Run harness adapters (Claude Code, Codex, OpenCode) INSIDE isolated sandboxes via defineSandbox + withSandbox + a provider (localProcessSandbox / dockerSandbox). Covers declarative provisioning: createSecrets + secret/bearer, skills (agentSkill/gitSkill/mcpSkill/ fileSkill), plugins, instructions → canonical AGENTS.md + symlinks projected per harness; shallow-clone default with depth opt-out; serial/parallel setup callback over a persistent shell; snapshot-after-setup default with snapshotMaxAge TTL; defineWorkspace (git/setup/scripts/skills/secrets/ instructions/plugins), defineSandboxPolicy (allow/ask/deny), lifecycle/resume, the SandboxHandle (fs/git/process/ports), capability tokens, defineSandbox hooks (onFile/onFileCreate/onFileChange/onFileDelete/onReady/onError/ onDestroy) + fileEvents flag, chat middleware sandbox group (defineChatMiddleware sandbox hooks), the sandbox debug category, watchWorkspace as a low-level building block, and the file.changed / sandbox.file / claude-code.session-id events. Use whenever a harness adapter needs a sandbox or when building sandbox providers.

AI becomes provocative and mocks the opponent during gameplay

Codex-only team orchestration: run a named team of cooperating Codex workers with durable, script-managed state. MUST USE when the user asks Codex to create, run, coordinate, inspect, archive, or delete a team of agents/threads/sessions, or to work on something as a team in parallel. FIRST inspects the active tool surface (checking tool_search for deferred tools) and tells the user the route: native MultiAgentV2 agents (flat spawn_agent with task_name) when available, Codex App threads as the fallback, or a plain-subagent split when neither set exists. The main session is always the leader; members are defined by a concrete part, ownership area, or perspective - never a vague job role; a bundled cross-platform script writes the .omo/teams state plus an auto-generated member field manual. Use a team when the work is not perfectly isolated but parallelizing helps; use plain subagents when scope is perfectly isolated or the goal is ambiguous. Triggers: team mode, teammode, make a team, run as a team, team of agents, coordinate threads, parallel Codex threads, archive the team.

Long-running task loop using /loop (Claude Code dynamic mode) and ScheduleWakeup to re-enter with fresh context on each wake-up. Internally invokes harness-work through Agent. Trigger: long-running, loop, wake-up, autonomous. Do NOT load for: one-shot task execution, review, release, planning.

Manage SSOT, memory, and cross-tool memory search. Guardian of decisions.md and patterns.md. Use when user mentions memory, SSOT, decisions.md, patterns.md, merging, migration, SSOT promotion, sync memory, save learnings, memory search, harness-mem, past decisions, or record this. Do NOT load for: implementation work, reviews, ad-hoc notes, or in-session logging.

File cleanup and archiving. Tidies up bloated Plans.md, session-log.md, old logs, and state files. Trigger: /maintenance, cleanup, archive, organize, split session-log. Do NOT load for: implementation, review, release, new feature development.

HAR: Execute Plans.md tasks from single task to full parallel team run. Trigger: implement, execute, do everything, breezing, team run, parallel, composer, composer 2.5. Do NOT load for: planning, review, release, setup.

HAR: Sync Plans.md with implementation. Drift detect, marker update, retrospective. Trigger: sync-status, where am I, check progress. --snapshot for snapshots. Do NOT load for: planning, implementation, review, release.

HAR: Project init, tool setup, agent config, memory setup, skill mirror sync. Trigger: setup, init, new project, CI/Codex setup, harness-mem, mirror. Do NOT load for: implementation, review, release, planning.

HAR: Multi-angle code, plan, scope review. Security/quality check. Trigger: review, code review, plan review, scope analysis. Do NOT load for: implementation, new features, bugfix, setup, release.

Generic release automation for projects using Keep a Changelog + GitHub. Single confirmation gate then end-to-end automation: bump detection, CHANGELOG promotion, PR/main merge, tag, GitHub Release. Trigger: release, version bump, publish. Do NOT load for: implementation, review, planning, setup.

Generate a Progress Tracker HTML for non-engineer vibecoders to glance at session progress (cc:WIP / cc:TODO / cc:完了 counts, percentage, elapsed/estimated minutes, cost so far/estimate, drift alerts). Uses Plans.md as source of truth, renders a single-file HTML with auto-regeneration support. Use when user asks for progress overview, session status snapshot, dashboard, or says: progress tracker, 進捗確認, 進捗ボード, dashboard. Do NOT load for: actual implementation, code review, release work.

Generate a Plan Brief HTML for non-engineer vibecoders before implementation starts. Searches harness-mem (project-only) for relevant past decisions, patterns, and Plans archive entries, then renders a single-file HTML artifact summarizing understanding, options, risks, acceptance criteria, and confidence. Use when the user requests a planning preview, a non-engineer-friendly summary before approval, or says: plan brief, planning preview, 計画概要, 計画レビュー. Do NOT load for: actual implementation, code review, release work.

HAR: Codex-native long-running loop runner. Uses a real background runner that executes one ready batch per cycle through Breezing by default, with status/stop controls. Trigger: long-running, loop, autonomous, background, Codex. Do NOT load for: one-shot implementation, normal review, release.

Generate an Acceptance Demo HTML for non-engineer vibecoders right before ship/wait/reject decision. Reads back the acceptance_criteria that were stored as personal-preference.v1 by harness-plan-brief (joined by user_request_hash), then renders a single-file HTML showing each criterion as verified or unverified along with a ship/wait/reject recommendation. Use when the user asks for an acceptance review, wants to decide whether to ship a delivered task, or says: acceptance demo, accept demo, 受け入れ判断, 受入レビュー, ship/wait/reject 判定, 検収レビュー. Do NOT load for: implementation, code review, release work.

Extract recurring failure patterns from breezing orchestration logs and Judgment Ledger, emit failure-rule.v1 proposals with confidence scores. SSOT promotion to patterns.md or decisions.md is proposal-only — human-approval-required. Use when user mentions failure codifier, failure patterns, self-learning loop, codify failures, or failure-rule proposals. Do NOT load for: direct SSOT edits, auto-promotion, or implementation unrelated to failure analysis.

Configure and verify the Cursor backend for Claude Code Harness. Use when user invokes cursor:setup, wants Cursor as the local default implementation backend, or asks to check Cursor plugin/agent readiness. Distribution default remains opt-in; only local env/user settings are changed when explicitly requested.

Run a Cursor Composer review as an advisory second opinion while keeping the primary review verdict on the host brain. Use when user invokes cursor:review, asks Cursor to review, or wants composer to sanity-check a diff. Cursor never owns APPROVE/REQUEST_CHANGES.

Delegate a single write task to Cursor Composer via cursor-companion.sh inside an isolated worktree, then Lead-review the diff and cherry-pick. Use when user invokes cursor:do, says delegate to cursor, have composer write it, refactor with cursor, hand a file edit to Composer. Do NOT load for: planning, code review only, read-only investigation, or multi-task team runs (use breezing --cursor or cursor:ask instead).

Read-only delegate to cursor-agent (Composer) for questions, investigation, design discussion, and adversarial sanity checks. No worktree, no cherry-pick, no Lead diff review — cursor-agent is locked to ask mode and cannot write. Use when user says: ask cursor, cursor sanity check, get a second opinion, adversarial review, design discussion, investigate with cursor, cursor:ask. Do NOT load for: implementation, refactor, file edits, commit/push work, anything requiring write access (use cursor:do or breezing --cursor instead).