메뉴
Install or verify the implementer, reviewer, and saver agent profiles required by Herder. Use when setting up the Herder plugin in a repository, when Herder reports a missing Codex role, or when the user asks to refresh Herder's host-native agent definitions.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Execute, resume, inspect, or safely clean a validated herder-plans/ backlog as a dependency-aware multi-agent run with per-attempt token accounting. Use when the user asks to fire, run, resume, automatically complete Herder plans, clean Herder run worktrees or branches, or report a Fire run's status and token coverage. Do not use to create plans, repair plan formatting, or implement one ordinary task directly.
Survey a codebase as a senior advisor and write prioritized, self-contained Herder plans from verified repository findings without changing source code. Use when asked to audit code, find bugs or improvement opportunities, suggest evidence-backed product direction, review or reconcile existing plans, or produce a herder-plans/ backlog for $herder:fire. Route user-defined new features that require intent clarification to $herder:grill.
Initialize, validate, inspect, and manage Herder Markdown plan backlogs and execution-token ledgers. Use when creating or repairing herder-plans/, checking plan dependencies and readiness, changing plan tracking policy, inspecting execution status, reporting token coverage, or preparing plans for $herder:fire. Do not use to implement plans or orchestrate subagents.
Validate a Herder plan directory against the canonical mechanical and semantic contracts, report plan/index/dependency/drift issues without changing files, and conservatively repair safe issues with --fix. Use when the user invokes herder:validate, asks whether herder-plans/ is Fire-ready or executable, wants a cold plan-quality audit, or asks to repair malformed, incomplete, or drifted plans. Do not use to execute plans, modify source code, or decide missing product intent.
Interview the user one decision at a time, investigate repository facts, maintain domain-model and ADR obligations, and create one validated Herder implementation plan from confirmed intent. Also refine an existing TODO or decision-blocked plan when invoked with --plan. Use when the user invokes herder:grill, wants to clarify a feature or change before implementation, asks to turn product intent into a plan, or wants to grill or stress-test an existing plan before Herder Fire executes it.
| name | install |
| description | Install or verify the implementer, reviewer, and saver agent profiles required by Herder. Use when setting up the Herder plugin in a repository, when Herder reports a missing Codex role, or when the user asks to refresh Herder's host-native agent definitions. |
Install or verify the three host-native agent profiles required by Herder. The profiles are bundled with the installed plugin, so the plugin version is the single verified release unit and no runtime network fetch is required.
Codex Fire uses these profiles as native custom agent types. It requires Codex Multi-Agent V2 and never falls back to nested codex exec processes. Claude Fire uses the bundled host-native agents directly.
Interpret tokens after the skill name as installer arguments:
herder:install [--host codex|claude|all] [--scope project|user]
[--dry-run] [--force]
project.--force: replace differing installed profiles after preserving backups. Never add it unless the user explicitly asks to replace or refresh customized profiles.Codex users invoke $herder:install; Claude Code users invoke /herder:install.
SKILL.md.codex or claude. If the user requests both, use all.node <skill-dir>/scripts/install-herder.mjs \
--host <codex|claude|all> \
--scope <project|user> \
<remaining user arguments>
codex features enable multi_agent_v2 only with the user's authorization to change Codex configuration. Then, with the same authorization, replace the boolean feature entry with the namespaced configuration below. If the current release does not expose the flag, report that it is unsupported rather than suggesting an execution fallback.Codex project scope installs to <repo>/.codex/agents/; user scope installs to ~/.codex/agents/. Claude agents load directly from the plugin's agents/ directory, so the installer only verifies their bundled definitions.
$herder:fire requires a live Multi-Agent V2 spawn interface that accepts a custom agent_type. Codex's reserved generic spawn schema can hide custom-agent metadata, so Herder uses a dedicated namespace. Configure exactly one form of multi_agent_v2; replace multi_agent_v2 = true under [features] with:
[features.multi_agent_v2]
enabled = true
hide_spawn_agent_metadata = false
tool_namespace = "herder_agents"
Do not set legacy agents.max_threads with Multi-Agent V2. Use Fire's --max-parallel for scheduling; if a host-level limit is needed, use max_concurrent_threads_per_session inside the block above. The installer checks the effective feature state through the current Codex executable, but it cannot inspect a session's already-frozen tool schema. Installing profiles does not silently edit config.toml; it prints the required configuration and a feature-enable command when configuration is missing.
Install automatically when a target is absent or byte-identical. If a target differs, stop without changing any profile and show the conflict. Preserve user customization by default. Use --force only with explicit user authorization; the installer saves replaced files under <Codex config root>/.plan-herder-backups/<timestamp>/, alongside rather than inside agents/, before writing the new set. On any install or verification run, migrate the unsafe legacy agents/.herder-backups/ directory to a timestamped directory under that safe root; --dry-run reports the migration without changing files.
Do not execute profile contents. Read only files named in the bundled manifest, require paths to remain inside the expected plugin profile trees, and verify every SHA-256 digest before writing anything.