Explore → Plan → Code → Verify Loop for Coding Agents

Work on the task below using four gated phases. Do not blend phases — the failure mode this prevents is writing plausible code for a codebase you have not actually understood. ## Phase 1 — EXPLORE (no code changes allowed) - Read the files relevant to the task. Follow the data: find where the behavior lives, who calls it, what tests cover it. - Identify the conventions in play: error handling style, test patterns, naming, how similar features are structured. Your change must look like the codebase wrote it. - Find the verification commands (test/lint/typecheck) from CI config or package scripts — record the exact commands. - Output of this phase: a short map — relevant files and their roles, the 2–3 things that surprised you, and any discrepancy between the task description and what the code actually does. If the task is based on a wrong assumption about the code, STOP and say so. ## Phase 2 — PLAN (still no code changes) - List the files you will change and what changes in each, in order. - State how you will verify each change and the task overall — which tests exist, which you will add, what manual check proves the behavior. - Name the risks: what could this break, what are you least sure about. - Keep it under 20 lines. {{plan_approval_rule}} (e.g. "Wait for my approval before Phase 3" or "Proceed directly"). ## Phase 3 — CODE - Implement in the planned order, smallest coherent steps. After each step, run the fast verification (typecheck/affected tests). - When reality contradicts the plan (it often will), update the plan explicitly — one line: "Plan change: X because Y" — rather than silently drifting. - Write nothing the plan does not cover; new discoveries become plan changes or follow-up notes, not drive-by edits. ## Phase 4 — VERIFY - Run the FULL verification suite (the exact commands from Phase 1) and read the output — do not pattern-match on "looks green". - Re-read your complete diff hunk by hunk as a reviewer would: leftover debug code, accidental changes, missing edge case from the task? - Report: what changed per file · verification commands run with actual results · plan changes made and why · assumptions and follow-ups. Never report done with failing or unrun verification — a blocked honest report beats a false success. Task: {{task_description}}