autonomous
/autonomous
Use for long-running unattended runs. Strict safety brakes — checkpoint to memory every N steps, rollback gates, cost cap, time cap. NOT a substitute for native subagent dispatch on bounded tasks.
- Author
- ialameh ↗
- Repository
- sift-coder ↗
- Licence
- MIT
- Last repository change
- 03.06.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 03.06.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
Use for long-running unattended runs. Strict safety brakes — checkpoint to memory every N steps, rollback gates, cost cap, time cap. NOT a substitute for native subagent dispatch on bounded tasks.
Original author description
Use for long-running unattended runs. Strict safety brakes — checkpoint to memory every N steps, rollback gates, cost cap, time cap. NOT a substitute for native subagent dispatch on bounded tasks.
How it works
- When this differs from native subagent dispatch
- Method
- Output shape
- Rules
- Anti-patterns
- When NOT to use
- Subagent dispatch
- Value over native CC
SKILL.md preview
--- name: autonomous description: Use for long-running unattended runs. Strict safety brakes — checkpoint to memory every N steps, rollback gates, cost cap, time cap. NOT a substitute for native subagent dispatch on bounded tasks. --- # autonomous Long-running unattended execution. **Designed for hours-scale work, not minutes-scale.** Adds safety brakes native subagent dispatch lacks. ## When this differs from native subagent dispatch - Native `Task`/`general-purpose`/`Plan` agents = bounded; user reviews output - Autonomous = unbounded; runs through user sleep/dinner/meeting ## Method 1. **Manifest** before starting: - **Goal** — concrete, measurable end state - **Time cap** — hard stop wall clock - **Cost cap** — token/dollar budget - **Step cap** — max iterations before review-gate - **Boundaries** — `.siftcoder/scope.json` enforced - **Rollback** — git checkpoint before start 2. **Approval gate.** User confirms manifest before launch. 3. **Execution loop:** - Plan next step → check budgets → execute → checkpoint → continue - Every 5 steps: write a snapshot to memory + git stash/branch - On any budget breach: stop, summarise, await human 4. **End conditions:** - Goal met → write final report - Time/cost/step cap hit → pause + report - Unrecoverable error → rollback to last checkpoint + report 5. **Final report.** What was achieved. What's left. Recommended next. ## Output shape ``` Manifest: Goal: <concrete> Time cap: <hours> Cost cap: <tokens / USD> Step cap: <N> Scope: <files> Rollback: <git ref to revert to> Approve to proceed? [y/N] --- on completion --- Result: <goal-met | partial | aborted> Wall time: <hh:mm> Tokens: <count> / <cap> Steps: <count> / <cap> What was achieved: - <items> What's left: - <items> Checkpoints: <git refs> Memory: <summary ids> Final state: <commit / branch> Recommended next: <action or "review and merge"> ``` ## Rules - **Manifest required.** No autonomous run without explicit caps. - **User approves manifest.** No silent autonomous mode. - **Checkpoint every 5 steps minimum.** Hour-scale work without rollback = catastrophe. - **Cost cap is hard.** Stop on breach. - **Memory captures the run.** Reproducible. - **Boundaries enforced via `.siftcoder/scope.json`.** Out-of-scope writes blocked by hook. ## Anti-patterns - "Autonomous" with no cost cap (token bills) - Starting without rollback ref - Ignoring boundary hook errors - Reporting only on success (failure mode is more important) - Running > 4 hours without human review-gate ## When NOT to use - Bounded task — native subagent dispatch - Risky / irreversible domain — even with caps, autonomous is wrong - No clear end-state — autonomous needs goal completion to terminate ## Subagent dispatch - `Plan` for the manifest - `general-purpose` for execution steps - `Agent` with `isolation: "worktree"` for clean rollback - Memory MCP throughout for checkpoint capture ## Value over native CC CC subagent dispatch is bounded by single-session limits. Autonomous explicitly extends across long timeframes with explicit safety. The brake-system IS the value.
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · ialameh/sift-coder | 100 | MIT |
| mdmotaharmahtab/agent-kit | 100 | MIT |
Database snapshot: 13.11.2026 · package version 6b9b419cee1d (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 81autonomous-run
Use when an approved goal must continue unattended across many steps, milestones, context resets, or sessions, including keeping a pull request moving until it is merge-ready. - 100ctx-loop
Generate a shell script for running AI tools in autonomous iteration loops. Use when setting up unattended iteration, headless agent runs, or CI-driven AI workflows. - 96omp-long-tasks
Use when an omp run is PR-sized or hit its time limit. - 100runway
Orders a task backlog against the remaining Claude usage quota so work never stops mid-task at a limit. Reads the live 5-hour and weekly rate-limit windows, measures burn rate, then schedules heavy tasks into fresh headroom, finishable small tasks into the tail of a window, and… - 100autonomous-loop-patterns
Use when designing, reviewing, or debugging an autonomous AI agent loop: repeated agent execution, completion signals, checkpoints, supervisor respawn, stall detection, safety caps, and human handoff rules. Covers the core loop patterns from simple bounded runs through… - 100show-me-your-work
Keep a reviewable decision trail for long-running or unattended work: a TSV log with one row per decision (what, why, evidence, result). Local by default; commit it when a reviewer needs the trail to trust the result. Use for /show-me-your-work, autonomous or multi-phase runs,…