Rufler has 14 tools with basic descriptions and parameter schemas, but significant gaps in LLM-optimization and error handling. Tool names are action-oriented (check, init, dashboard, agents, skills, mcp, stop, rm, status, watch, logs, follow, ps, projects) but descriptions vary widely in quality. Most descriptions are 50-200 chars and explain what the tool does, but lack guidance on WHEN to use it vs similar tools. Parameters have type definitions and descriptions, but many lack constraints (enums, ranges, patterns). No output schemas are documented. Error handling is minimal, tools return exit codes but provide no recovery guidance. Security concerns: the 'skills' tool has a --delete flag that is destructive, and 'stop' kills processes, but neither tool has confirmation patterns or detailed error messages. Composition is reasonable, each tool has one responsibility, but chaining is unclear (e.g., does 'ps' output IDs that 'stop' accepts?). Overall, definitions are functional but not LLM-optimized.
List agents declared in a rufler flow (name, type, role, prompt preview).
Verify node, claude code and ruflo are available.
Live TUI dashboard with runs, tasks, and log tail. Arrow keys to navigate runs. Selected run's tasks and log tail update in real-time. Press q to quit, r to force-refresh.
Live TUI dashboard with task progress, AI conversation, and system log. Tails all per-task NDJSON logs and renders a dashboard with four panels: tasks list, session stats, AI conversation stream, and system events. Replays existing log content for context, then streams live. Press q to quit.
Create a sample rufler_flow.yml in the current directory.
Tail recent autopilot events, or the raw NDJSON run log with --raw. With `-f` / `--follow` the run log is streamed live (Ctrl+C to stop). Pass a task sub-id (e.g. a1b2c3d4.01) to show only that task's log slice.
No output schemas documented. Tools return data but LLMs cannot predict structure (fields, types, pagination). Agents cannot plan downstream calls or extract required IDs.
Destructive tools (skills --delete, stop, rm) lack confirmation patterns or dry-run modes. Agents can accidentally delete data without recovery guidance.
Error handling is minimal. Tools exit with codes but provide no actionable recovery guidance. LLMs cannot self-correct or know whether to retry.
Parameter constraints missing. 'interval' (watch), 'last' (logs), 'grace' (stop), 'teardown_timeout' (stop) lack min/max bounds. LLMs can pass absurd values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
List MCP servers declared in rufler_flow.yml or registered with Claude Code.
List all projects ever run by rufler with last-run timestamps. This is the "project history" rollup — it survives `rufler rm` and `rufler ps --prune` because it's stored separately from individual runs.
Docker-style list of rufler runs. No args → currently running only (like `docker ps`). -a → everything ever run (like `docker ps -a`). <ID> → detailed view of one run: status, tasks, pids, log tail. --prune / --prune-older-than-days → clean up stale entries.
Remove rufler run entries from the central registry. This only deletes the registry bookkeeping — it does NOT touch log files, project directories or ruflo state. Use `rufler stop <id>` first if the run is still alive.
List skills installed under `<project>/.claude/skills/` plus the skills configuration declared in rufler_flow.yml. Use --available to see what's shipped inside ruflo's source tree (useful for picking `extra:` names). Use --delete to wipe installed skills (symlinks are preserved).
Show full ruflo status: system + swarm + hive-mind + autopilot.
Shutdown autopilot, hive-mind, daemon; SIGTERM the supervisor pid tree; record post-task outcome; end session. By default this DOES kill the supervisor and every descendant claude proc. Pass --no-kill to skip the kill step (e.g. when you only want the ruflo teardown side-effects but the run is already dead).
Poll `rufler status` on a loop until Ctrl-C.
Tool chaining unclear. 'ps' lists runs with IDs, but does 'stop' accept the same ID format? Does 'logs' accept task sub-IDs from 'ps' output? Undocumented dependencies.