Codeagent

Codeagent Wrapper Integration

Overview

Execute codeagent-wrapper commands with pluggable AI backends (Codex, Claude, Gemini, OpenCode), agent presets, auto-detected skill injection, and parallel task orchestration. Default to background execution, and prefer --parallel whenever work can be split into independent tasks.

When to Use

• Complex code analysis requiring deep understanding
• Large-scale refactoring across multiple files
• Multi-agent orchestration (explore → design → implement → review)
• Automated code generation with backend/agent selection
• Parallel task execution with dependency management

Quick Reference

codeagent-wrapper [flags] <task|-> [workdir]
codeagent-wrapper [flags] resume <session_id> <task|-> [workdir]
codeagent-wrapper --parallel [flags] < tasks_config

CLI Flags

Backends

Agent Presets

Agent presets bundle backend, model, prompt, and tool control into a reusable name. Use --agent <name> to select.

Sources (checked in order): 1. ~/.codeagent/models.json → agents.<name> object 2. ~/.codeagent/agents/<name>.md → markdown file becomes the prompt

Agent config fields (in models.json):

{
  "agents": {
    "develop": {
      "backend": "codex",
      "model": "gpt-4.1",
      "prompt_file": "~/.codeagent/prompts/develop.md",
      "reasoning": "high",
      "yolo": true,
      "allowed_tools": ["Read", "Write", "Bash"],
      "disallowed_tools": ["WebFetch"]
    }
  }
}

Common agent presets:

Skill Injection

Auto-Detection

When --skills is not specified, skills are auto-detected from the working directory:

Manual Override

codeagent-wrapper --agent develop --skills golang-base-practices,frontend-design - . <<'EOF'
Implement full-stack feature...
EOF

Skills are loaded from ~/.claude/skills/{name}/SKILL.md, stripped of YAML frontmatter, and injected into the task prompt.

Usage Patterns

Single Task (HEREDOC recommended)

codeagent-wrapper --backend codex - [workdir] <<'EOF'
<task content here>
EOF

Default execution mode is background. Run in foreground only when the next step requires the full response immediately.

When running in foreground, codeagent-wrapper emits liveness frames on stdout before the final answer:

[codeagent-progress] status=started ...
[codeagent-progress] status=streaming ...
[codeagent-progress] status=running ...
[codeagent-progress] status=completed ...

Treat these lines as progress only. Do not conclude "no data returned" and do not start doing the task yourself while progress frames are still arriving.

With Agent Preset

codeagent-wrapper --agent develop --skills golang-base-practices - . <<'EOF'
Implement the authentication middleware following existing patterns.
EOF

Simple Task (short prompts only)

codeagent-wrapper --backend codex "simple task description" [workdir]

Auto-stdin detection: When task length exceeds 800 characters or contains special characters (\n, \, ", ', ` `, $), stdin mode is used automatically. Use -` to force stdin mode explicitly.

Resume Session

codeagent-wrapper --backend codex resume <session_id> - <<'EOF'
<follow-up task>
EOF

# Or with agent preset
codeagent-wrapper --agent develop resume <session_id> - <<'EOF'
<follow-up task>
EOF

Worktree Isolation

Execute in an isolated git worktree to keep changes separate from the main branch:

codeagent-wrapper --agent develop --worktree - . <<'EOF'
Implement feature in isolation...
EOF

Rules:

• Read-only agents (code-explorer, code-architect, code-reviewer) do NOT need worktree
• Only develop agent needs worktree when making changes

Parallel Execution

Use --parallel by default for multi-step or multi-agent work. Fall back to single-task mode only when the work is truly linear or the next step depends on the full output of the current task.

Task Config Format

codeagent-wrapper --parallel <<'EOF'
---TASK---
id: <unique_id>
agent: <agent_name>
workdir: <path>
backend: <name>
model: <model_name>
reasoning_effort: <low|medium|high>
skills: <skill1>, <skill2>
dependencies: <id1>, <id2>
session_id: <id>
skip_permissions: true|false
worktree: true|false
---CONTENT---
<task content>
EOF

Task header fields (all optional except id):

Multi-Agent Orchestration Example

codeagent-wrapper --parallel <<'EOF'
---TASK---
id: p1_architecture
agent: code-explorer
workdir: .
---CONTENT---
Map architecture for the authentication subsystem. Return: module map + key files with line numbers.

