bereniketech/copilot-autonomous-tasks

Copilot Autonomous Tasks

Use this skill when GitHub Copilot agent mode should complete a project's task queue end-to-end: read task → implement → verify → hand off → clear → next task, with no mid-task questions or manual approvals.

---

1. VSCode Setup (One-Time)

Create .vscode/settings.json at the project root (commit this file):

{
  "github.copilot.chat.agent.autoApprove": true,
  "github.copilot.chat.agent.terminal.allowList": ["*"]
}

Switch the Copilot Chat input to Agent mode via the dropdown next to the chat input.

Rule: "terminal.allowList": ["*"] allows every terminal command without prompting. Tighten to ["npm *", "git *", "npx *"] if the project warrants tighter control.

---

2. copilot-instructions.md — Autonomous Loop Directive

The loop lives in .github/copilot-instructions.md, not the activation prompt. Copilot auto-loads this file on every new chat, so the loop persists across /clear context resets.

Replace ## Start Here in every project's copilot-instructions.md with:

## Autonomous Loop — Begin Immediately on Session Start
On session start, do not wait for user input. Start from step 1 now.

1. Read `Current task:` from `## Active Feature` above. If it reads `ALL TASKS COMPLETE`, stop and notify the user.
2. Open that task file — it is self-contained with all acceptance criteria.
3. Read the skill files listed in `## Skills` above (once per session, not per task).
4. Implement all acceptance criteria. Zero clarifying questions mid-task — resolve unknowns from the task file, skills, or existing code. If genuinely blocked, write `BLOCKED: <reason>` in the task file's `## Status` and stop.
5. Run `/verify` — all phases must pass before continuing. On failure, fix the root cause. If the same phase fails 3× in a row, write `VERIFY_FAIL: <phase> — <error>` in `## Status` and stop.
6. Run `/task-handoff` — updates `Current task:` in this file and commits.
7. Run `/clear` — mandatory between every task to reset context. This file auto-loads in the new chat; the loop restarts from step 1 automatically.

Rule: The "Begin Immediately on Session Start" header is what drives auto-continuation after /clear. Copilot re-reads this file in the new chat and starts the loop without any user re-prompt.

Rule: Always read Current task: from copilot-instructions.md — it is auto-loaded and always reflects the latest pointer after /task-handoff.

---

3. Activation Prompt

Paste this once to start a session. The loop then runs autonomously:

Read .github/copilot-instructions.md. Follow the ## Autonomous Loop instructions exactly.

That's all. The full loop logic is in copilot-instructions.md and persists across context resets.

---

4. Behavioral Rules

When operating as an autonomous Copilot agent:

• Never ask a clarifying question mid-task. Resolve unknowns from the task file, skill files, or existing code. If genuinely unresolvable, write BLOCKED: <reason> in the task file's ## Status section and stop.
• Never skip /verify. All 6 phases must pass before /task-handoff.
• On verify failure: fix the root cause. If the same phase fails 3 consecutive times, write VERIFY_FAIL: <phase> — <error summary> in ## Status and stop. Do not commit.
• On destructive operations (drop table, force push to main, delete untracked files): stop and ask the user explicitly. Never auto-approve destructive ops regardless of agent mode settings.
• Always /clear between tasks. Mandatory after every /task-handoff. copilot-instructions.md auto-loads in the new chat and restarts the loop.

Rule: Never modify test files to make tests pass. Fix the code under test.

---

5. Loading Skills

.github/copilot-instructions.md lists absolute skill file paths under ## Skills. Read each listed skill file before writing any code. Skills contain the coding standards, test patterns, and security rules for this codebase.

Do NOT use @ imports — Copilot cannot resolve them. Use the explicit paths in ## Skills only.

---

6. Termination Conditions

Stop the loop and notify the user when:

| Condition | Action | |---|---| | Current task: ALL TASKS COMPLETE | Output "ALL TASKS COMPLETE." Stop. | | Same verify phase fails 3× in a row | Write VERIFY_FAIL to task file. Stop. Output failure summary. | | Task file ## Status is BLOCKED | Output "BLOCKED on <task-file>: <reason>". Stop. | | Required env var is missing | Output "MISSING: <VAR> needed for <task>". Stop. | | Destructive operation required | Stop. Ask the user before proceeding. |

---

7. After Each Task — Clear and Continue

After /task-handoff commits:

1. Run /clear — mandatory. Resets stale context from the completed task.
2. Copilot auto-loads copilot-instructions.md in the new chat.
3. The "Begin Immediately on Session Start" directive triggers the loop from step 1 — no user re-prompt needed.

Rule: Do not re-paste the activation prompt after /clear. copilot-instructions.md carries the full loop and auto-starts it.

---

8. Resume Mid-Task

If interrupted (context window hit, network error, manual stop), paste this to resume:

Read .claude/CLAUDE.md → Current task. Read that task file. Check ## Status for any
BLOCKED notes or partial work. Continue from where work stopped. When done, run
/verify then /task-handoff.

If the task file has no status notes, run git status and check recent file modification times to reconstruct where implementation stopped before resuming.