtinSkills: true,
  planningMode: 'auto',
  autoDelegation: { enabled: true, maxTasks: 4 },
  maxParallelTasks: 8,
  autoParallel: false,
});

const result = await session.send(
  'Use @review to inspect the authentication changes and summarize release blockers',
);

console.log(result.text);
console.log(result.verificationSummaryText);

Current Surface

• Conversation: send(), run(), and stream() for synchronous and streaming turns.
• Delegation: task() and tasks() wrap the core task / parallel_task tools and return compact child summaries.
• Orchestration: session.parallel(...), session.pipeline(...), and session.parallelResumable(...) express programmable fan-out, barrier-free pipelines, and resumable/migratable workflows over the AgentExecutor seam.
• Automatic subagents: enabled with autoDelegation or auto_delegation, bounded by maxParallelTasks / max_parallel_tasks, and suppressible with autoParallel: false / auto_parallel = false.
• Observability: runs(), runSnapshot(), runEvents(), activeTools(), and cancelRun() for live run/tool tracking.
• PTC: session.program(...) runs bounded JavaScript in QuickJS for deterministic tool chains.
• Structured output: generate_object produces schema-validated JSON with repair attempts.
• Configuration: .acl file paths or inline ACL strings are accepted by Agent.create().
• Persistence: file and memory stores support resumable sessions, explicit save(), autoSave, memory recall, and session IDs.

Start with API Contract, Sessions, Tasks, and Providers.