autonomous-skill（自主技能）

Autonomous Skill - Long-Running Task Execution

Execute complex, long-running tasks across multiple sessions using a dual-agent pattern (Initializer + Executor) with automatic session continuation via Codex non-interactive mode.

Quick Start

Use the run-session.sh script to manage autonomous tasks:

# Start a new autonomous task
~/.codex/skills/autonomous-skill/scripts/run-session.sh "Build a REST API for todo app"

# Continue an existing task
~/.codex/skills/autonomous-skill/scripts/run-session.sh --task-name build-rest-api-todo --continue

# List all tasks and their progress
~/.codex/skills/autonomous-skill/scripts/run-session.sh --list

# Show help
~/.codex/skills/autonomous-skill/scripts/run-session.sh --help

The runner intentionally leaves --model unset so Codex uses the active config.toml or selected profile model by default.

Directory Structure

All task data is stored in .autonomous/<task-name>/ under the project root:

project-root/
└── .autonomous/
    ├── build-rest-api/
    │   ├── task_list.md        # Master task checklist
    │   ├── progress.md         # Session-by-session notes
    │   ├── session.id          # Last Codex session ID for resumption
    │   └── session.log         # Codex stdout/stderr transcript with JSON events
    ├── refactor-auth/
    │   ├── task_list.md
    │   ├── progress.md
    │   └── session.id
    └── ...

This allows multiple autonomous tasks to run in parallel without conflicts.

Script Options

Usage:
  run-session.sh "task description"           Start new task (auto-generates name)
  run-session.sh --task-name <name> --continue Continue specific task
  run-session.sh --list                        List all tasks
  run-session.sh --help                        Show help

Options:
  --task-name <name>       Specify task name explicitly
  --continue, -c           Continue existing task
  --no-auto-continue       Don't auto-continue after session
  --max-sessions N         Limit to N sessions
  --list                   List all existing tasks
  --resume-last            Resume the most recent Codex session
  --network                Enable danger-full-access sandbox for tasks that need broader shell access

Workflow Overview

User Request → Generate Task Name → Create .autonomous/<task-name>/ → Execute Codex Sessions
                                                                            ↓
                                                                    ┌───────────────┐
                                                                    │ task_list.md  │
                                                                    │ exists?       │
                                                                    └───────┬───────┘
                                                                            │
                                                    ┌───────────────────────┴───────────────────────┐
                                                    │ NO                                        YES │
                                                    ▼                                              ▼
                                            ┌───────────────┐                            ┌───────────────┐
                                            │  INITIALIZER  │                            │   EXECUTOR    │
                                            │  - Analyze    │                            │  - Read state │
                                            │  - Break down │                            │  - Next task  │
                                            │  - Create     │                            │  - Implement  │
                                            │    task_list  │                            │  - Mark done  │
                                            └───────────────┘                            └───────────────┘
                                                                            │
                                                                            ▼
                                                                    ┌───────────────┐
                                                                    │ All complete? │
                                                                    └───────┬───────┘
                                                                            │
                                                            ┌───────────────┴───────────────┐
                                                            │ NO                        YES │
                                                            ▼                              ▼
                                                    Auto-continue               Exit with success
                                                    (3 sec delay)

Usage Examples

Example 1: Start New Task

~/.codex/skills/autonomous-skill/scripts/run-session.sh "Build a REST API for todo app"

Output:

ℹ Generated task name: build-rest-api-todo
==========================================
  SESSION 1 - build-rest-api-todo
==========================================

==========================================
  INITIALIZER SESSION
==========================================
Task: Build a REST API for todo app
Task Name: build-rest-api-todo
Task Directory: .autonomous/build-rest-api-todo

[Codex creates task_list.md with 25 tasks...]

✓ Initializer session complete
ℹ Session ID saved: 550e8400-e29b-41d4-a716-446655440000

=== Progress: 0/25 ===

Continuing in 3 seconds... (Press Ctrl+C to pause)

Example 2: Continue Existing Task

~/.codex/skills/autonomous-skill/scripts/run-session.sh --task-name build-rest-api-todo --continue

Example 3: Resume with Session Context

# Resume the Codex session (preserves conversation context)
~/.codex/skills/autonomous-skill/scripts/run-session.sh --task-name build-rest-api-todo --continue --resume-last

Example 4: List All Tasks

~/.codex/skills/autonomous-skill/scripts/run-session.sh --list

Output:

==========================================
  AUTONOMOUS TASKS
==========================================
  ✓ build-rest-api-todo (25/25 - 100% complete) [session: 550e8400...]
  ○ refactor-auth (12/30 - 40%) [session: 661f9511...]
  ? incomplete-task (no task_list.md)

Example 5: With Network Access

# Enable network access for tasks that need API calls
~/.codex/skills/autonomous-skill/scripts/run-session.sh --network "Fetch data from GitHub API and analyze"

Key Files

For each task in .autonomous/<task-name>/:

Important Notes

1. Task Isolation: Each task has its own directory, no conflicts
2. Task Naming: Auto-generated from description (lowercase, hyphens, max 30 chars)
3. Task List is Sacred: Never delete or modify task descriptions, only mark [x]
4. One Task at a Time per Session: Focus on completing tasks thoroughly
5. Auto-Continue: Sessions auto-continue with 3s delay; Ctrl+C to pause
6. Session Resumption: Use --resume-last to preserve Codex conversation context
7. Configured Model: The runner does not pass --model; it uses the active Codex config/profile model
8. Network Mode: --network switches the sandbox override to danger-full-access while keeping approval policy non-interactive
9. Git Hygiene: Consider adding .autonomous/ to .gitignore to avoid committing logs

Codex CLI Reference

The script uses these Codex commands internally. It intentionally uses config overrides instead of --full-auto so unattended runs do not inherit on-request approvals from the current CLI:

# Non-interactive execution with file edits (fully autonomous)
# Uses the configured model from the active Codex config/profile
codex exec \
  -c 'approval_policy="never"' \
  -c 'sandbox_mode="workspace-write"' \
  --skip-git-repo-check \
  --json \
  "prompt"

# Resume previous session
codex exec resume \
  -c 'approval_policy="never"' \
  -c 'sandbox_mode="workspace-write"' \
  --skip-git-repo-check \
  --json \
  <SESSION_ID> \
  "prompt"

# Full access (file edits + shell network / unrestricted filesystem) - use with caution!
codex exec \
  -c 'approval_policy="never"' \
  -c 'sandbox_mode="danger-full-access"' \
  --skip-git-repo-check \
  --json \
  "prompt"

Troubleshooting