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

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

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]

--lightweight

--task-name <name>

--continue, -c

--list, -l

--completion-promise TEXT

--max-sessions N

--max-budget N.NN

--model <model>

--fallback-model <m>

--effort <level>

--no-auto-continue

--permission-mode <m>

--add-dir <dirs>

Hook Mode — Setup

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

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

--mode structured|lightweight

--max-iterations N

--completion-promise TEXT

--task-name NAME

The hook is registered in

hooks/hooks.json

.claude/autonomous-loop.local.md

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

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

.autonomous/<task>/run.lock

sessions/session-NNN.log

.claude/autonomous-loop.local.md

--list

--max-budget

--model sonnet