anti-hall-devswarm

---

name: anti-hall-devswarm description: Explain anti-hall's optional DevSwarm integration from Codex — the hivecontrol reference KB, the (designed, unbuilt) workspace-tier orchestration, and the shipped layered recovery model (Claude-side companion: child self-report → supervisor poke → escalate-to-parent, automatic path NEVER kills, plus the on-demand devswarm-recover CLI — the only kill path). Use when the user asks about DevSwarm, hivecontrol, or the anti-hall liveness supervisor while working in Codex.

anti-hall DevSwarm integration (Codex view)

anti-hall's DevSwarm support is optional and feature-detected, same model as the OMC/OMX integration: dormant, zero effect, unless DevSwarm is actually in use. anti-hall ships only generic pieces; any DevSwarm-consumer glue (inbox daemon, done-report contract, hivecontrol wiring) is the downstream project's, not this plugin's.

DevSwarm itself is agent-agnostic — DEVSWARM_AI_AGENT names the active in-workspace agent (claude, codex, …), so a Codex session can be running inside a DevSwarm workspace too. What follows is what a Codex user needs to know about the four addons.

The four addons

Addon	Status	Where
hivecontrol reference KB	Reference doc	docs/KB-devswarm-hivecontrol.md — the hivecontrol CLI surface, .devswarm/config.json schema, DEVSWARM_* env vars, and the parent/child async message-passing model. Includes a parallel OMC/OMX table for the workspace tier (§8.4 of the KB): the workspace-create/merge surface is CLI-identical either way; only the in-workspace fan-out engine differs (Workflow tool + subagents for Claude, omx team/workers for Codex). Repo-clone-only — does not ship with a plugin install.
Workspace-tier orchestration	Designed, NOT built	docs/superpowers/specs/2026-07-05-devswarm-orchestration-design.md + docs/superpowers/plans/2026-07-06-devswarm-orchestration.md. Nothing from this plan exists in the repo yet (no devswarm-guard.js/devswarm-children.js) — do not describe it as active in either OMC or OMX.
Liveness supervisor (detect → poke → escalate, never kills)	Shipped — Claude-only	companion/devswarm-supervisor.js + install-devswarm-supervisor.js + companion/lib/{liveness,recovery,target-session,doctor-devswarm}.js + hooks/lib/devswarm-detect.js + hooks/lib/devswarm-role.js + hooks/devswarm-child-role.js. The automatic background sweep. Recovers wedged claude sessions specifically (workaround for claude-code#39755) — it identity-binds to claude -p --resume <uuid> processes by argv, not Codex sessions. This section explains it for awareness; it is not a Codex-side capability.
On-demand recovery CLI (the ONLY kill path)	Shipped — targets claude processes only	companion/devswarm-recover.js, invoked explicitly per workspace id. Still Claude-only in what it targets (see below), but a Codex-side operator can run it.

The layered recovery model, for Codex users

Why it exists: a claude session inside a DevSwarm child workspace can go wedged — process alive, listener dead, no crash for a restarter to catch, no timeout event to generate a new turn (anthropics/claude-code#39755, related #28482/#33949). The automatic path handles this in three escalating layers and never kills anything:

1. Child self-report — the child workspace's own Claude-side SessionStart hook (hooks/devswarm-child-role.js) reminds an idle child to proactively message its parent via hivecontrol workspace message-parent. Cooperative only — doesn't help a truly wedged child, which is what the next two layers are for. Codex has no equivalent hook; this only fires for a Claude session with DEVSWARM_SOURCE_BRANCH set.
2. Supervisor poke — the background sweep detects a stale workspace (both outbound signals idle past threshold, plus a pending unread backlog) and fires an optional descriptor-supplied nudgeCommand, persisting verdict nudged.
3. Escalate-to-parent — once the poke budget is exhausted, the sweep persists verdict escalated (terminal — never re-targeted) and fires an optional escalateCommand.

If your DevSwarm setup mixes Claude and Codex workspaces (DEVSWARM_AI_AGENT differs per workspace), all of the above only ever touches the claude ones — it identity-binds to a claude process's argv session id and cwd. A Codex workspace is simply outside its target surface; nothing here signals or resumes a Codex process.

command-guard's destructive-read redirect (shipped, v0.53.0 — applies to Codex too)

Unlike the recovery model above, this one is not Claude-only: command-guard.js is a single file registered in both codex/hooks/hooks.json and the Claude hooks.json, so its DevSwarm branch fires identically for a Codex workspace. Under a DevSwarm-active session it redirects the two CONSUMING native hivecontrol inbox reads, in ALL contexts (a delegated read drains the queue identically): hivecontrol workspace monitor blocks UNCONDITIONALLY (a no-timeout long-poll that hangs the shell and consumes the queue); hivecontrol workspace read-messages blocks ONLY when durable-inbox evidence exists — ANTIHALL_DEVSWARM_INBOX_CMD set, or a ~/.anti-hall/devswarm/workspaces/*.json descriptor with a truthy inboxPath — so a harmless single-consumer read-messages is still allowed. Non-destructive message-count/message-parent/message-child are untouched. Own devswarm-read-guard skip name (in skip-guard's DESTRUCTIVE set — a blanket all skip does not cover it). Full detail: docs/KB-devswarm-hivecontrol.md §8.5.

