zeus-methodology-bootstrap - SKILL.md Agent Skill

name: zeus-methodology-bootstrap description: Auto-loads the Zeus adversarial-debate methodology on session boot when the task involves multi-agent coordination, adversarial debate, contamination remediation, or any cycle ≥ 2 of structured pro/con dialogue. Triggers on compound-phrase keywords (per critic MED-CAVEAT-S4-1 fix to avoid over-firing on routine Zeus drift/verdict work): "adversarial debate", "critic-gate", "debate verdict", "verdict drift", "round verdict", "concession bank", "contamination remediation", "anti-drift methodology", "phase discipline", "longlast teammate", "cross-session merge", "stage-gated revert", "process gate", "5th outcome", "audit-first mode", "verdict erratum". Note: bare keywords "drift" and "verdict" intentionally REMOVED — empirical 138 commits in 30 days mention "drift" (data drift, signal drift, calibration drift) which would over-fire the auto-load. Returns the methodology doc as required reading + the 5-for-5 case study summary + the 5th outcome category from contamination remediation 2026-04-28.

Zeus Methodology Bootstrap

Purpose

Ensure the adversarial-debate methodology (docs/methodology/adversarial_debate_for_project_evaluation.md, ~785 lines) loads into the active session's context BEFORE any debate-class work begins. Prevents the cross-session knowledge loss pattern that contributed to the 2026-04-28 contamination event (other session did not know about §5.Z2 codified gates → bypassed bidirectional grep → produced 6 drift items including 815k mislabeled production rows).

When this SKILL fires

Auto-triggers on session boot when ANY of these hold:

Task description contains keywords listed in the description frontmatter
Active task is a Mode C (longlast multi-batch) or Mode D (adversarial debate) per .agents/skills/zeus-ai-handoff/SKILL.md §3
Operator references "another cycle", "next round", "the methodology", "critic gate", "stage-gated revert", "5th outcome"
Recent commits in active branch reference task_*_harness_debate/ or task_*_contamination_remediation/
New session resuming work after compaction in a debate-class context

What this SKILL provides

When fired, this SKILL ensures the agent reads:

Methodology doc: docs/methodology/adversarial_debate_for_project_evaluation.md (master text)
5-for-5 case study summary (the empirical track record — load these into immediate context):
- §5.X BATCH D INV-16/17: prescribed DELETE → audit found 9 hidden tests → REVERT + add tests:; first case study of "apparent gap may be intentional"
- §5.Y Phase 2 auto-gen registries: prescribed AUTO-GEN → audit found 95% intentional curation
- §5.Z Phase 3 module_manifest: prescribed REPLACE → audit 21 KEEP / 4 HYBRID / 0 REPLACE
- §5.Z2 Phase 4 @enforced_by: prototype works for uniform domain; HYBRID for cross-cutting; bounded confirmation
- §5.Z3.1 (NEW from cycle 5) Cycle 5 contamination remediation: 5th outcome category formalized; quantitative ≥3 errata/cycle audit-first trigger added
§5.Z3 4-outcome categories + 5th meta-level outcome:
- 1. Falsified — don't change; erratum upstream
- 1. Confirmed bounded — change at bounded scope with discipline
- 1. Confirmed unbounded — change at full scope (rare; multiple pass-gates required)
- 1. Inconclusive — defer; iterate on the audit
- 1. CONDITIONAL-REVERT-PENDING-OTHER-SESSION-COMPLETION (Stage-gated revert) — for already-contaminated state across session boundaries; revert+quarantine NOW + condition future restoration on independent critic gate + staged restoration as evidence accumulates
§5.Z2 codified gates (mandatory before locking ANY % claim):
- Bidirectional grep (forward: manifest cites field? + reverse: target system back-cites identity?)
- Intent-aware audit (does the code's stated intent match its observed behavior?)
- Default-deny on grep-only evidence
§5.Z3.1 quantitative erratum-frequency trigger: ≥3 errata/cycle → next cycle MUST start audit-first mode

Cross-references

.Codex/skills/multi-agent-debate-and-execution/SKILL.md (NEW v1 2026-05-03) — generic reusable EXECUTION + CLOSURE phase skill. Covers per-packet boot template, GO/DONE/REVIEW cycle, 3-batch decomposition, PATH A precision-favored framing, K3-adjacent surface pre-flag, HONEST DISCLOSURE pattern, 10-ATTACK template, co-tenant safe staging, cross-module orchestration seam. Embeds 5-packet 32-cycle Zeus R3 §1 #2 case study. Mirror at ~/.Codex/skills/multi-agent-debate-and-execution/SKILL.md for cross-project reuse. Methodology doc covers DEBATE phase; this new skill covers EXECUTION + CLOSURE.
.agents/skills/zeus-ai-handoff/SKILL.md (Mode C/D selection)
.Codex/skills/zeus-phase-discipline/SKILL.md (Mode C per-batch; 14 anti-drift mechanisms compressed)
Root AGENTS.md (Required Reads for adversarial work; §3 "What to read by task" → Adversarial debate / multi-agent / contamination remediation bullet)
docs/operations/task_2026-04-27_harness_debate/ (4-cycle case studies: R1+R2+R3+Tier 2)
docs/operations/task_2026-04-28_contamination_remediation/ (5th cycle + 5th outcome category instantiation)

Memory references (load if relevant to current task)

feedback_critic_prompt_adversarial_template — 10-attack template; never write "narrow scope self-validating" or "pattern proven"
feedback_executor_commit_boundary_gate — executor cannot self-approve over multi-batch work; independent critic mandatory
feedback_zeus_plan_citations_rot_fast — file:line citations rot ~20-30%/week; grep-verify within 10 min before lock
feedback_converged_results_to_disk — SendMessage drop pattern is empirical; disk is canonical
feedback_idle_only_bootstrap — spawn longlast teammates with idle-only boot prompt
feedback_no_git_add_all_with_cotenant — git add SPECIFIC files; never -A with co-tenant active
feedback_multi_angle_review_at_packet_close — 5 parallel sub-agents (architect/critic/explore/scientist/verifier) before declaring DEBATE_CLOSED

Maintenance

This SKILL is living. Update after each methodology cycle that surfaces new patterns or new outcome categories. Cite this SKILL in new methodology cycles' TOPIC.md as the boot context loader.

When promoting to OMC/global skill: copy to ~/.Codex/skills/zeus-methodology-bootstrap/SKILL.md and adjust paths.

Lineage

v1 (2026-04-28): created during contamination remediation Stage 4 Gate D. Replaces the implicit "methodology will be loaded if agent thinks to" pattern with explicit auto-load on debate-class tasks. Born from the 2026-04-28 contamination event where another session did not know §5.Z2 codified gates and bypassed bidirectional-grep audit, producing 6 drift items.