Autonomous Skill - Multi-Session Task Execution

Execute complex tasks across multiple Claude Code sessions with automatic continuation, progress tracking, and two completion mechanisms (promise tags + checkbox counting).

Two Execution Modes

Headless Mode (default)

Spawns claude -p child sessions in a bash loop. Best for background/unattended work.

bash <skill-dir>/scripts/run-session.sh "Build a REST API" --max-sessions 10

Hook Mode (in-session)

Uses a Stop hook to intercept session exit and feed the prompt back. Runs inside the current interactive session — no nesting issues.

bash <skill-dir>/scripts/setup-loop.sh "Build a REST API" --max-iterations 10

Two Task Strategies

Structured (default)

Full task decomposition: Initializer creates task_list.md with phased sub-tasks, Executor picks up and completes them one by one. Best for complex, multi-phase projects.

bash <skill-dir>/scripts/run-session.sh "Build a REST API for todo app"

Lightweight (--lightweight)

Ralph-style iteration: same prompt repeated each session, no task decomposition. Best for iterative tasks with clear success criteria (TDD, bug fixing, refactoring).

bash <skill-dir>/scripts/run-session.sh "Fix all failing tests in src/" --lightweight

Completion Detection

Two complementary mechanisms — whichever triggers first wins:

1. Promise tags (both modes): The agent outputs <promise>DONE</promise> when work is genuinely complete. Default promise is DONE; customize with --completion-promise. The agent is instructed to only output the promise when the work is truly finished — not to escape the loop.

2. Checkbox counting (structured mode only): All [ ] items in task_list.md are marked [x].

Directory Layout

project-root/
├── .autonomous/
│   └── <task-name>/
│       ├── task_list.md      # Master checklist (structured mode)
│       ├── progress.md       # Per-session progress log
│       ├── .mode             # "structured" or "lightweight"
│       ├── sessions/         # Transcript logs per session
│       │   ├── session-001.log
│       │   └── session-002.log
│       └── run.lock          # Prevents concurrent runs
└── .claude/
    └── autonomous-loop.local.md  # Hook mode state (when active)

Headless Mode — CLI Reference

bash <skill-dir>/scripts/run-session.sh "task description" [OPTIONS]

Flag	Description	Default
--lightweight	Ralph-style iteration (no task decomposition)	structured
--task-name <name>	Explicit task name	Auto-generated
--continue, -c	Continue most recent or named task	—
--list, -l	List all tasks with progress	—
--completion-promise TEXT	Promise phrase for completion	DONE
--max-sessions N	Stop after N sessions	Unlimited
--max-budget N.NN	Per-session dollar budget	5.00
--model <model>	Model alias or full name	sonnet
--fallback-model <m>	Fallback if primary overloaded	—
--effort <level>	Thinking effort (low/medium/high)	high
--no-auto-continue	Run one session only	—
--permission-mode <m>	Permission mode	auto
--add-dir <dirs>	Extra directories to allow	—

Hook Mode — Setup

For in-session loops (no child process spawning):

bash <skill-dir>/scripts/setup-loop.sh "task description" [OPTIONS]

Flag	Description	Default
--mode structured|lightweight	Task strategy	structured
--max-iterations N	Max loop iterations	Unlimited
--completion-promise TEXT	Promise phrase	DONE
--task-name NAME	Explicit task name	Auto-generated

The hook is registered in hooks/hooks.json. When active, the Stop hook reads .claude/autonomous-loop.local.md and blocks exit until the promise is detected or max iterations reached.

To cancel an active hook-mode loop: rm .claude/autonomous-loop.local.md

Workflow Detail

Structured Mode

1. Initializer Agent — analyzes task, creates phased task_list.md, begins work
2. Executor Agent — reads task list + progress, verifies previous work, completes next task
3. Auto-Continue — checks promise tags + checkboxes; if not done, spawns next session

Lightweight Mode

1. Same prompt fed each iteration
2. Agent sees its previous work in files and git history
3. Iterates until work is complete and promise tag is output
4. No task_list.md — completion is purely promise-based

When to Use Which

Scenario	Strategy	Mode
Build a full application	Structured	Headless
Fix all failing tests	Lightweight	Either
Refactor a module	Lightweight	Either
Multi-phase project	Structured	Headless
Quick iterative fix	Lightweight	Hook
Overnight batch work	Structured	Headless

Important Constraints

1. task_list.md is append-only for completions: Only change [ ] → [x]
2. One runner per task: Lock file prevents concurrent sessions on same task
3. Project files stay in project root: .autonomous/ is only for tracking
4. Promise integrity: The agent must not output <promise>DONE</promise> until genuinely complete
5. Cost awareness: Default per-session budget is $5. Adjust with --max-budget

Troubleshooting

Issue	Solution
"Lock file exists"	Previous run crashed. Remove .autonomous/<task>/run.lock
Session keeps failing	Check sessions/session-NNN.log for errors
Nested session error	Script auto-unsets CLAUDECODE; use hook mode as alternative
Hook loop won't stop	Delete .claude/autonomous-loop.local.md
Task not found	Run --list to see available tasks
Want to restart	Delete the task directory and start fresh
Cost too high	Lower --max-budget or use --model sonnet