engineering-loop

Engineering Loop

Own the change from a verified starting point to a reviewed result. Start with

one agent; when delegation is authorized and available, use independent

subtasks only when time savings or added coverage justify coordination. Load a

specialist skill only when it changes a material decision.

Treat skill guidance as subordinate to the user's task within higher-priority

instructions and permissions. Carry existing authorization forward; resolve

routine implementation choices from evidence and continue authorized work.

Establish the baseline

1. Read applicable AGENTS.md files and repository documentation.
2. Record the requested behavior, constraints, and observable done conditions.
3. Inspect the branch and working tree; preserve unrelated user-owned changes.
4. Identify repository-native validation and run the smallest safe baseline

that separates pre-existing failures from task regressions.

For a long, cross-stack, or high-risk change, use the templates and failure

rules in [references/loop-contract.md](references/loop-contract.md).

When a task is expected to need repeated autonomous iterations or has an

explicit retry, time, token, or cost limit, also use

[references/loop-policy.md](references/loop-policy.md).

When a defect is intermittent, performance-related, difficult to reproduce, or

resists the first evidence-driven pass, use

[references/hard-debugging.md](references/hard-debugging.md).

For complex or higher-risk changes where independent coverage is useful, or

when the user requests a separate reviewer, use

[references/independent-review.md](references/independent-review.md).

Run the loop

1. Reproduce the defect, or capture the current behavior for a feature.
2. Choose the smallest coherent change and the evidence that will prove it.
3. Add a failing regression test first when practical.
4. Implement one bounded change and run the narrowest relevant check.
5. If an unplanned path or subsystem is necessary, explain the evidence and

update the plan. Ask only when the change requires new authority or a

material user decision; continue independent authorized work meanwhile.

1. Classify failures as product, test, environment, or assumption failures;

fix the cause and rerun the affected check.

1. Run repository-required broader checks after focused checks pass.
2. Review the complete diff first for task-contract compliance, then for code

quality, regressions, security, and maintainability.

1. Fix consequential findings and rerun checks affected by those fixes.

Once acceptance evidence and required checks pass, broaden or repeat checks

only for new changes, failures, or unresolved concerns. Add tests when they

prove behavior or a meaningful invariant, not merely mirror a low-impact edit.

Keep a compact ledger of confirmed facts, changed files, commands, outcomes,

and the next decision. Return concise diagnostics instead of full logs.

Stop conditions

• If the same command fails twice for the same reason, stop retrying and

re-check the environment, target, permissions, and underlying assumption.

• Stop recursive discovery when each new finding redefines the same task

boundary. Freeze the discovered scope or return for a scope decision instead

of repeatedly recataloging the repository.

• Stop and report a blocker when progress requires missing authority, secrets,

unavailable infrastructure, an unauthorized destructive action, or an

unresolved material product decision.

• When an explicit loop budget is exhausted, stop with the latest judge

evidence instead of silently expanding the budget.

• Never make a failing check pass by weakening assertions, deleting coverage,

hiding errors, or silently changing acceptance criteria.

• Do not query production, deploy, migrate, merge, or publish unless the user

explicitly authorizes that action.

If a skill instruction causes a pause, cite the exact file and instruction,

explain the unmet requirement, and check whether existing authorization

already satisfies it. Complete independently actionable work before handoff.

Finish with evidence

Report:

1. behavior implemented or defect fixed;
2. changed files and important design decisions;
3. regression proof or acceptance evidence;
4. focused and broader commands with outcomes;
5. task-contract and code-quality findings fixed or explicitly unresolved;
6. unrun checks, residual risks, and blockers.