A3S Code
  • English
  • v9.1.0

      • Filesystem conventions

      A3S Code loads durable project behavior from the workspace filesystem. Roles, skills, and worker definitions that need to last are written as files, then loaded by convention. The goal is reviewable, versioned, reusable behavior — not a separate product runtime.

      Text
      repo/

      ├── AGENTS.md          # durable project instructions

      └── .a3s/

          ├── config.acl     # model, provider, queue, and delegation config

          ├── agents/        # worker agents for task / autoDelegation

          └── skills/        # reusable skills

      Path Map

      PathPurposeHow it loads
      AGENTS.mdStable workspace instructionsAutomatically, from the Git root down to the workspace. See AGENTS.md.
      .a3s/config.aclModel, provider, queue, skill dirs, and delegation configThe TUI discovers it. SDK hosts pass an ACL path or string to Agent.create.
      .a3s/agents/Worker/subagent definitionsAutomatically, together with ~/.a3s/agents and the .claude/agents compatibility directories.
      .a3s/skills/Reusable skillsOnly when listed in ACL skill_dirs or session skillDirs; there is no implicit skills scan.

      These conventions are not a new prompt system. AGENTS.md and skills enter A3S Code's context composition path; tool visibility, permission gates, HITL, response contracts, and verification remain harness-controlled.

      Loading Order

      Agent.create parses the ACL config and loads the configured skill_dirs into an agent-wide skill registry. Each session then binds a workspace and loads project instructions, session skill directories, agent definitions, worker agents, direct tools, MCP connections, and runtime policy.

      Design Rules

      1. Prefer files that can be reviewed in a PR over opaque in-memory defaults.
      2. Keep harness authority: files supply content; Core still owns safety and verification.
      3. Put worker definitions under .a3s/agents/.