---TASK---
id: p1_conventions
agent: code-explorer
workdir: .
---CONTENT---
Identify testing patterns, conventions, config. Return: test commands + file locations.

---TASK---
id: p2_design
agent: code-architect
workdir: .
dependencies: p1_architecture, p1_conventions
---CONTENT---
Design minimal-change implementation plan based on architecture analysis.

---TASK---
id: p3_backend
agent: develop
workdir: .
skills: golang-base-practices
dependencies: p2_design
---CONTENT---
Implement backend changes following the design plan.

---TASK---
id: p3_frontend
agent: develop
workdir: .
skills: vercel-react-best-practices,frontend-design
dependencies: p2_design
---CONTENT---
Implement frontend changes following the design plan.

---TASK---
id: p4_review
agent: code-reviewer
workdir: .
dependencies: p3_backend, p3_frontend
---CONTENT---
Review all changes for correctness, edge cases, and KISS compliance.
Classify each issue as BLOCKING or MINOR.
EOF

Dependency Resolution

• Tasks are topologically sorted (Kahn's algorithm)
• Circular dependencies are detected and reported
• Failed parent tasks cause dependent tasks to be skipped
• Independent tasks at the same level run concurrently

Output Modes

Summary (default): Structured report per task with extracted fields:

=== Execution Report ===
3 tasks | 2 passed | 1 failed

### task_id PASS 92%
Did: Brief description of work done
Files: file1.ts, file2.ts
Tests: 12 passed
Log: /tmp/codeagent-xxx.log

### task_id FAIL
Exit code: 1
Error: Assertion failed
Detail: Expected status 200 but got 401
Log: /tmp/codeagent-zzz.log

Full output (--full-output): Complete task messages included. Use only for debugging specific failures.

Structured JSON Output

codeagent-wrapper --parallel --output results.json <<'EOF'
...
EOF

Produces:

{
  "results": [
    {
      "task_id": "task_1",
      "exit_code": 0,
      "message": "...",
      "session_id": "...",
      "coverage": "92%",
      "files_changed": ["file.ts"],
      "tests_passed": 12,
      "log_path": "/tmp/..."
    }
  ],
  "summary": { "total": 3, "success": 2, "failed": 1 }
}

Return Format

Single task output:

Agent response text here...

---
SESSION_ID: 019a7247-ac9d-71f3-89e2-a823dbd8fd14

Environment Variables

Config file: Supports ~/.codeagent/config.(yaml|yml|json|toml) with the same keys as CLI flags (kebab-case). Env vars use CODEAGENT_ prefix with underscores.

Exit Codes

Invocation Pattern

Single Task:

Bash tool parameters:
- command: codeagent-wrapper --agent <agent> --skills <skills> - [workdir] <<'EOF'
  <task content>
  EOF
- background: true
- description: <brief description>

Parallel Tasks:

Bash tool parameters:
- command: codeagent-wrapper --parallel <<'EOF'
  ---TASK---
  id: task_id
  agent: <agent>
  workdir: /path
  skills: <skill1>, <skill2>
  dependencies: dep1, dep2
  ---CONTENT---
  task content
  EOF
- background: true
- description: <brief description>

Critical Rules

NEVER kill codeagent processes. Long-running tasks are normal. Default to background execution and inspect progress instead of blocking on the command.

1. Check task status via log file:

   tail -f /tmp/claude/<workdir>/tasks/<task_id>.output

2. Prefer wrapper progress frames over ad-hoc log scraping: If stdout is still emitting [codeagent-progress] ..., the task is alive and has not stalled.

3. Check process without killing:

   ps aux | grep codeagent-wrapper | grep -v grep

Why: Killing wastes API costs and loses progress.

Tool Control (Claude Backend)

When using Claude backend with agent presets, control available tools:

{
  "agents": {
    "safe-develop": {
      "backend": "claude",
      "allowed_tools": ["Read", "Write", "Bash", "Grep", "Glob"],
      "disallowed_tools": ["WebFetch", "WebSearch"]
    }
  }
}

Passed as --allowedTools and --disallowedTools to Claude CLI. Explicit enumeration only (no wildcards).