Context Bundler

Dependencies

This skill requires Python 3.8+ and standard library only. No external packages needed.

Context Bundler Skill 📦

Overview

This skill centralizes workflows for compiling codebase files, documentation, and instructions into portable payloads (.md for AI chat UIs or .zip for offline/agent handoffs).

Storage & Git Discipline: All bundle manifests, markdown payloads, and archives MUST be generated inside a gitignored temp/ subfolder (e.g., temp/bundles/ or temp/context-bundle-[name]/). Never write bundle outputs to project root or non-ignored directories, which would pollute git status with massive generated review packages.

context-bundler supports 4 Execution Modes:

1. Standard Bundle Mode: Custom interactive selection of files/directories for general review or context sharing.
2. Persona-Driven Review / Handoff Mode: Injects specialized review or delegation persona prompts (prompt.md) ahead of codebase files.
3. Monorepo Segmented Mode: Full monorepo context packaging partitioned by domain (/skills, /agents, /scripts, /docs).
4. Multi-Persona Fan-Out Mode: Packages the same target content against multiple persona templates in one pass, producing one bundle per persona for parallel adversarial dispatch. This is the mechanism graph-planning-superpowers-policy.md §2.2-2.3 refers to for Phase 1 plan review.

🎭 Persona Template Library (assets/templates/)

When bundling context for external models, sub-agents, or human reviews, select or recommend a template from assets/templates/:

1. Security & Quality Review Personas

• adversarial-security-auditor.md: OWASP Top 10, auth bypasses, injection vectors, exploit scenarios, CVSS risk ratings.
• refactoring-quality-specialist.md: DRY violations, code smells, cyclomatic complexity reduction, before/after code diffs.
• compliance-standards-reviewer.md: Project conventions, 20-line purpose headers, ADR compliance, type annotations.

2. Architecture & Design Planning Personas

• structural-architecture-reviewer.md: C4 model, SOLID principles, coupling, interface abstraction leaks, component boundaries.
• plan-critique-reviewer.md: Stress-tests implementation plans, unstated dependencies, execution friction, rollback mechanisms.

3. Agent Task Handoff & Documentation Personas

• agent-task-delegator.md: Builds turnkey task prompts for sub-agents (Copilot CLI, Claude Code, Gemini CLI) with explicit tool gates & test criteria.
• docs-synthesis-specialist.md: Generates ADRs, C4 Mermaid diagrams, README guides, and API specs from raw codebase context.

4. Graph Planning Phase 1 Fan-Out Trio

The canonical three-persona set for graph-planning-superpowers-policy.md §2.3's adversarial plan review — use all three together via Multi-Persona Fan-Out Mode, not individually:

• structural-architecture-reviewer.md — Architecture Skeptic role: interfaces, dependency cycles, missing contracts.
• adversarial-security-auditor.md — Security / Edge-Case Auditor role: injection, auth, failure paths, race conditions.
• tdd-contract-reviewer.md — TDD Contract Reviewer role: deterministic test fixtures and assertion validity.

(For CLI-dispatched review instead of a bundle-to-external-chat handoff, the equivalent cli-agents sub-agents are architect-review, security-auditor, and tdd-contract-reviewer.)

Core Workflow

Phase 1: Mode & Target Discovery

Evaluate the request and negotiate mode and format:

1. Mode: Standard Bundle, Persona-Driven Review/Handoff (select persona template), Monorepo Segmented, or Multi-Persona Fan-Out (select persona set — see below).
2. Format: Single Markdown payload (.md) or Portable ZIP archive (.zip).
3. Targets: Directories or file paths to package.

Phase 2: Recap & Pre-Execution Confirmation

Present execution plan to user before running scripts:

Context Bundle Plan:
- Mode: [Standard / Persona Review (Selected Persona) / Monorepo Segmented / Multi-Persona Fan-Out (Persona Set)]
- Format: [.md or .zip]
- Persona Prompt: assets/templates/[selected-persona].md
- Included Paths:
  1. plugins/dev-utils/
  2. docs/architecture.md
- Output Target: temp/context-bundle-[name]/payload.[md|zip]

Proceed? (yes / adjust)

Phase 3: Manifest Construction

Generate file-manifest.json in temporary directory (temp/context-bundle-[name]/). For Persona-Driven mode, prompt.md (containing the selected persona prompt) MUST be listed as the first file entry in files.

{
  "title": "Sub-Agent Handoff Bundle",
  "description": "Task delegation bundle for Copilot CLI.",
  "excludes": ["**/*.png", "**/node_modules/**"],
  "files": [
    {
      "path": "temp/context-bundle-task/prompt.md",
      "note": "Primary Persona & Handoff Instructions"
    },
    {
      "path": "plugins/dev-utils/skills/github-issue-agent/",
      "note": "Target codebase"
    }
  ]
}

Phase 4: Bundler Script Execution

• Markdown (.md):

  python3 ./scripts/bundle.py --manifest temp/context-bundle-[name]/file-manifest.json --bundle temp/context-bundle-[name]/payload.md
  

• ZIP Archive (.zip):

  python3 ./scripts/bundle_zip.py --manifest temp/context-bundle-[name]/file-manifest.json --bundle temp/context-bundle-[name]/payload.zip
  

Inform user when payload is ready for handoff or clipboard copying.

Multi-Persona Fan-Out Mode (Mode 4)

Used when the same target content (typically a Phase 1 plan draft, per graph-planning-superpowers-policy.md §2.2-2.3) needs parallel review by more than one persona in a single pass, with a bounded convergence loop.

1. Select the persona set — default to the Graph Planning Phase 1 Fan-Out Trio (structural-architecture-reviewer.md, adversarial-security-auditor.md, tdd-contract-reviewer.md) unless the caller specifies a different set.
2. Build one manifest per persona, each following the standard Phase 3 manifest format above, with that persona's prompt.md as the first file entry and the same target content following it:

   temp/context-bundle-[name]/architecture-skeptic/file-manifest.json
   temp/context-bundle-[name]/security-edge-case-auditor/file-manifest.json
   temp/context-bundle-[name]/tdd-contract-reviewer/file-manifest.json
   

3. Run Phase 4 (bundler script execution) once per manifest — this produces one payload per persona, ready for parallel dispatch (CLI sub-agents, or paste-to-chat/browser for each).
4. Create a responses/ subfolder inside each persona's own directory, so responses collected back from external models land next to the prompt/payload that produced them:

   temp/context-bundle-[name]/architecture-skeptic/responses/round1.md
   temp/context-bundle-[name]/security-edge-case-auditor/responses/round1.md
   

   On a later round, regenerate prompt.md/payload.md in place (same paths) rather than creating round-numbered prompt/payload files — only responses/ files are round-numbered.
5. Convergence cap: the caller (typically red-team-review) tracks round count across all personas combined and caps at 2-3 rounds total, per graph-planning-superpowers-policy.md §2.3 — context-bundler only packages each round's bundles, it does not track rounds itself.