Child-side reception — devswarm.js inbox pull (shipped, v0.54.2)

Since the native reads are guard-redirected, the safe way a child RECEIVES parent messages is node scripts/devswarm.js inbox pull <DEVSWARM_BUILDER_ID> — a bounded, guard-safe one-shot drain (command-guard's root-anchored LIGHT_EXCEPTION for scripts/devswarm.js allows it inline; the CLI wrapper is invokable from either agent). It auto-ensures the descriptor, then runs ONE pull: a non-destructive message-count gate first (count 0 → never calls read-messages), and only on count >0 a single bounded read-messages (finite 10 s timeout, never monitor), appended to the durable inbox NDJSON in one atomic, hash-idempotent write plus a store-parity feed. Residual limitations (honest): (1) read-messages marks-read BEFORE the durable append, so a crash in that window loses the native messages — the count-gate minimizes but cannot close it (hivecontrol has no non-destructive full read); a failed append surfaces ok:false, no partial NDJSON. (2) Pull-not-push: reception latency = one turn (no background child drainer). The drain module (companion/lib/devswarm-pull.js) and the durable inbox are Claude-side companion state, but the CLI itself runs identically either way. Full detail: docs/KB-devswarm-hivecontrol.md (v0.54.2 note).

On-demand recovery — devswarm-recover CLI

# SKILL_FILE = the absolute path Codex showed you for this SKILL.md.
ANTI_HALL_ROOT="$(cd "$(dirname "$SKILL_FILE")/../../.." && pwd)"
test -f "$ANTI_HALL_ROOT/.codex-plugin/plugin.json" || { echo "anti-hall plugin root not found relative to $SKILL_FILE — aborting" >&2; exit 1; }
node "$ANTI_HALL_ROOT/companion/devswarm-recover.js" <workspace-id>

This is the only place in DevSwarm that ever kills a process — the automatic sweep above never does. It's Claude-side tooling (it identity-binds to and resumes a claude process), but a Codex-side operator in a mixed OMC/OMX DevSwarm setup can run it directly against a claude workspace's id — there's nothing Claude-only about invoking the script itself, only about what it targets. Naming the id on the command line is a deliberate override: unlike the automatic sweep, this CLI will also target an interactive claude session (not just headless), under the same confirm-gate safety (exactly-one-or-abstain, identity + cwd re-confirmed immediately before every signal, single-writer lock, a recovery cap before it escalates instead). Windows: escalate-only, never kills. Full detail (safety invariants, resume guardrail, env) lives in plugins/anti-hall/skills/devswarm/SKILL.md (the Claude mirror of this skill).

Activation (Claude-side; for awareness)

If you're the one wiring up a mixed OMC/OMX DevSwarm setup and need the automatic supervisor active for the Claude side, the install command and full activation contract (per-workspace descriptor at ~/.anti-hall/devswarm/workspaces/<id>.json — id, worktreePath, sessionId required; inboxPath/cursorPath load-bearing; optional nudgeCommand/escalateCommand argv arrays for Layers 2–3; env gates DISABLE_ANTIHALL_DEVSWARM=1 / ANTIHALL_DEVSWARM_SUPERVISOR; the ANTIHALL_DEVSWARM_NUDGE_* tuning vars) live in plugins/anti-hall/skills/devswarm/SKILL.md. The update skill on the Claude side autonomously installs/refreshes this supervisor whenever it detects it's running inside an active DevSwarm session — no separate step needed once that side is set up. As of 0.54.1, the same autonomous-refresh posture also covers companion/install-devswarm-ingest.js — the one supervised daemon wrapping hivecontrol workspace monitor into the substrate store (companion/devswarm-ingest.js, shipped 0.54.0). It runs continuously (macOS LaunchAgent KeepAlive / Linux systemd --user Restart=always .service, cron fallback with a ~60s worst-case revive gap on a cron-only Linux host), distinct label/log from the supervisor, and is Claude-side tooling like the rest of this section — for awareness on a mixed OMC/OMX setup, not a Codex-side capability.

Outputs to watch either way: ~/.anti-hall/devswarm/liveness/<id>.json (per-workspace verdict — alive/stale/nudged/ambiguous/escalated), ~/.anti-hall/devswarm/recovery.log (append-only attempt/poke/escalate log), and node hooks/doctor.js (silent unless DevSwarm is active; nudged reports as WARN, not a failure).

Anti-hall + OMX workflow mapping

• Workspace tier (unbuilt): would be CLI-identical between agents — hivecontrol workspace create … -a codex vs … -a claude — only the in-workspace fan-out differs (omx team/workers vs Workflow tool + subagents).
• Layered recovery model + on-demand CLI: Claude-only in what they target, as above. No Codex-side equivalent exists in this plugin, though the CLI is invokable from either side.

Anti-hall does not replace OMX. Anti-hall supplies verify-first policy, the reference KB, and (for the Claude side) the layered recovery model; OMX supplies Codex-native orchestration/workflow runtime.