---
name: autonomous-execution-mode
description: Use to enable safe one-shot autonomous execution for an approved slice. Defines the pre-approved command list, the auto-continue rules, and the human-confirmation boundary so the agent does not stop to ask about already-authorized actions, but still pauses on destructive, cross-tenant, or out-of-scope operations.
version: 1.0.0
origin: captured
generation: 0
status: trial
trigger_phases: ["planning", "implementation", "validation", "closure"]
applicable_agents: ["Copilot Orchestrator", "Copilot Implementation", "Copilot Drift Check", "Copilot QA Automation", "Copilot Frontend Developer"]
---

# Autonomous Execution Mode

This skill complements `agent-execution-governance`. It does **not** weaken safety boundaries; it removes redundant prompts for actions that have already been approved by the slice plan or by repeated user authorization.

## Activation

Autonomous Execution Mode is active when **all** of the following hold:

1. The slice has an approved plan (Plan Mode artifact, blade-plan/blade-spec output, or DomainCompletion ledger entry).
2. The user has not explicitly requested step-by-step confirmation in the current turn.
3. No `BLOCKED` verdict is open from Reality Fit Gate / Schema Guardian / Invariant Sentinel / Drift Check.
4. The pending command falls inside the **Pre-Approved Command List** below or is a strict subset of an action the user previously authorized in the current conversation.

If any condition fails, fall back to the default confirmation behavior.

## Pre-Approved Command List (no re-confirmation needed)

These are read-only or local, reversible, and non-destructive. They may be executed without asking the user again, even if invoked repeatedly.

### Read / inspect
- `git status`, `git diff`, `git log --oneline -n <N>`, `git show <ref>`, `git branch`, `git remote -v`
- `ls`, `cat`, `head`, `tail`, `wc`, `find`, `grep`, `rg` (read-only filters)
- `pwd`, `which`, `command -v`, `printenv`

### Project tooling (read / validate)
- `pnpm install` (no `--force`), `pnpm i`, `npm ci`
- `pnpm run lint`, `pnpm run test`, `pnpm run build`, `pnpm exec playwright test`
- `mvn -q -DskipTests=true compile`, `mvn -q test`, `mvn -q dependency:tree`
- Any task defined in `.vscode/tasks.json` of this workspace, including:
  - `BladeX: Toolchain Doctor`
  - `BladeX: Backend Compile`
  - `BladeX: Frontend Install`
  - `BladeX: Frontend Build`
  - `BladeX: Frontend Health Check`
  - `BladeX: Full Compile`

### Repository helpers
- `python tools/agent-tools/<name>.py [args]` (validators / generators only; see `tools/agent-tools/README.md`)
- `bash tools/blade/cli.sh <subcommand>` for the subcommands wrapped by the workspace tasks above

### Local edits
- File creation / edits inside `.github/agents/**`, `.github/skills/**`, `.github/agentLog/**`, `M-DOC/**`, `mes-enreach-mom-web/src/**`, `mes-enreach-mom-base-server/**`, `tools/**` when they fall inside the approved slice scope.

## Always-Confirm List (never auto-run)

Even in Autonomous Execution Mode, the agent must pause and ask before running:

- Any `rm -rf`, `git reset --hard`, `git clean -fd`, `git push --force`, `git push` to a protected branch
- `git commit --amend` against an already-pushed commit, force-push, branch deletion (`git branch -D`, `git push --delete`)
- DROP / TRUNCATE / DELETE without WHERE on production-like databases; any DB schema change outside the approved slice
- Package publishing (`pnpm publish`, `npm publish`, `mvn deploy`)
- Network calls to external services not pre-listed (PR comments, issue comments, message webhooks)
- Container registry pushes, infra-as-code apply, kubectl apply against shared clusters
- Any command that touches another tenant's data or another workspace folder outside `${workspaceFolder}`

## Auto-Continue Rules

When activation conditions hold and the next step is implied by the accepted plan:

1. **Implementation continuity** — once Implementation starts on an approved slice, continue to field coverage, validation mirroring, permission integration, and self-check without pausing only to ask "shall I continue".
2. **Bundle continuity** — under `mode: DeliveryBundle`, run the stage sequence end-to-end and emit the final bundle index in one pass.
3. **DomainCompletion continuity** — advance one active module at a time per ledger; do not pause between sub-steps of the same active module.
4. **Validation continuity** — running lint / tests / health checks listed above does not require re-confirmation per run.

Stop conditions: real blocker hit, a gate returns `BLOCKED`, the user explicitly changes scope, or the next step would require an Always-Confirm List action.

## Output discipline

When auto-continuing, keep agent narration brief: state the current step, the command class (read / build / test / edit), and the next step. Do not re-explain the entire plan after every command.

## Safety net

- Each auto-run command must still be logged in the slice's agentLog summary at closure.
- If the agent is uncertain whether a command is in the Pre-Approved list, default to confirm; do not over-extend.
- This skill does not bypass `skill-safety-scanner` or `agent-execution-governance` write-boundary rules.