القائمة
Use BEFORE bulk/iterative AO spawns (>3). Hard cap 20 active workers; spawn ≤5 at a time; never use loadavg as the gate.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Token-efficient second opinion slash command /advice. Extracts decision point + artifact (≤150 lines), then fans out in parallel: (1) Opus subagent reviewer with fallback chain codex→agy→cursor, (2) /research on the decision topic, (3) /secondo multi-model opinion. Use instead of advisor() which ships the full conversation uncached.
Lazy senior-dev mode — the seven-rung ladder that decides whether to write code at all, whether to reuse code that already exists, and whether to prefer stdlib/platform/installed deps over a new dependency. Use before writing any code, every PR diff, every fix, every feature. Source attribution included.
Use when making any "X is enabled" / "feature X works" / "cache works" claim in a running session — requires runtime probe output from the standard harness startup path, NOT a launchd/cron test that explicitly sets the activation env var.
Pre-flight checklist for self-hosted runner failures. Use when investigating low disk, missing runner, or stuck Green Gate on the jleechanorg/worldarchitect.ai fleet. Always verifies host-level container state, not just GitHub API.
When asked to verify whether a Claude Code feature works (especially slash commands, dialogs, pickers, status indicators), spawn a real interactive TUI session in cmux — never use `claude --print "/feature"` as a test, because --print is non-interactive and will always return "isn't available in this environment" regardless of whether the feature actually works.
Read ~/.hermes/agent-orchestrator.yaml BEFORE declaring an --agent <X> value unsupported. Lists plugin shorthands (wafer, minimax, agy).
| name | ao-spawn-safety |
| description | Use BEFORE bulk/iterative AO spawns (>3). Hard cap 20 active workers; spawn ≤5 at a time; never use loadavg as the gate. |
Scope: AO worker spawning only. Do NOT apply these checks to Hermes's own incoming message handling — that blocks Hermes from responding when AO workers are running.
Before issuing any bulk or iterative AO spawn sequence (more than 3 spawns in a loop or in response to an open-ended instruction like "do all" / "take it all the way"):
ao list shows ≥20 active sessions total, decline to spawn and report the count instead. This is the single spawn limit — there is no lower "soft pause" threshold. The cap is on AO worker count, not system load average — do NOT use sysctl vm.loadavg as the spawn gate.Why: 2026-05-15 incident — "take it all the way" spawned 517 sessions, loadavg=205, DNS starved, gateway down 2h. The 20-worker hard cap is the guard. (Earlier revisions also carried a soft "pause and ask at 8" rung tied to a kanban.max_spawn config that was never actually set in config.yaml or agent-orchestrator.yaml — removed 2026-06-12 as redundant friction; 20 is the only cap.)
backfillAllPRs: true spawns via backfillUncoveredPRs each health-cron cycle (every 3600s by default). Its cap is project.spawnQueue.maxActiveSessions — default 20 (independent of the interactive 20-worker cap above).
Always set this companion when enabling backfill:
backfillAllPRs: true
spawnQueue:
enabled: true
maxActiveSessions: 5 # prevents zombie accumulation from dead-tmux sessions
maxConcurrentSessions is NOT a valid config field and is silently ignored by Zod. Using it instead of spawnQueue.maxActiveSessions has no effect.
Why 9 zombies accumulated (2026-06-18): Dead tmux sessions → hasSession()=false → PR marked "uncovered" → new worker spawned each cycle. Without a per-project cap (or with a silently-ignored field), the default 20-cap was never hit and accumulation continued unchecked.