CC + Codex Spec Bootstrap Pipeline

A multi-agent pipeline where Claude Code (CC) orchestrates and Codex executes in parallel. CC analyzes the repo with GitNexus + ABCoder, creates Trellis task PRDs, then Codex agents fill the coding specs — each with access to the same code intelligence MCP tools.

Why This Exists

AI coding agents produce better code when they have project-specific coding guidelines (not generic templates). But filling those guidelines manually is tedious. This skill automates the bootstrap:

• You (Claude Code) analyze the repo architecture using GitNexus + ABCoder

• You create Trellis tasks with rich PRDs containing architectural context + MCP tool instructions

• Codex agents run those tasks in parallel, each filling one spec directory using the same MCP tools

The result: every spec file contains real code examples, actual patterns, and project-specific anti-patterns — not placeholder text.

Prerequisites

Before running this skill, ensure these tools are set up. See references/mcp-setup.md for detailed installation instructions.

Tool Purpose Required

Trellis Workflow framework with .trellis/spec/ structure Yes

GitNexus Code → knowledge graph (Tree-sitter + KuzuDB) Yes

ABCoder Code → UniAST (ts-morph for TS, tree-sitter for others) Yes

Codex CLI Parallel task execution agent Yes

Quick check:

Verify all tools

npx gitnexus status # GitNexus indexed? abcoder list-repos # ABCoder has ASTs? codex mcp list # Codex has MCP servers? python3 .trellis/scripts/get_context.py # Trellis initialized?

Phase 1: Analyze the Repository

Step 1: Index with GitNexus

npx gitnexus analyze

This builds a knowledge graph: nodes (symbols), edges (dependencies), clusters (module groups), and execution flows. Takes ~5s for a typical monorepo.

After indexing, use GitNexus MCP tools to understand the architecture:

gitnexus_query({query: "plugin system"}) # Find execution flows gitnexus_context({name: "SomeClass"}) # 360-degree symbol view gitnexus_cypher({query: "MATCH (n:Class) RETURN n.name, n.file LIMIT 30"}) # Graph queries

Step 2: Parse with ABCoder

ABCoder provides precise AST analysis — function signatures, type definitions, cross-file dependency chains.

Parse each package

abcoder parse /path/to/package --lang typescript --name package-name --output ~/abcoder-asts

Then use ABCoder MCP tools:

get_repo_structure({repo_name: "package-name"}) get_file_structure({repo_name: "package-name", file_path: "src/core/types.ts"}) get_ast_node({repo_name: "package-name", node_ids: [{mod_path: "...", pkg_path: "...", name: "ClassName"}]})

Step 3: Map the Architecture

Combine insights from both tools to understand:

• Package boundaries — which packages exist, what each one does

• Module clusters — GitNexus clusters resource shows functional groupings

• Key patterns — Fetcher/Provider/Plugin/Adapter/Router patterns

• Cross-package data flows — how data moves between packages

• Error handling patterns — how errors propagate

• State management — what's stateless vs stateful

Write down your findings — they go into the PRDs.

Phase 2: Create Trellis Tasks

Task Decomposition Strategy

Create one task per (package, layer) combination. Each task is independently executable by a Codex agent.

Typical decomposition for a monorepo:

package-a/backend → Task 1 package-a/frontend → Task 2 package-b/backend → Task 3 package-b/frontend → Task 4 cross-layer-guide → Task 5

Skip layers that don't apply (e.g., no frontend task for a pure CLI library).

Create Task Directories

python3 .trellis/scripts/task.py create "Fill <package> <layer> spec" --slug <package>-<layer>-spec

Write PRDs

Each PRD must contain these sections. This is the critical part — the PRD is the entire context a Codex agent receives.

Fill <package> <layer> spec

Goal

One sentence: what to analyze, what files to fill.

Context

Project-specific architectural knowledge you gathered in Phase 1. Key concepts, patterns, abstractions — everything the agent needs to understand the codebase without reading every file.

Tools Available

[Use the MCP Tools Template below]

Files to Fill

List each spec file with bullet points on what to document. Include hints about which source files to analyze.

Important Rules

Spec files are NOT fixed — adapt to reality

• Delete template files that don't apply
• Create new files for patterns templates don't cover
• Rename files if template names don't fit
• Update index.md to reflect the final set

Parallel agents — stay in your lane

• ONLY modify files under your assigned spec directory
• DO NOT modify source code, other spec directories, or task files
• DO NOT run git commands
• You may read any file for analysis

Acceptance Criteria

• Real code examples from the actual codebase (with file paths)
• Anti-patterns documented
• No placeholder text remaining
• index.md reflects actual file set

Technical Notes

Package path, language, framework, build tools, key deps.

MCP Tools Template for PRDs

Include this in every PRD so Codex knows how to call the tools:

Tools Available

You have two MCP servers configured — use both for accurate specs:

GitNexus MCP (architecture-level: clusters, execution flows, impact)

ABCoder MCP (symbol-level: AST nodes, signatures, cross-file deps)

Recommended Workflow

1. GitNexus first — find relevant execution flows and clusters
2. ABCoder second — get exact code patterns and signatures
3. Read source files — for full context where needed
4. Write specs — with real code examples from steps 2-3

Phase 3: Launch Codex Agents

Run in Parallel

Each task is independent — launch all agents simultaneously:

One terminal per task

codex -q "Read .trellis/tasks/<task-slug>/prd.md and execute the task. Use GitNexus and ABCoder MCP tools to analyze the codebase, then fill all spec files listed in the PRD."

Monitor Progress

Check which spec files have been filled:

Line counts — 0 or ~50 means still template

find .trellis/spec -name "*.md" -exec sh -c 'echo "$(wc -l < "$1") $1"' _ {} ; | sort -rn

Check for remaining placeholders

grep -rl "To be filled" .trellis/spec/

Newly created or modified files

find .trellis/spec -name "*.md" -newer .trellis/tasks/ -exec ls -la {} ;

Review Results

After all agents complete:

• Check line counts — substantive files should be 80+ lines

• Grep for leftover placeholders

• Spot-check a few files for real code examples vs generic advice

• Verify index.md in each directory reflects actual files

Checklist

• GitNexus analyzed (npx gitnexus analyze )

• ABCoder parsed all packages

• GitNexus + ABCoder MCP configured for both Claude Code and Codex

• Architecture mapped (packages, patterns, boundaries)

• One task per (package, layer) created with task.py create

• Each PRD has: Context, MCP Tools, Files to Fill, Rules, Acceptance Criteria

• Codex agents launched in parallel

• Results reviewed — no placeholders, real code examples present