Default Task Loop (Bounded Local Agent Loop)¶
The single, repeatable loop used for every coding task in this workspace, calibrated to keep autonomy bounded and verifiable.
The 6-step loop¶
1. Ask (intent capture)¶
- State the goal in 1-2 sentences. If unclear, the agent must use the Ask tool before planning.
- Identify allowed-edit scope (specific files / folders / repos). Anything outside scope requires explicit re-approval.
- Mark data sensitivity (
private_sensitive,internal,low_sensitive).
Outputs: scope statement + sensitivity tag.
2. Plan¶
- Plan-mode review of approach. Multiple options weighed only if non-trivial.
- Mermaid diagram or numbered steps if ≥ 3 distinct edits or files.
- Verifier criteria written before implementation: how do we know it's done?
Outputs: plan + acceptance criteria.
Trigger to switch to plan mode: ambiguity, > 3 files touched, architectural impact, or unfamiliar territory.
3. Bounded edit¶
- Smallest possible code change satisfying the plan.
- Stay strictly within the scope statement.
- No "while we're here" cleanups.
- All risky operations (shell, MCP write, file ops outside scope) require explicit approval.
Outputs: diff confined to declared scope.
4. Test¶
- Run existing tests touched by the change; add new ones for new behavior.
- Lint / type-check what changed.
- For docs / config / infra changes, do a dry-run or
--checkequivalent.
Outputs: green test run; any flakes triaged or noted.
5. Verifier pass¶
- Invoke
/verifiersubagent (.cursor/agents/verifier.md). - Verifier checks: implementation actually exists; acceptance criteria met; tests truly pass; no obvious gaps.
- If verifier reports incomplete → return to step 3 with the gap, do not skip.
Outputs: verifier report attached to the task in daily-log.md.
6. Commit¶
- Atomic commit message, conventional-commit style:
feat:,fix:,refactor:,docs:,test:,chore:,security:.- Optional: trigger Bugbot manually with
cursor reviewon PR. - Update
instrumentation/daily-log.mdrow: outcome + rework flag (still empty unless rework needed).
Outputs: commit (or staged change ready for PR).
Approval gates¶
These actions always require an explicit user approval, even with autorun:
- Any shell command outside the workspace.
- Any MCP write.
- Network egress beyond Cursor docs / package managers / GitHub.
- Edit to anything in the user-only secret store.
- Edit to
.cursor/permissions.json,mcp.json, or governance files. - Any agent-credential provisioning step.
Anti-patterns to avoid¶
- Multi-file refactors disguised as a "small fix".
- Skipping verifier "because it's small".
- Marking task done before tests are green.
- Editing the same file in two passes when one was possible.
- Generating a plan after editing.
Success gate for Phase 1¶
5 consecutive tasks completed through this exact loop, logged in daily-log.md, with verifier reports attached, before adopting any new capability.