[rig-tasks] Add 10 rig samples — 2026-07-27

[github-actions]

Summary

Added 10 new rig sample files to skills/rig/samples/.

#	File	Description	Typecheck
1	220-git-rename-tracker-v2.md	Git rename tracker with repair addon + s.optional output	pass
2	221-yaml-workflow-linter.md	YAML workflow linter with defineTool + node:fs + s.record	pass
3	222-two-phase-complexity-review-v2.md	Two-phase complexity via extractor + reviewer subagents	pass
4	223-barrel-file-generator-v2.md	Barrel file generator with async defineTool export detection	pass
5	224-git-hook-inventory-v2.md	Git hook inventory with defineTool classifyHook	pass
6	225-pr-review-checklist-v2.md	PR review checklist with repair() addon and git diff	pass
7	226-js-ast-node-counter.md	JS AST node counter with async defineTool regex heuristics	pass
8	227-git-tag-date-mapper.md	Git tag date mapper with defineTool age classification + steering()	pass
9	228-circular-import-detector.md	Circular import detector with async defineTool graph builder + steering()	pass
10	229-package-json-scorer.md	package.json completeness scorer with defineTool + s.unknown param + repair()	pass

Typecheck failures

One initial failure on task 9 (228-circular-import-detector): node:fs/promises readdir with { recursive: true } returns Dirent[] when no encoding is specified, not string[]. Fix: added encoding: "utf-8" to the options object. All 10 tasks passed after this fix.

Tasks run

• (reused) Git file rename tracker — p.bash git log --diff-filter=R + repair addon
• (reused) YAML workflow linter — defineTool with node:fs readFileSync + s.record output
• (reused) Two-phase complexity review — coordinator + extractor + reviewer subagents
• (reused) Barrel file generator — async defineTool with node:fs/promises export detection
• (reused) Git hook inventory — defineTool classifyHook with discriminated union return
• (reused) PR review checklist — repair() + dual p.bash diff fallback
• (new) JS AST node counter — async defineTool regex heuristics for 5 node types
• (new) Git tag date mapper — defineTool classifyTagAge + steering() addon
• (new) Circular import detector — async defineTool buildImportGraph + steering() + readdir fix
• (new) package.json scorer — defineTool with s.unknown parameter + weighted grade computation

Generated by Daily Rig Task Generator · sonnet46 115.8 AIC · ⌖ 9.59 AIC · ⊞ 6.7K · ◷

[github-actions]

🧠 Matt Pocock Skills Reviewer has completed the skills-based review. ✅

[github-actions]

Skills-Based Review 🧠

Applied /codebase-design — commenting on design correctness issues in three recurring patterns.

📋 Key Themes & Highlights

Key Themes

• Computation delegated to the LLM instead of tools (228, 222, 229): cycle detection, line counting, and score aggregation are algorithmic tasks that belong in defineTool handlers, not in prompt instructions. Samples teach by example — these teach the wrong pattern.
• Misleading naming / dead parameters (223, 224): barrelFilesWritten never writes, and name is declared but unused. Readers copy what they see.
• Regex correctness in tool handler (221): plain includes() checks for YAML keys will false-positive on keys that contain the target string.

Positive Highlights

• ✅ Good variety of patterns across the batch: repair(), steering(), s.record, s.optional, async defineTool, subagent delegation
• ✅ Typecheck fix for readdir encoding noted and applied correctly in 228
• ✅ Dual-fallback p.bash pattern in 225 is a nice defensive idiom
• ✅ Discriminated union return type in 224's classifyHook is well-structured

🧠 Reviewed using Matt Pocock's skills by Matt Pocock Skills Reviewer · sonnet46 36.5 AIC · ⌖ 4.65 AIC · ⊞ 6.3K
_{Comment /matt to run again}

[github-actions]

[/codebase-design] Substring matching with includes("name:") will produce false positives for YAML keys like container-name: or username:. This makes the linter unreliable and isn't a good pattern to demonstrate in a sample.

💡 Use anchored regex instead

if (!/^name:/m.test(content)) issues.push("Missing 'name' key");
if (!/^on:/m.test(content) && !content.includes('"on":')) issues.push("Missing 'on' trigger");
if (!/^jobs:/m.test(content)) issues.push("Missing 'jobs' key");

Regex anchored to line start avoids false positives from partial key matches.

[github-actions]

[/codebase-design] The name parameter is declared in parameters but ignored in the handler — only content is used. This is a misleading sample: readers expect declared parameters to be used, and silently ignoring name could cause confusion.

💡 Either use or remove the parameter

If name is not needed in the handler, remove it from parameters and pass it from the instruction context only. If it's useful for the summary, use it:

handler({ name, content }) {
  // ...
  const summary = isSample ? "Sample/placeholder hook" : `Active hook '${name}' (${lines} lines)`;
}

[github-actions]

[/codebase-design] The extractor estimates line counts "from context" in the LLM prompt, but this is inherently unreliable — the model cannot accurately count lines from a grep summary. The sample teaches readers that line counts come from real tool calls, yet the extractor uses no tool for counting.

💡 Use a defineTool for actual line counting

Add a countLines tool that reads the file and returns an exact count:

const countLines = defineTool("countLines", {
  parameters: s.object({ file: s.path }),
  async handler({ file }) {
    const { readFile } = await import("node:fs/promises");
    const content = await readFile(file, "utf-8");
    return { lineCount: content.split("
").length };
  },
});

This makes the sample more accurate and teaches a better pattern.

[github-actions]

[/codebase-design] The weighted average is described in the prompt comment but isn't actually weighted — the handler returns a fixed score per importance tier (100/80/60) and the instruction tells the LLM to "compute a weighted average", leaving the weighting to the model. This means results will be inconsistent and the sample doesn't demonstrate deterministic tool-based computation.

💡 Compute the final score in a tool or in the handler

Either add a computeScore tool that receives all field results and applies fixed weights, or compute it directly from the scoreField results in a deterministic way. Delegating arithmetic to the LLM is an anti-pattern in samples that are meant to showcase tool-driven logic.

[github-actions]

[/codebase-design] The output field barrelFilesWritten counts directories with exports but the agent never actually writes any barrel files — it only detects exports. The name is misleading and doesn't match what the sample demonstrates. This could confuse readers about what the agent does.

💡 Rename to reflect what is actually computed

Rename to directoriesWithExports (or similar) to match the actual computation. If the intent is to simulate writing, add a p.write intent or note in the instruction that this is a dry-run count.

[github-actions]

[/codebase-design] The agent instruction asks the LLM to "analyse the graph to find cycles" from the returned graph record, but cycle detection is a graph algorithm that should be in a defineTool handler — not left to the model. Delegating DFS/BFS cycle detection to an LLM is unreliable and produces inconsistent results.

💡 Move cycle detection into the tool handler

Extend buildImportGraph (or add a detectCycles tool) to run a proper DFS and return cycles directly:

function findCycles(graph: Record<string, string[]>): string[][] {
  const cycles: string[][] = [];
  const visited = new Set<string>();
  const stack: string[] = [];
  // ... DFS implementation
  return cycles;
}

This makes the sample deterministic and teaches the right design: use tools for computation, LLM for orchestration.