القائمة
Codex-native equivalent of anti-hall deadly-loop. Use to harden risky changes with repeated adversarial review and fix waves until no new P0/P1 blockers remain.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | anti-hall-deadly-loop |
| description | Codex-native equivalent of anti-hall deadly-loop. Use to harden risky changes with repeated adversarial review and fix waves until no new P0/P1 blockers remain. |
This is the Codex-native protocol. Do not run the Claude deadly-loop.workflow.js; Codex does not expose that Workflow runtime.
Use for:
Round structure:
docs/<date>-<topic>-session-handoff.md with current state.gpt-5.6-solgpt-5.6-solgpt-5.6-sol~/.anti-hall/approvals/<repo>@<HEAD-sha>.json — fields: repo, sha,
timestamp, roundTrend (NEW confirmed P0/P1 count per round), seatVerdicts,
and "proof": false. This file is an ADVISORY AUDIT RECORD, never proof of
correctness and never an authorization token. Deadly-loop is agent-followed
guidance, not a hook enforced by the harness — an agent that fakes convergence
can just as easily fake this file. Any reader that consumes it (a CI push gate,
DevSwarm's merge gate, etc.) MUST treat "proof": false as load-bearing and
enforce its own real gating rather than trusting this record as sufficient
evidence on its own.Each review must cite file/line evidence and distinguish:
NEW: newly found issueREDISCOVERED: known issue still presentRESOLVED: previous issue fixed with evidenceREFUTED: reported issue disproven with evidenceCodex model rule: debate and validation seats use gpt-5.6-sol. Do not use mini models for debate roles.
This roster is three gpt-5.6-sol seats — same-model, not the canonical
cross-model TRIO. MODEL-POLICY.md's availability fallback matrix (Claude
side) lists 3-of-one-family as the explicit worst-case last-resort row,
with the stated reason "same-model agents share systematic blind spots":
| Roster | Status |
|---|---|
| Claude-side TRIO (Sonnet 5 Reviewer + Opus Auditor + Codex Critic) | intended default |
This Codex-native roster (3× gpt-5.6-sol) | degraded / last-resort — Codex has no Workflow runtime to fan a round out across model families on its own |
Treat this file's all-gpt-5.6-sol roster as the Codex-only floor, not the goal.
For L-tier hard-risk work, use the opt-in escalation below to recover one
cross-model seat.
For L-tier hard-risk work only — security/auth/signing/redaction/prompt-injection,
schema/migration/production-data, shell scripts, CI/workflow YAML — optionally
shell out to the claude CLI for ONE seat (Reviewer is the recommended pick)
to get a genuinely different model's perspective:
| Aspect | Detail |
|---|---|
| Requirement | claude CLI installed and on PATH (command -v claude / where claude) |
| Invocation | claude -p "<REVIEWER_PROMPT>" — one-shot, non-interactive (-p/--print) |
| Scope | ONE seat only (Reviewer recommended); Auditor and Critic stay on gpt-5.6-sol |
| Requirement level | OPT-IN, not a hard requirement — if claude is unavailable, stay on the plain 3×gpt-5.6-sol roster and note the degradation |
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.
Explain and activate anti-hall's optional DevSwarm integration — the hivecontrol reference KB, the (designed, unbuilt) workspace-tier orchestration, and the shipped layered recovery model (child self-report → supervisor poke → escalate-to-parent, automatic path NEVER kills) plus the on-demand devswarm-recover CLI (the only path that ever kills). Use when the user asks "explain the anti-hall DevSwarm integration", "how do I activate the DevSwarm supervisor", "what DevSwarm addons does anti-hall have", "tune the liveness supervisor", "recover a stuck DevSwarm workspace", or anything about hivecontrol / DevSwarm workspaces from an anti-hall angle.
Check or update anti-hall from the local marketplace clone. Use when the user asks to update anti-hall, check whether anti-hall is current, or refresh the Codex port files.
Update anti-hall to the latest released version and show the changelog delta. Use when the user says "update anti-hall", "/anti-hall:update", "upgrade the plugin", "is anti-hall up to date", or "check for an anti-hall update". Fast-forward-pulls the marketplace clone, mirrors the new version into the plugin cache, prints what changed, then has the user reload in-session via /reload-plugins (rarely, a harness build may require a restart — the skill says so when relevant).
Brief the agent that installs or operates anti-hall (in Codex/OMX) on how the whole system works — a DERIVED, live enumeration of every installed hook, the shipped skills, the DevSwarm coordination substrate (mechanical triggers, store, CLI, auto-safe migration), and a docs/KB map. Generated from the real hooks.json and files on disk, never a hardcoded list, so it can't drift. Use when the user asks "brief me on anti-hall", "what's in this build", "how does the whole system work", or when onboarding an agent to install/run anti-hall. For "is it actually installed in Codex / do the guards fire", use anti-hall-doctor.
Brief the agent that installs or operates anti-hall on how the whole system works — a DERIVED, live enumeration of every installed hook (grouped by event, each with its own one-line purpose), the shipped skills, the DevSwarm coordination substrate (mechanical triggers, store, CLI, auto-safe migration), and a docs/KB map. Generated from the actual hooks.json and the files on disk, never a hardcoded list, so it can't drift. Use when the user asks "brief me on anti-hall", "what's in this build", "how does the whole system work", "give me the anti-hall system overview", "what hooks/skills/substrate ship here", or when onboarding an agent to install or run anti-hall. For "is it actually working / do the guards fire", use the `doctor` skill instead.