metaharness 0.1.0

package/templates/vertical_agentics/package.json.tmpl

@@ -0,0 +1,30 @@
+{
+  "name": "{{name}}",
+  "version": "0.1.0",
+  "description": "{{description}}",
+  "type": "module",
+  "bin": {
+    "{{name}}": "./bin/{{name}}.js"
+  },
+  "files": ["bin/**", "dist/**", ".claude/**", "CLAUDE.md", "README.md", "LICENSE"],
+  "scripts": {
+    "build": "tsc",
+    "test": "vitest run --passWithNoTests",
+    "init": "node ./dist/init.js"
+  },
+  "dependencies": {
+    "@ruflo/kernel": "^0.1.0",
+    "@ruflo/host-{{host}}": "^0.1.0"
+  },
+  "devDependencies": {
+    "@types/node": "^20.0.0",
+    "typescript": "^5.4.0",
+    "vitest": "^2.0.0"
+  },
+  "engines": {
+    "node": ">=20.0.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}

package/templates/vertical_agentics/src/agents/critic.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Critic agent — Reviews outputs before they land.
+
+export const SYSTEM_PROMPT = `You review worker outputs against the task's success criteria before they are accepted into shared state. Reject work that is plausible but wrong, and say exactly why. You are the swarm's quality gate — without you, errors compound across agents. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'critic';
+export const TIER = 'opus' as const;

package/templates/vertical_agentics/src/agents/orchestrator.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Orchestrator agent — Routes work and owns the goal state.
+
+export const SYSTEM_PROMPT = `You own the goal. Decompose it, dispatch sub-tasks to workers over the swarm bus, and hold the shared state of what is done, blocked, and in flight. You route by capability and re-plan when a worker fails rather than restarting. You do the work of coordination, not the tasks themselves. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'orchestrator';
+export const TIER = 'opus' as const;

package/templates/vertical_agentics/src/agents/planner.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Planner agent — Builds the dependency-aware plan.
+
+export const SYSTEM_PROMPT = `You turn the goal into a dependency-aware plan: tasks, their preconditions and effects, and the order that respects dependencies. You expose the critical path and the tasks that can run in parallel. You replan from the current state on failure — never from scratch. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'planner';
+export const TIER = 'opus' as const;

package/templates/vertical_agentics/src/agents/worker.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Worker agent — Executes one task and reports.
+
+export const SYSTEM_PROMPT = `You execute exactly one assigned task, write the result and any new facts to shared memory, and report success or a precise failure to the orchestrator. You stay in your lane: you do not re-plan or grab another task. A crisp failure report is more useful than a heroic overreach. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'worker';
+export const TIER = 'sonnet' as const;

package/templates/vertical_agentics/src/init.ts.tmpl

@@ -0,0 +1,21 @@
+// SPDX-License-Identifier: MIT
+// Generated by create-agent-harness — your harness's `{{name}} init` entry.
+
+import { loadKernel } from '@ruflo/kernel';
+import adapter from '@ruflo/host-{{host}}';
+
+const HARNESS_NAME = '{{name}}';
+
+async function main(): Promise<number> {
+  const kernel = await loadKernel();
+  const info = kernel.kernelInfo();
+  console.log(`${HARNESS_NAME} — kernel ${info.version} (${kernel.backend})`);
+  console.log(`Host adapter: ${adapter.name}`);
+  console.log(`Run \`${HARNESS_NAME} doctor\` to verify the install.`);
+  return 0;
+}
+
+main().then(c => process.exit(c)).catch(err => {
+  console.error(err);
+  process.exit(1);
+});

package/templates/vertical_ai/.claude/commands/doctor.md.tmpl

@@ -0,0 +1,12 @@
+---
+description: "Health-check the harness: kernel load, MCP wiring, memory backend, host adapter."
+---
+
+Run a full health check and print a PASS/FAIL table.
+
+1. Kernel loads and `kernelInfo().version` matches package.json.
+2. The MCP server starts and lists its tools.
+3. The memory backend is reachable.
+4. The configured host adapter is present.
+
+Exit non-zero if any check fails.

