- human-approval-loop.md 45 lines · 2.4 KB
Human-in-the-Loop Approval Loop
Loop Engineering — same loop shape as the autonomous ones, but the stop condition is your approval instead of a passing test. The agent does the work, then waits at a gate: approve, revise, or skip.
🎯 Goal
Run the task, then pause and route the result to you for a decision before anything ships — treating human review as its own queue with clear actions.
⏱️ Schedule
Trigger: on-demand (each item pauses for your call before shipping).
▶️ Run it
/loop run the next task, then pause and send me approve / revise / skip before anything ships. On approve, continue and ship. On revise, take my note and redo. On skip, move to the next item. Never ship without my approval.🔁 Iteration steps
- Act — complete the task up to the point of shipping.
- Gate — the
plan-gatehook halts before any irreversible/ship step. - Ask — send approve / revise / skip via
telegram-notifications. - Branch — approve → ship (e.g.
/create-pr); revise → apply note and redo; skip → next item. - Observe — record the decision; nothing ships unattended.
🛑 Stopping condition
You approve (ship) or skip/redirect — the loop never ships without an explicit human decision.
💰 Budget & guardrails
The gate is the safety: irreversible actions wait for a human. Add a reminder/deadline so pending items don't pile up silently, and a budget for the work done before each gate.
🧩 Referenced components
hook:quality-gates/plan-gate— halts before irreversible/ship steps.hook:automation/telegram-notifications— sends the approve/revise/skip prompt.command:git-workflow/create-pr— the ship action on approval.
💡 Example
The agent drafts a refund-policy change, pauses, and pings you on Telegram. You hit "revise" with a note; it redoes the draft and pings again — and only ships after you approve.
Scanned by
NVIDIA SkillSpector
Automated static analysis (no LLM) · v2.12.0 · 2026-11-19
1 observation in 1 category
- Excessive Agency 1
SkillSpector looks for patterns such as shell commands, credential access, network calls or unpinned dependencies. An observation is something worth reading in the files before you install, not a confirmed vulnerability. Components that run code locally, like hooks, MCP servers and security tooling, naturally show more of them. Open one to see the lines it points at.
Autonomous Decision MakingExcessive Agency · human-approval-loop.md:41
| 39 | ## 🧩 Referenced components |
| 40 | - `hook:quality-gates/plan-gate` — halts before irreversible/ship steps. |
| 41 | - `hook:automation/telegram-notifications` — sends the approve/revise/skip prompt. |
| 42 | - `command:git-workflow/create-pr` — the ship action on approval. |
| 43 |
Learn more about NVIDIA SkillSpector or how components are reviewed.