Fusion Package Scribe

Package Scribe

Experimental caveat

This skill is experimental and not yet stable. Behavior, structure, and outputs may change between versions.

When to use

Use when you need to systematically add or improve documentation across TypeScript packages — whether a full monorepo sweep or a single package.

Typical triggers:

• "document all packages"
• "improve docs for packages/utils/observable"
• "add TSDoc to all public exports in this package"
• "rewrite the README for this package"
• "run a documentation pass on the monorepo"
• "review the TSDoc quality in this package"

Implicit triggers:

• A package has public exports with missing or stub TSDoc
• A package README is outdated, inconsistent, or missing standard sections
• A documentation sweep is planned across multiple packages

When not to use

• Modifying runtime code (only doc comments and README files)
• Generating API reference sites or TypeDoc output
• Non-TypeScript languages (initial scope is TS/TSX only)
• Auto-merging PRs or bypassing review workflows
• Security vulnerability scanning or performance profiling
• Replacing existing CI-based doc generation pipelines

Required inputs

If required inputs are missing or ambiguous, ask before proceeding.

• Scope: monorepo root path or specific package path(s)
• Mode: single (one package) or sweep (multiple packages)

Conditional inputs

• Package filter: glob or list when processing a subset of packages
• README template path: repo-specific template if not using the built-in default (for example under .github/instructions/)
• Tracking issue: issue number for lifecycle updates (assign, status, close)
• Commit style: repo-specific conventional commit format (defaults to docs(<package>): )
• Batch size: number of packages per batch for sweep mode (defaults to 5)

Defaults

• Mode: single (if only one package path provided)
• Commit prefix: docs(<package-name>):
• Batch size: 5 packages per sweep batch
• README structure: built-in template from references/readme-template.md
• Review council: enabled (runs after every package)

Precedence and standards discovery

This skill discovers and follows the target repository's own standards:

1. Repository instructions — applicable files under .github/instructions/, CONTRIBUTING.md, contribute/
2. Tooling configuration — tsconfig.json, biome.json, .editorconfig
3. Companion skill — fusion-code-conventions for TSDoc rules, naming conventions, and intent quality (when installed)
4. Built-in defaults — references/tsdoc-checklist.md and references/readme-template.md

Repository-level standards always win. When no repo standards exist, the built-in defaults apply.

Agent modes

In single-agent runtimes, all three roles run inline sequentially.

Instructions

Step 1 — Discover repository standards

Before generating any documentation:

1. Search for repo-level documentation instructions:
   • applicable files under .github/instructions/
   • CONTRIBUTING.md, contribute/
2. Read tsconfig.json to understand module structure, path aliases, and strict mode settings
3. Read biome.json or equivalent linter config for style expectations
4. Check if fusion-code-conventions is available — if so, defer to its TSDoc rules from references/typescript.conventions.md
5. If no repo-level standards exist, use references/tsdoc-checklist.md as the quality baseline

Step 2 — Discover packages

1. For single mode: validate the provided package path exists and has TypeScript source files
2. For sweep mode:
   • Read the root package.json or workspace config (pnpm-workspace.yaml, lerna.json, turbo.json) to find all packages
   • Filter by any provided glob/list
   • Sort packages by estimated size (file count) for batch planning
   • Write the discovery summary to .tmp/scribe-discovery-<context>.md

Step 3 — Plan execution (sweep mode only)

Activate agents/orchestrator.agent.md (or run inline):

1. Group packages into batches of the configured batch size
2. Order batches: smaller packages (fewer source files) first to maximize early throughput
3. Estimate token budget per package: ~100 tokens per source file for reading, ~200 tokens per export for TSDoc generation
4. If a single package exceeds 60% of the estimated context window, flag it for special handling (barrel exports first, defer internal modules)
5. Write the execution plan to .tmp/scribe-plan-<context>.md

Step 4 — Process each package

Activate agents/documenter.agent.md (or run inline) for each package:

4a — Scan public API surface

1. Find the barrel export file (index.ts, index.tsx, or main field in package.json)
2. Trace all re-exports to identify the full public API surface
3. Categorize exports: functions, classes, types/interfaces, constants, hooks, enums
4. Prioritize: barrel exports first, then direct public exports, then internally-consumed-but-exported items

4b — Generate or improve TSDoc

For each public export:

1. Read the existing implementation to understand intent, parameters, return values, error paths, and side effects
2. Check for existing TSDoc — improve rather than replace when present
3. Apply the TSDoc checklist (repo standards or references/tsdoc-checklist.md):
   • Summary line: explain why and what problem it solves, not just what it does
   • @param for every parameter with meaningful descriptions
   • @returns for every non-void function
   • @template for every generic type parameter
   • @throws for meaningful error paths
   • @example for user-facing and non-trivial public APIs
   • @deprecated with replacement guidance when applicable
4. Flag and rewrite "name-echo" patterns (for example /** Gets the value. */ getValue())
5. Do not modify runtime code — only doc comments

4c — Rewrite or improve README

1. Read the existing README (if any) to preserve valuable content
2. Apply the README structure from repo instructions or references/readme-template.md:
   • Package name and description
   • Features / key exports
   • Installation
   • Usage with code examples
   • API reference (summary of key exports with links or inline docs)
   • Configuration (if applicable)
3. Ensure the README is useful to a developer discovering the package for the first time
4. Optimize for retrieval: use clear headings, keyword-rich descriptions, and concrete examples

Step 5 — Review council

Activate agents/reviewer.agent.md (or run inline) after each package:

1. Intent extraction — Does the TSDoc accurately describe what the code does and why? Flag any comment that merely restates the function/type name.
2. Code comprehension — Are complex algorithms, state machines, or side effects explained? Would a new developer understand the code from the docs alone?
3. User-facing quality — Is the README useful to someone discovering the package for the first time? Does it have working examples?
4. Retrieval fitness — Will the documentation produce good hits in RAG / semantic search? Are key terms present in headings and summaries?

The reviewer produces a pass/fail per criterion. Failures loop back to Step 4 for the specific package.

Step 6 — Commit

After the review council passes for a package:

1. Stage only documentation files (.ts/.tsx files for TSDoc changes, README.md)
2. Commit using the repo's conventional commit format, defaulting to:

   docs(<package-name>): improve TSDoc and README documentation
   

3. Do not push — leave that to the user or a PR workflow

Step 7 — Report

After all packages are processed:

1. Write a summary to .tmp/scribe-report-<context>.md:
   • Packages processed and status (pass/fail/skipped)
   • Total exports documented
   • Review council pass rates per criterion
   • Any packages flagged for manual review
   • Commits created
2. If a tracking issue was provided, update it with the summary

Expected output

• TSDoc comments on all public exports in processed packages
• Consistent README for each processed package
• One commit per package with documentation-only changes
• Summary report in .tmp/
• Tracking issue update (when issue number provided)

Safety & constraints

This skill is mutation-capable. Repository-local workflow instructions take precedence over inline guidance when they conflict.

• Only modify doc comments and README files — never touch runtime code
• Do not push commits; leave push decisions to the user
• Do not auto-merge PRs or bypass review workflows
• Do not invent API behavior — document only what the code actually does
• Do not delete or replace existing valuable documentation without preserving its content
• When a tracking issue is referenced, update it only after explicit confirmation
• Respect .gitignore and do not commit temporary files from .tmp/
• If the review council fails a package twice, flag it for manual review instead of looping indefinitely