Launch an autonomous coding agent for implementation, bug fixes, and refactoring. Runs in COMPLETE ISOLATION — the prompt + context_files are its ONLY context. The coder is always the final implementation stage — investigate and plan before coding for non-trivial tasks.

context_files are MANDATORY and ONLY .md files are accepted. Pass .ts/.js/.json and it WILL fail. Create .md specs via launch-super-planner first, or write one yourself.

Workflow: researcher → planner → CODER → tester After planner completes, read task:///{id} to get workspace path, then pass ALL .md files from that workspace as context_files here. Don't cherry-pick — send everything. Coder writes detailed testing notes to .agent-workspace/implementation/[topic]/HANDOFF.md — including Playwright hints for UI or curl commands for APIs — which the tester consumes.

Status: Read task:///all every ~30s to monitor all tasks (status, deps, questions). Statuses: running, waiting → <dep>, waiting_answer ⏸, completed, failed.

Launch an autonomous planning agent. Designs architecture and creates implementation plans as .md files. Always uses claude-opus-4.6 regardless of model parameter. Use this for any non-trivial task — if the work touches 3+ files or has ambiguous requirements, plan first.

Workflow: researcher → PLANNER → coder → tester Output goes to .agent-workspace/plans/[topic]/. After completion, read task:///{id} to get workspace path, then pass ALL .md files from that workspace as context_files to launch-super-coder. Send everything — builder-briefing.md, tester-checklist.md, task specs, the full workspace.

Status: Read task:///all every ~30s to monitor all tasks (status, deps, questions). Statuses: running, waiting → <dep>, waiting_answer ⏸, completed, failed.

Launch an autonomous testing agent. Primarily E2E testing with Playwright (browser flows, visual, interactions) but also handles API testing (curl + jq), running existing test suites, and any verification that proves the code works in the real world. Runs in COMPLETE ISOLATION.

context_files are MANDATORY — any file type accepted (source, tests, handoff docs). Pass ALL files from the coder's agent workspace — especially HANDOFF.md which contains testing instructions, curl examples, and Playwright hints.

Workflow: researcher → planner → coder → TESTER Chain with depends_on after coder.

Status: Read task:///all every ~30s to monitor all tasks (status, deps, questions). Statuses: running, waiting → <dep>, waiting_answer ⏸, completed, failed.

Launch an autonomous research agent. Investigates codebases, APIs, libraries, and technical topics. Produces .md research documents for downstream agents. Investigate before you solve — use this before planning or coding when the problem space is unclear.

Workflow: RESEARCHER → planner → coder → tester Output goes to .agent-workspace/researches/[topic]/. After completion, read task:///{id} to get the workspace path, then pass ALL .md files from that workspace as context_files to the next agent (planner or coder). Don't cherry-pick — send everything.

Status: Read task:///all every ~30s to monitor all tasks (status, deps, questions). Statuses: running, waiting → <dep>, waiting_answer ⏸, completed, failed.

Launch a general-purpose autonomous agent for non-specialized tasks: documentation, analysis, file organization, data processing, report generation, or any automation that doesn't fit coder/planner/tester/researcher. Investigate first — have the agent explore and understand before producing output. Keeps context across the session — good for iterative or exploratory work.

Status: Read task:///all every ~30s to monitor all tasks (status, deps, questions). Statuses: running, waiting → <dep>, waiting_answer ⏸, completed, failed.

Send a follow-up message to an existing agent session. Resumes the session — the agent continues from where it left off.

Returns a NEW task_id — the original task stays terminal. Monitor the new ID for progress.

When to call: Continue a completed/failed/rate-limited agent with follow-up instructions, or resume with default "continue".

Find task_id: Read task:///all — pick a terminal task (completed, failed, rate_limited, timed_out) to resume.

Cancel running agents or clear the entire workspace.

Cancel: { "task_id": "abc123" } or { "task_id": ["abc", "def"] } (max 50). Clear all: { "task_id": "all", "clear": true } — kills active agents, removes all state.

Running agents are killed (SIGTERM → SIGKILL). Terminal agents are removed from memory.

Find task_id: Read task:///all.

Submit an answer to a pending question from an agent. When an agent pauses because it asked a question, use this to respond and resume execution.

When to call: Read task:///all — tasks with status waiting_answer have a "Pending Questions" section showing the question, choices, and an example answer-agent call.

Single-question flows (Copilot / Claude): Use the answer field.

• Choice by number: "1", "2", "3" — selects the corresponding option

• Choice by text: Exact text of a choice option

• Custom answer: "OTHER: your custom text" — for freeform responses when choices don't fit

Multi-question flows (Codex): Use the answers field with a map of question IDs to answers. Read task:///{id} → pending_question.structured_questions to get question IDs.

Examples:

answer-agent { "task_id": "abc123", "answer": "2" }
answer-agent { "task_id": "abc123", "answer": "OTHER: Use TypeScript instead" }
answer-agent { "task_id": "abc123", "answers": { "q_build_system": "1", "q_language": "TypeScript" } }

Find pending questions: Read task:///all — look for the "Pending Questions" section.