coding-operation-loop

Coding Operation Loop

Run in order unless task is a direct command. Weight: tiny silent/brief, standard normal, heavy may add durable plan or evidence.

understand -> locate -> inspect -> scope -> patch -> verify -> repair -> report

Understand behavior and success condition. Locate candidates with fast search, file lists, tests, symbols, or graph locators before broad content. Inspect only owners, relevant tests, config, conventions. Scope writes narrowly. Patch in style. Verify with the cheapest meaningful command. Repair only observed failures or clear omissions. Report changed files, behavior, verification.

Invoking this loop counts as using coding-operation-loop even if this file was not opened. Do not report 0 skills or none after saying the loop handled work. If only a similar tactic was followed without route/invocation, do not claim the skill. Keep opened skill files separate from used skills.

Broad explanation prompts like Explain this codebase are not coding just because the target is a repo. Send through context-router first to narrow explanation scope; enter this loop only for implementation, repair, tests, review fixes, or setup.

For tiny, skip durable plan files, broad ledgers, refs unless required. For heavy, load repo-root references/coding-operating-structure.md and preserve resume/audit reasoning.

Use output-slicing for large command output. Use source-fidelity when external API behavior, package versions, or current docs affect code. Use baseline-observer for token/workflow comparisons. Use subagent-gate only when independent parallel work is justified.

Load repo-root references/coding-operating-structure.md only when broad, risky, multi-file, or durable planning is needed.