package/templates/vertical_ai/.claude/settings.json.tmpl

@@ -0,0 +1,35 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(npx {{name}}*)",
+      "mcp__{{name}}__*",
+      "mcp__experiments__*",
+      "Bash(python *)"
+    ],
+    "deny": [
+      "Read(./.env)",
+      "Read(./.env.*)",
+      "Bash(rm -rf*)"
+    ]
+  },
+  "mcpServers": {
+    "{{name}}": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "{{name}}@latest",
+        "mcp",
+        "start"
+      ]
+    },
+    "experiments": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "{{name}}@latest",
+        "mcp",
+        "experiments"
+      ]
+    }
+  }
+}

package/templates/vertical_ai/.claude/skills/eval-report/SKILL.md.tmpl

@@ -0,0 +1,13 @@
+---
+name: eval-report
+description: "Produce an honest eval report: metrics, subgroup slices, baseline delta, ship/no-ship."
+---
+
+# eval-report
+
+Produce an evaluation report.
+
+1. Evaluate on the held-out set with objective-aligned metrics.
+2. Slice by subgroup and report the worst slice.
+3. Compare against the baseline; show the delta.
+4. End with SHIP or NO-SHIP and the number behind it.

package/templates/vertical_ai/CLAUDE.md.tmpl

@@ -0,0 +1,31 @@
+# {{name}}
+
+{{description}}
+
+> AI / ML Engineering harness · domain: `ai/ml-lifecycle`. Generated with [create-agent-harness](https://github.com/ruvnet/agent-harness-generator).
+
+## Behavioral rules
+
+- Use the harness's MCP tools (`mcp__{{name}}__*`) for orchestration
+- Memory and routing are handled by the kernel — you don't need to learn them
+- Defer destructive operations to the user
+
+## Agents
+
+| Agent | Tier | Role |
+|---|---|---|
+| `data-curator` | sonnet | Builds and documents the dataset. |
+| `trainer` | sonnet | Runs reproducible training jobs. |
+| `evaluator` | opus | The honest eval gate. |
+| `deployer` | sonnet | Ships behind a guardrail. |
+## Skills
+
+- `/eval-report` — Produce an honest eval report: metrics, subgroup slices, baseline delta, ship/no-ship.
+
+## Commands
+
+- `doctor` — Health-check the harness: kernel load, MCP wiring, memory backend, host adapter.
+
+## Architecture
+
+This harness uses [@ruflo/kernel](https://www.npmjs.com/package/@ruflo/kernel) — a Rust-compiled WASM module with a NAPI-RS native fallback — so the same code runs identically on every platform.

package/templates/vertical_ai/README.md.tmpl

@@ -0,0 +1,30 @@
+# {{name}}
+
+{{description}}
+
+> **AI / ML Engineering** — Curate → train → evaluate → deploy, with an experiment-tracking MCP and eval gates.
+>
+> Generated with [`create-agent-harness`](https://github.com/ruvnet/agent-harness-generator). WASM kernel, multi-host support, witness-signed releases.
+
+## Install
+
+```bash
+npm install -g {{name}}
+{{name}} init
+{{name}} doctor
+```
+
+## Agents
+
+| Agent | Role |
+|---|---|
+| `data-curator` | Builds and documents the dataset. |
+| `trainer` | Runs reproducible training jobs. |
+| `evaluator` | The honest eval gate. |
+| `deployer` | Ships behind a guardrail. |
+
+This harness ships with the **{{host}}** adapter.
+
+## License
+
+MIT

package/templates/vertical_ai/manifest.json

@@ -0,0 +1,86 @@
+{
+  "id": "vertical:ai",
+  "description": "An ML pod — data-curator, trainer, evaluator, and deployer over the model lifecycle.",
+  "domain": "ai/ml-lifecycle",
+  "category": "Engineering",
+  "files": [
+    {
+      "src": "package.json.tmpl",
+      "dst": "package.json",
+      "render": true
+    },
+    {
+      "src": "CLAUDE.md.tmpl",
+      "dst": "CLAUDE.md",
+      "render": true
+    },
+    {
+      "src": "README.md.tmpl",
+      "dst": "README.md",
+      "render": true
+    },
+    {
+      "src": ".claude/settings.json.tmpl",
+      "dst": ".claude/settings.json",
+      "render": true
+    },
+    {
+      "src": "src/init.ts.tmpl",
+      "dst": "src/init.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/data-curator.ts.tmpl",
+      "dst": "src/agents/data-curator.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/trainer.ts.tmpl",
+      "dst": "src/agents/trainer.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/evaluator.ts.tmpl",
+      "dst": "src/agents/evaluator.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/deployer.ts.tmpl",
+      "dst": "src/agents/deployer.ts",
+      "render": true
+    },
+    {
+      "src": ".claude/skills/eval-report/SKILL.md.tmpl",
+      "dst": ".claude/skills/eval-report/SKILL.md",
+      "render": true
+    },
+    {
+      "src": ".claude/commands/doctor.md.tmpl",
+      "dst": ".claude/commands/doctor.md",
+      "render": true
+    }
+  ],
+  "vars": [
+    {
+      "name": "name",
+      "prompt": "Harness name (kebab-case)",
+      "validate": "^[a-z0-9-]+$"
+    },
+    {
+      "name": "description",
+      "prompt": "One-line description",
+      "default": "Curate data, train, evaluate, and deploy models"
+    },
+    {
+      "name": "host",
+      "prompt": "Host adapter",
+      "default": "claude-code",
+      "choices": [
+        "claude-code",
+        "codex",
+        "pi-dev",
+        "hermes"
+      ]
+    }
+  ]
+}

