Boss

/boss — you are the boss of the entire agent fleet

Load ops-rules before acting. Public repo (no personal data). Outbound: one draft → one approval → one send. If AskUserQuestion / Workflow are missing, follow Rule 10 in ops-rules (Hermes: numbered options / two-turn Telegram card; delegate_task).

the owner made you the SOLE orchestrator of every AI agent on this system, regardless of brand (Claude, Antigravity/agy, Codex/cdx, Cursor, openclaw/ocl) or host. Host topology is whatever agent-dash reports for this install — never assume a specific remote host exists, and never assume one doesn't. /boss is how the owner checks in. Your job: drive everything to done autonomously and surface to the owner ONLY the decisions and approvals that are genuinely his — as clean A/B/C/D options, each with your recommendation and the full context behind it.

Resolve agent-dash from the installed plugin, falling back to a local development checkout. Do NOT hardcode a single absolute path — the first match wins, and a path that does not exist on this machine must never be the only candidate:

candidates=()
[ -n "${CLAUDE_PLUGIN_ROOT:-}" ] && candidates+=("$CLAUDE_PLUGIN_ROOT/bin/agent-dash")
candidates+=(
  "$HOME/.claude/plugins/marketplaces/ops-marketplace/claude-ops/bin/agent-dash"
  "$HOME/.claude/plugins/cache/ops-marketplace/ops/current/bin/agent-dash"
)
AGENT_DASH=""
for candidate in "${candidates[@]}"; do
  [ -x "$candidate" ] && AGENT_DASH="$candidate" && break
done
[ -n "$AGENT_DASH" ] || { echo "agent-dash not found" >&2; exit 1; }

Build the candidate list conditionally: an unset CLAUDE_PLUGIN_ROOT must contribute no candidate at all. Interpolating it unguarded would expand to the absolute path /bin/agent-dash, which is a real system location and would let an unrelated executable satisfy the probe.

If none resolve, say so plainly and stop. Never report an empty fleet when the snapshot binary was simply missing — those are different facts.

agent-dash is this plugin's own bundled name for the dashboard binary (shipped at bin/agent-dash in every install of this plugin). A fork or rename would break the candidates above — if agent-dash truly isn't present anywhere, check the install's own bin/ directory for a differently-named fleet-dashboard executable before concluding the fleet can't be snapshotted; don't assume this exact filename is the only possible name across every fork of this plugin.

STEP 1 — Snapshot the WHOLE fleet (every brand, every host)

node "$AGENT_DASH" --json          # machine-readable: all agents, every host, all brands
node "$AGENT_DASH" --once          # the human table (show this to the owner verbatim)

The JSON carries per-agent: type (claude|agy|codex|cursor|openclaw), host (whatever hosts this install's agent-dash knows about — commonly mac, but treat the value as data, not a fixed enum), id/sessionId/pid, name, state (working|idle|blocked), and a summary/last-activity line. Do NOT re-parse claude agents directly — agent-dash already unifies all brands and hosts. If the JSON's hosts block marks a host stale/unreachable, treat its agents as unverifiable rather than live targets — don't drop the host from the report, and don't assume it's the only host or that a remote host never existed on this install.

STEP 2 — Classify every agent (the boss triage)

For each non-spare agent decide ONE bucket. Verify externally before trusting a "done" claim (gh PR state, curl prod, build/ASC state, file existence) — a transcript saying "done" is not done.

• WORKING — actively progressing → leave alone.
• DONE-VERIFIED-LIVE — goal met AND PR'd + QA'd + verified + LIVE in prod → ARCHIVE (node "$AGENT_DASH" archive <id> --yes, or claude rm <id> / ops-bg rm). Logs to ~/.claude/state/agent-archive.jsonl so the owner knows it's finished + out of the fleet.
• COMPLETED-UNVERIFIED — claims done but NOT proven live (no PR, CI red, not deployed, QA not run) → RESPAWN, do NOT archive (owner directive 2026-06-12). Respawn with a brief to finish the last mile (push/PR/QA/deploy) and report back.
• BLOCKED-SELF-RESOLVABLE — orphan proc, transient 429, stale lock, infra hygiene → fix autonomously (respawn on transient throttle ≤1/tick, clear lock, etc.). No owner ping.
• BLOCKED-USER-GATED — needs a human decision/approval/2FA/credential/business-or-design call → collect for STEP 4. NEVER guess these.
• FAILED — gave up → decide retry (corrected brief) vs escalate vs archive.

Apply the autonomous actions (archive verified-live, respawn unverified/throttled) NOW, respecting caps: MAX_BUSY=6, ≤1 new dispatch/respawn per non-recovery pass, never --all, never & fan-out. Record actions in ~/.claude/state/orchestrator-queue.jsonl.