sase-agents-status

name	sase_agents_status
description	Report on currently-running SASE agents. Use when the user asks "what's running?", "agent status", "status report", or any question about live/background agents.

Before doing anything else, run this command to record that you are using this skill:

sase skill use sase_agents_status --reason "<one-line reason for using this skill>"

Quick reference for answering "what's running?" questions about SASE agents.

Primary command

sase agent list -j

This prints a stable-shape JSON array to stdout. Each row has: name, project, pid, model, provider, workspace_num, status, duration_seconds, started_at, approve, prompt_snippet, artifacts_dir, wait and runner slot context, parent_agent_name, and agent_family. Clan members appear as their own agent rows; the clan container itself is an ACE projection rather than a running process. Waiting rows include runner_slots_in_use and runner_slot_holders, which identify the agents keeping them parked.

An implicit global-cap waiter is reported as QUEUED; authored %wait(runners=N) threshold waits remain WAITING. runner_slot_queue_position and runner_slot_queue_size cover every live slot waiter in priority-then-request-FIFO admission order, including while the pool is full. Use the position on QUEUED rows when explaining what will start next; do not describe their stale dependency or time-wait marker fields as active blockers.

How to summarize

• Group rows by project and render a compact table with name, duration_seconds, provider, and a short prompt_snippet.
• If more than 10 agents are running, show the 10 most recent (they come first — the list is already sorted by start time, newest first) and say "N more".
• If the list is empty, report "no agents are running" plainly — do not fabricate rows.

Other useful forms

• sase agent list — pretty rich table (for direct terminal use, not machine consumption).
• sase agent list -a — include recently-completed DONE/FAILED agents.
• sase agent list -p <project> — filter by project name.
• sase agent show -n <name> — full detail panel (prompt, pid, artifacts dir, live-tail hint).
• sase agent kill -n <name> — SIGTERM an agent by name. No confirmation prompt; use with care.
• For completed agents' transcripts, use the /sase_chats skill — sase chat show --agent <name> resolves to the saved chat for any agent that has run.

Retrying an agent

For ordinary retries, locate the source with sase agent list -a -j or sase agent show -n <name>, prefer <artifacts_dir>/raw_xprompt.md as the prompt source, allocate a fresh name with sase.agent.names.allocate_retry_name("<name>"), then rewrite the prompt through sase.agent.retry_prompt.rewrite_retry_prompt_name(raw_prompt, retry_name) before launching with sase run "$rewritten_prompt". Confirm the new run with sase agent list -a -j or sase agent show -n <retry-name>. Do not use %id:!<name> from non-TUI surfaces for ordinary retries; reserve forced same-name reuse for explicitly approved reruns through code paths that call wipe_names_for_forced_reuse.

Artifacts directory

Each row's artifacts_dir is the on-disk state for that agent. Files you may read mid-run:

• live_reply.md — streaming buffer for the agent's in-flight response. Treat as draft; only quote with an "in-progress" caveat.
• agent_meta.json — static metadata, including chat_path (where the transcript will live once saved).
• workflow_state.json, prompt_step_*.json — per-step checkpoints, written once on each step transition. Stable to read while the agent is still running.
• done.json — only present after the agent has stopped. When present it is authoritative; it has outcome, response_path, and plan_path (if the agent submitted a plan).

workspace_num resolves to <parent-of-this-repo>/sase_<N>/ — the agent's ephemeral workspace clone. Pre-submission plan drafts live there as <workspace>/sase_plan_*.md; once the agent has run sase plan propose, the submitted copy is at ~/.sase/plans/<YYMM>/<descriptive>.md and won't be rewritten.

Stable vs streaming

Rule of thumb: if a file is the agent's in-flight response and the agent is RUNNING, treat it as draft. If the file is a checkpoint, submission, or per-step output, treat it as stable — safe to read and quote even while the agent runs.

When you use live state to answer, cite the artifact paths you read and label each piece of evidence as draft/live or stable/completed. For review or comparison questions across multiple active agents, do not treat the absence of a completed transcript as the absence of useful evidence.

Implementation notes

The JSON shape is stable — do not reorder keys or rename fields when summarizing. started_at is ISO 8601 with timezone. duration_seconds is an integer. workspace_num is null for home-project agents.