package/templates/vertical_ai/package.json.tmpl

@@ -0,0 +1,30 @@
+{
+  "name": "{{name}}",
+  "version": "0.1.0",
+  "description": "{{description}}",
+  "type": "module",
+  "bin": {
+    "{{name}}": "./bin/{{name}}.js"
+  },
+  "files": ["bin/**", "dist/**", ".claude/**", "CLAUDE.md", "README.md", "LICENSE"],
+  "scripts": {
+    "build": "tsc",
+    "test": "vitest run --passWithNoTests",
+    "init": "node ./dist/init.js"
+  },
+  "dependencies": {
+    "@ruflo/kernel": "^0.1.0",
+    "@ruflo/host-{{host}}": "^0.1.0"
+  },
+  "devDependencies": {
+    "@types/node": "^20.0.0",
+    "typescript": "^5.4.0",
+    "vitest": "^2.0.0"
+  },
+  "engines": {
+    "node": ">=20.0.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}

package/templates/vertical_ai/src/agents/data-curator.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Data Curator agent — Builds and documents the dataset.
+
+export const SYSTEM_PROMPT = `You curate the dataset: source it, clean it, split it without leakage, and document its provenance and biases in a datasheet. The split is sacred — any leakage between train and eval invalidates everything downstream. You flag class imbalance and distribution shift before training starts. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'data-curator';
+export const TIER = 'sonnet' as const;

package/templates/vertical_ai/src/agents/deployer.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Deployer agent — Ships behind a guardrail.
+
+export const SYSTEM_PROMPT = `You deploy only models that passed the evaluator. Ship behind a canary or shadow first, wire up monitoring for the eval metric in production, and define the rollback trigger before traffic arrives. A model with no monitoring is not deployed — it is abandoned. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'deployer';
+export const TIER = 'sonnet' as const;

package/templates/vertical_ai/src/agents/evaluator.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Evaluator agent — The honest eval gate.
+
+export const SYSTEM_PROMPT = `You are the eval gate. Evaluate on the held-out set with metrics that match the real objective, slice by subgroup to catch hidden failure, and compare against a real baseline. You report the number that matters, including where the model is worse. No model ships on a cherry-picked metric. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'evaluator';
+export const TIER = 'opus' as const;

package/templates/vertical_ai/src/agents/trainer.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Trainer agent — Runs reproducible training jobs.
+
+export const SYSTEM_PROMPT = `You run training jobs reproducibly: fixed seeds, logged hyperparameters, and every run tracked in the experiments MCP. You change one variable at a time so results are attributable. You report training/val curves and stop early on overfitting. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'trainer';
+export const TIER = 'sonnet' as const;

package/templates/vertical_ai/src/init.ts.tmpl

@@ -0,0 +1,21 @@
+// SPDX-License-Identifier: MIT
+// Generated by create-agent-harness — your harness's `{{name}} init` entry.
+
+import { loadKernel } from '@ruflo/kernel';
+import adapter from '@ruflo/host-{{host}}';
+
+const HARNESS_NAME = '{{name}}';
+
+async function main(): Promise<number> {
+  const kernel = await loadKernel();
+  const info = kernel.kernelInfo();
+  console.log(`${HARNESS_NAME} — kernel ${info.version} (${kernel.backend})`);
+  console.log(`Host adapter: ${adapter.name}`);
+  console.log(`Run \`${HARNESS_NAME} doctor\` to verify the install.`);
+  return 0;
+}
+
+main().then(c => process.exit(c)).catch(err => {
+  console.error(err);
+  process.exit(1);
+});

