Menu
Templates and risk-tier evolution policy for authoring new agents, skills, and hooks โ the 3 artifact kinds this harness produces (no "mcp" kind; this harness has no MCP). Used primarily by meta-agent when extending the harness. Auto-invocable.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Reference-only skill documenting where artifacts live, the 3-verdict vocabulary (pass/pass-with-notes/reject), and the Preserved-Invariants-vs-Changed-Behaviors contract for revising an existing artifact. Load before writing or revising any artifact under .fable/. Auto-invocable.
Show CONSTITUTION.md's content and invariant list on request, re-run the SHA-256 hash check against CONSTITUTION.sha256 on demand, and document the amendment procedure. Useful for Claude to consult proactively as well as for a user to invoke directly.
Merge-safe installer that makes FABLE-HARNESS's agents/skills/hooks available at Claude Code's user scope (~/.claude), so they work across every project, not just H:\FABLE-HARNESS. Manual-only โ never auto-invoked, since this touches shared, hard-to-reverse state outside this project.
Canonical model-tier reference and the Fable-5 manual-only enforcement protocol. Consult before writing or changing any model field (agent frontmatter, skill frontmatter, Task dispatch parameter, settings.json availableModels, env vars). Auto-invocable โ any agent or skill authoring a new artifact should load this first.
Highest-stakes planning skill using Fable-5. Manual-only โ Claude can never auto-invoke this; only a user typing /plan-deep can. Use for the small number of asks big enough to warrant Fable-level structure and thinking budget (major-scope architecture, cross-cutting rewrites, irreversible decisions).
Everyday, auto-invocable planning skill for routine feature/bug/refactor work. Sonnet-tier โ the non-Fable default. Use for most planning needs; escalate to /plan-deep (manual-only, Fable-5) only for the rare highest-stakes plan.
| name | agent-factory |
| description | Templates and risk-tier evolution policy for authoring new agents, skills, and hooks โ the 3 artifact kinds this harness produces (no "mcp" kind; this harness has no MCP). Used primarily by meta-agent when extending the harness. Auto-invocable. |
| model | sonnet |
FABLE-HARNESS authors exactly three kinds of artifact: agent, skill, hook. (Sibling-repo research references a fourth "mcp" kind โ not applicable here; CONSTITUTION.md N8 means this harness never registers an MCP server, so there is no mcp-config artifact kind to template.)
.claude/agents/<name>.md)---
name: <lowercase-hyphen-name>
description: <dense description including trigger conditions โ this drives delegation>
tools: Read, Write, Edit, Glob, Grep, Bash
model: sonnet # or haiku / opus per model-routing skill's tier table โ NEVER fable
---
Required: name, description. tools: should be an explicit allowlist built from Read/Write/Edit/Glob/Grep/Bash + Task dispatch only โ no MCP tool names anywhere (N8). Consult the model-routing skill's tier table before picking model:.
.claude/skills/<name>/SKILL.md)---
name: <name>
description: <dense description โ drives auto-invocation unless disable-model-invocation is set>
model: sonnet # optional; omit to inherit, or haiku/opus/sonnet per tier table
# disable-model-invocation: true # only for plan-deep and install-user-scope โ manual-only skills
---
Required: name, description. Add disable-model-invocation: true only when the skill must be unreachable by Claude's own judgment (currently: plan-deep, install-user-scope). context: fork, allowed-tools, argument-hint are optional per the skills reference.
.claude/settings.json hooks.<EventName>[], plus the script it points at under .claude/hooks/){
"matcher": "Bash",
"hooks": [
{ "type": "command", "command": "pwsh -File .claude/hooks/pretooluse-bash-safety.ps1" }
]
}
Hook scripts must: fail open on internal error (exit 0 on unexpected parse/exception โ N5), fail closed on a genuine policy violation (exit 2, never exit 1 โ N5's explicit gotcha callout), and โ if enforcing a whitelist (e.g. the Bash safety gate) โ normalize input (strip quotes, collapse whitespace, split on &&/|/;) before matching each subcommand independently against an explicit allow-set (N6).
| risk_class | evolution_policy |
|---|---|
| low | Auto-commit the new artifact once it passes schema/invariant validation โ no human gate required. |
| medium | Require a human review before the artifact is considered final; may be drafted and staged, but not promoted, without sign-off. |
| high | Require a human review before final; treat any ambiguity as a reason to escalate rather than assume. |
| critical | Require a human review before final, and prefer pairing it with an explicit decision record under memory/decisions/ regardless of outcome. |
meta-agent (or any authoring agent) must classify every new artifact's risk_class before deciding whether it can auto-commit. When in doubt, round up to the more conservative tier โ this mirrors the evolution-handoff propose/evaluate/commit/HITL contract pattern from sibling-repo research, reimplemented here without any MCP/TheEights dependency (N8): "auto-commit" and "require human review" are just local git-commit-or-don't decisions, not calls to an external evolution service.
No agent file ever sets model: fable. This is non-negotiable and is CONSTITUTION.md N2. If meta-agent or any other authoring flow is about to write model: fable into anything other than .claude/skills/plan-deep/SKILL.md, it must stop and cite Refused per N2: no agent file may ever set model: fable. See the model-routing skill for the full six-surface reasoning behind why this rule exists.
N2 (Fable manual-only โ the hard rule above), N7 (typed-agent provenance โ never scaffold a "general-purpose" stand-in when a typed producer is what's needed), N8 (no MCP โ hook/agent tool allowlists never include mcp__* names).