菜单
Templates for the 8 artifact kinds (agent, skill, command, hook, team, squad, rubric, mcp) across the 6 sibling projects. Location conventions, frontmatter, validators, default risk class.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Known drift, injection, and runaway patterns plus mitigations. Reference before logging a Smith anomaly or wiring a detector.
The 10 immutable invariants (N1..N10) that AgentSmith enforces across the AI ecosystem. Reference before any refusal, override, or constitutional citation.
How AgentSmith isolates rogue artifacts and agents. Detect, suspend, sandbox, file HITL, await ruling, reinstate or purge. Includes what must NEVER be quarantined.
Refusal-appeal protocol between AgentSmith and Hydra's venom gate. How to file false-positive or false-negative appeals, immortal-head review path, and required audit trail.
Shared conventions across the 6 sibling AI projects. Frontmatter, .claude layout, team and command schemas, hook patterns, MCP registration, artifact directories. The default-when-unsure reference.
The propose/evaluate/commit/HITL contract with TheEights for evolving artifacts. Maps risk_class to evolution policy and documents the API surface AgentSmith calls.
| name | agent-factory-recipes |
| description | Templates for the 8 artifact kinds (agent, skill, command, hook, team, squad, rubric, mcp) across the 6 sibling projects. Location conventions, frontmatter, validators, default risk class. |
| allowed-tools | Read, Glob, Write |
The 8 artifact kinds Smith manufactures, validates, and ships across the 6 sibling projects.
Sibling projects: Hydra, TheEights, ExecutiveSuite, MarketBliss, RLM-Creative, pair-programmer.
Common root: Each sibling project's repository root (see ecosystem).
| Kind | Default risk_class | Why |
|---|---|---|
| agent | high | Defines an actor; large blast radius. |
| skill | low | Pure reference; no execution authority by itself. |
| command | medium | Operator entry point; can be misused. |
| hook | high | Runs without explicit invocation; trust-on-first-use. |
| team | medium | Composition of agents; inherits their authority. |
| squad | high | Composition with routing + budget authority. |
| rubric | low | Judging surface; affects verdicts not actions. |
| mcp | critical | Adds new tool surface; capability expansion is dangerous. |
Override only via evolution-handoff with TheEights.
| Project | Location |
|---|---|
| Hydra | <root>\agents\<slug>.yaml |
| TheEights | <root>\agents\<slug>\agent.yaml |
| ExecutiveSuite | <root>\execs\<role>\agent.yaml |
| MarketBliss | <root>\analysts\<slug>.yaml |
| RLM-Creative | <root>\studio\agents\<slug>.yaml |
| pair-programmer | <root>\agents\<slug>.yaml |
Required frontmatter: name, role, model_tier, allowed_tools, scope, owner.
Body skeleton: ## Purpose / ## Inputs / ## Outputs / ## Tools / ## Refusal rules.
Validators: agent.schema.json, model_tier ∈ {opus, sonnet, haiku}, allowed_tools ⊆ project allowlist.
All projects: <root>\.claude\skills\<slug>\SKILL.md.
Required frontmatter: name, description, allowed-tools (optional).
Body: progressive-disclosure markdown. No execution logic.
Validators: frontmatter present, name == folder slug, no allowed-tools with elevated permissions unless risk_class >= medium.
All projects: <root>\.claude\commands\<name>.md (and <root>\.claude\commands\<ns>\<name>.md for namespaced).
Required frontmatter: description, argument-hint (optional), allowed-tools (optional).
Body: the prompt body; supports $ARGUMENTS.
Validators: command.schema, slash-name regex ^[a-z][a-z0-9-]*$.
Registered in <root>\.claude\settings.json under hooks.{PreToolUse|PostToolUse|Stop|...}.
Required entry fields: matcher, command, timeout (optional), description.
Body (the command itself): PowerShell or bash script in <root>\.claude\hooks\<name>.ps1.
Validators: hook command exists on disk, exit code semantics documented, no unsanitized $-interpolation in PowerShell.
| Project | Location |
|---|---|
| pair-programmer | <root>\teams\<slug>.yaml |
| ExecutiveSuite | <root>\teams\<slug>.yaml |
| All others | not used |
Required: name, members (list of agent slugs), stages, rubric, budget_cap.
Validators: members exist, rubric exists, stages reference known stage kinds.
| Project | Location |
|---|---|
| Hydra | <root>\squads\<slug>\squad.yaml |
Required: name, fingerprints (keywords), agents, router, hitl_gates, budget.
Validators: squad.schema, fingerprints non-empty, at least one HITL gate placement.
| Project | Location |
|---|---|
| pair-programmer | <root>\rubrics\<slug>.yaml |
| Hydra | <root>\rubrics\<slug>.yaml |
| TheEights | <root>\evolution\rubrics\<slug>.yaml |
Required: name, criteria (list with weight), scoring_scale, alignment (e.g., ISO/NIST tag).
Validators: weights sum to 1.0, criteria >=3.
All projects: entry in <root>\.mcp.json (or <root>\.claude\mcp.json).
Required entry: command, args, env (optional), description, capability_class.
Validators: capability_class ∈ {read_only, read_write, network, exec}; exec requires HITL per N6.
daemon/src/factory/templates/
agent.template.yaml
skill.template.md
command.template.md
hook.template.ps1
team.template.yaml
squad.template.yaml
rubric.template.yaml
mcp.template.json
The daemon stamps generated_by: AgentSmith, generated_at: <iso>, constitution_hash: <hex> into every output. Missing stamps fail validation.
cross-project-conventions (the shared schemas referenced above)evolution-handoff (default risk_class overrides)anomaly-signatures (schema-frontmatter-drift catches violations)