package/templates/vertical_business/.claude/commands/doctor.md.tmpl

@@ -0,0 +1,12 @@
+---
+description: "Health-check the harness: kernel load, MCP wiring, memory backend, host adapter."
+---
+
+Run a full health check and print a PASS/FAIL table.
+
+1. Kernel loads and `kernelInfo().version` matches package.json.
+2. The MCP server starts and lists its tools.
+3. The memory backend is reachable.
+4. The configured host adapter is present.
+
+Exit non-zero if any check fails.

package/templates/vertical_business/.claude/settings.json.tmpl

@@ -0,0 +1,33 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(npx {{name}}*)",
+      "mcp__{{name}}__*",
+      "mcp__metrics__*"
+    ],
+    "deny": [
+      "Read(./.env)",
+      "Read(./.env.*)"
+    ]
+  },
+  "mcpServers": {
+    "{{name}}": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "{{name}}@latest",
+        "mcp",
+        "start"
+      ]
+    },
+    "metrics": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "{{name}}@latest",
+        "mcp",
+        "metrics"
+      ]
+    }
+  }
+}

package/templates/vertical_business/.claude/skills/quarterly-plan/SKILL.md.tmpl

@@ -0,0 +1,13 @@
+---
+name: quarterly-plan
+description: "Build a quarterly plan: findings → strategy → owned action items tied to KPIs."
+---
+
+# quarterly-plan
+
+Build a quarterly plan.
+
+1. Analyst pulls KPIs and reports what moved.
+2. Strategist frames options and recommends one, tied to a metric.
+3. Ops-coordinator breaks it into owned, dated action items.
+4. Output a one-page plan: goal, bet, metrics, owners, risks.

package/templates/vertical_business/CLAUDE.md.tmpl

@@ -0,0 +1,30 @@
+# {{name}}
+
+{{description}}
+
+> Business Operations harness · domain: `business/strategy`. Generated with [create-agent-harness](https://github.com/ruvnet/agent-harness-generator).
+
+## Behavioral rules
+
+- Use the harness's MCP tools (`mcp__{{name}}__*`) for orchestration
+- Memory and routing are handled by the kernel — you don't need to learn them
+- Defer destructive operations to the user
+
+## Agents
+
+| Agent | Tier | Role |
+|---|---|---|
+| `analyst` | sonnet | Turns raw metrics into findings. |
+| `strategist` | opus | Chooses the bet and the trade-offs. |
+| `ops-coordinator` | sonnet | Turns the chosen bet into owned actions. |
+## Skills
+
+- `/quarterly-plan` — Build a quarterly plan: findings → strategy → owned action items tied to KPIs.
+
+## Commands
+
+- `doctor` — Health-check the harness: kernel load, MCP wiring, memory backend, host adapter.
+
+## Architecture
+
+This harness uses [@ruflo/kernel](https://www.npmjs.com/package/@ruflo/kernel) — a Rust-compiled WASM module with a NAPI-RS native fallback — so the same code runs identically on every platform.

package/templates/vertical_business/README.md.tmpl

@@ -0,0 +1,29 @@
+# {{name}}
+
+{{description}}
+
+> **Business Operations** — Analyst → strategist → ops-coordinator, with a metrics MCP for KPI grounding.
+>
+> Generated with [`create-agent-harness`](https://github.com/ruvnet/agent-harness-generator). WASM kernel, multi-host support, witness-signed releases.
+
+## Install
+
+```bash
+npm install -g {{name}}
+{{name}} init
+{{name}} doctor
+```
+
+## Agents
+
+| Agent | Role |
+|---|---|
+| `analyst` | Turns raw metrics into findings. |
+| `strategist` | Chooses the bet and the trade-offs. |
+| `ops-coordinator` | Turns the chosen bet into owned actions. |
+
+This harness ships with the **{{host}}** adapter.
+
+## License
+
+MIT

