Break down a task and propose a scoped plan without editing files. Recommended timeout: 10 minutes (600s).

Investigate code, docs, and context in read-only mode and report findings. Recommended timeout: 20 minutes (1200s).

Evaluate an idea, scope an MVP, and recommend whether it is worth building. Recommended timeout: 10 minutes (600s).

Implement a change, run focused verification, and report results. Recommended timeout: 20 minutes (1200s).

Run existing repo validation commands (build, test, lint) and report the results. Recommended timeout: 10 minutes (600s).

Review code with one agent (default) or fan out to multiple specialist agents. Recommended timeout: 10 minutes single-agent, 20 minutes multi-agent (pass timeout_seconds=1200 when using multiple agents).

Use reverse-engineering tooling and broader roots to investigate binaries. Recommended timeout: 30 minutes (1800s).

Start a Dobby tool in the background and return immediately with a task id. Follow up with wait_run(task_id=...) to block on one run, or call start_run several times and then wait_run(task_ids=[...]) to be woken by whichever finishes first. get_run/list_runs remain available for a non-blocking peek. Recommended when your MCP client enforces short tools/call timeouts.

Get the status or final ToolResponse for a Dobby run by task id. This can recover results from .codex-dobby/runs even after a blocking tools/call timed out.

Call after start_run to block until a background run finishes. Pass task_id for one run, task_ids=[...] to wait for whichever of several finishes FIRST, or omit both to wait on every currently-live run for the repo. On completion returns the winning run's final ToolResponse (pending_task_ids lists any still-running siblings); on timeout returns a RUNNING lookup whose summary says to keep calling wait_run with pending_task_ids until one finishes. Background tasks are shielded from waiter cancellation, so re-calling never loses work. Default 540s (9 min); clamped to [1, 100_000s / ~27.8h], matching Claude Code's MCP_TOOL_TIMEOUT default. Pick timeout_seconds below your own MCP client's tools/call ceiling — Claude Code defaults to ~28h (the full clamp is usable); Codex CLI defaults to 60s per [mcp_servers.].tool_timeout_sec (raise that first); Claude Desktop / Cursor / Cline / Continue vary.

List recent Dobby runs for a repo. Useful for recovering task ids and results after a caller-side timeout.