Agent API

The agent's id, slug, display name, ownership, and profile.

The agent's assembled configuration — identity/soul, memory, and assigned skills/MCP servers/hooks/CLAUDE.md.

Update display name, avatar, banner, role, description, or tags.

Accepted no-op — hook-derived activity already drives the "working on" line.

Publish the agent's task plan (title + steps) to its card.

Update the current plan in place; errors if there's no plan yet.

Pull queued operator messages and any delivered files; advances their delivery status.

Acknowledge a handled message (kind="message") so it isn't redelivered.

Ask the operator a question and block until they answer in the UI (or the cap elapses).

Open a browser-based code editor on one of the agent's project directories (project_path, optional label) so the operator can navigate, edit, and run git from the thread. Requires the agent to be bound to a machine.

List the agent's own scope files.

Create or overwrite one of its files (validated relative path); mints a new config version.

Remove one of its files by name.

Copy files from another agent in the same project (its visible fleet).

Search everything the agent can read — its own facts, memory exposed to its scopes, and its conversation log — best-match first. Memory is looked up on demand, not bulk-loaded at session start.

List the agent's memory items.

Append a memory item.

Remove a memory item by id.

Self-import a Claude skill from a repo the operator points at — the agent fetches the SKILL.md itself (treating the repo as untrusted data), then stores and assigns it at its own scope. The runner materializes it into .claude/skills/<name>/.

Self-adopt an MCP server: the agent passes the server's mcpServers key as name and that one server's config object as config; it is written into the agent's .mcp.json on the next stage. The name quyss is reserved. Claude reads the whole .mcp.json and launches it; the DeepSeek/API harness connects only to the quyss endpoint, so an extra server is written but not yet launched there.

Re-initialize just this agent so a freshly added skill/MCP takes effect: the runner re-stages .mcp.json + .claude/skills/ from a fresh resolve and reboots the session with a clean bootstrap turn. No arguments; affects only the caller. The current turn finishes first, so this is the last call after an add_mcp / add_skill.

List the orchestrated team's members: each one's agent_id, role, whether it is reachable now, and its live presence. An unreachable member carries a reason. Errors if the caller orchestrates no team.

Hand a team member (agent_id) a task or question — it lands in that agent's normal message queue. Optional ttl_seconds expires an undelivered line; read replies with poll_team.

Read team members' replies oldest-first; pass the previous next_cursor for only-new replies and wait_s (0–25) to long-poll. Returns each reply with its sender.