package/templates/vertical_business/manifest.json

@@ -0,0 +1,81 @@
+{
+  "id": "vertical:business",
+  "description": "A business pod — analyst, strategist, and ops-coordinator for plans, metrics, and execution.",
+  "domain": "business/strategy",
+  "category": "Business",
+  "files": [
+    {
+      "src": "package.json.tmpl",
+      "dst": "package.json",
+      "render": true
+    },
+    {
+      "src": "CLAUDE.md.tmpl",
+      "dst": "CLAUDE.md",
+      "render": true
+    },
+    {
+      "src": "README.md.tmpl",
+      "dst": "README.md",
+      "render": true
+    },
+    {
+      "src": ".claude/settings.json.tmpl",
+      "dst": ".claude/settings.json",
+      "render": true
+    },
+    {
+      "src": "src/init.ts.tmpl",
+      "dst": "src/init.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/analyst.ts.tmpl",
+      "dst": "src/agents/analyst.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/strategist.ts.tmpl",
+      "dst": "src/agents/strategist.ts",
+      "render": true
+    },
+    {
+      "src": "src/agents/ops-coordinator.ts.tmpl",
+      "dst": "src/agents/ops-coordinator.ts",
+      "render": true
+    },
+    {
+      "src": ".claude/skills/quarterly-plan/SKILL.md.tmpl",
+      "dst": ".claude/skills/quarterly-plan/SKILL.md",
+      "render": true
+    },
+    {
+      "src": ".claude/commands/doctor.md.tmpl",
+      "dst": ".claude/commands/doctor.md",
+      "render": true
+    }
+  ],
+  "vars": [
+    {
+      "name": "name",
+      "prompt": "Harness name (kebab-case)",
+      "validate": "^[a-z0-9-]+$"
+    },
+    {
+      "name": "description",
+      "prompt": "One-line description",
+      "default": "Analyse, strategise, and coordinate execution"
+    },
+    {
+      "name": "host",
+      "prompt": "Host adapter",
+      "default": "claude-code",
+      "choices": [
+        "claude-code",
+        "codex",
+        "pi-dev",
+        "hermes"
+      ]
+    }
+  ]
+}

package/templates/vertical_business/package.json.tmpl

@@ -0,0 +1,30 @@
+{
+  "name": "{{name}}",
+  "version": "0.1.0",
+  "description": "{{description}}",
+  "type": "module",
+  "bin": {
+    "{{name}}": "./bin/{{name}}.js"
+  },
+  "files": ["bin/**", "dist/**", ".claude/**", "CLAUDE.md", "README.md", "LICENSE"],
+  "scripts": {
+    "build": "tsc",
+    "test": "vitest run --passWithNoTests",
+    "init": "node ./dist/init.js"
+  },
+  "dependencies": {
+    "@ruflo/kernel": "^0.1.0",
+    "@ruflo/host-{{host}}": "^0.1.0"
+  },
+  "devDependencies": {
+    "@types/node": "^20.0.0",
+    "typescript": "^5.4.0",
+    "vitest": "^2.0.0"
+  },
+  "engines": {
+    "node": ">=20.0.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}

package/templates/vertical_business/src/agents/analyst.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Analyst agent — Turns raw metrics into findings.
+
+export const SYSTEM_PROMPT = `You are the analyst. Pull the relevant KPIs from the metrics MCP and turn them into findings: what moved, by how much, and the most likely driver. Quantify everything; flag where the data is too thin to conclude. You report; you do not decide strategy. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'analyst';
+export const TIER = 'sonnet' as const;

package/templates/vertical_business/src/agents/ops-coordinator.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Ops Coordinator agent — Turns the chosen bet into owned actions.
+
+export const SYSTEM_PROMPT = `You convert the chosen strategy into execution: concrete, owned, dated action items with a success metric each. You surface dependencies and the first thing that will go wrong. No action item ships without an owner and a date. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'ops-coordinator';
+export const TIER = 'sonnet' as const;

package/templates/vertical_business/src/agents/strategist.ts.tmpl

@@ -0,0 +1,7 @@
+// SPDX-License-Identifier: MIT
+// Strategist agent — Chooses the bet and the trade-offs.
+
+export const SYSTEM_PROMPT = `You set strategy from the analyst's findings. Frame two or three real options, name the trade-off each makes, and recommend one with the reasoning. Tie every recommendation to a metric it should move and a time horizon. Avoid generic advice — be specific to this business's numbers. You operate inside the {{name}} harness; defer destructive actions to the user.`;
+
+export const NAME = 'strategist';
+export const TIER = 'opus' as const;