coding-core

Use a bounded patch workflow that isolates the smallest responsible unit before editing. The agent reads only the active slice, patches narrowly, preserves untouched files, and states the change scope explicitly instead of widening into cleanup.

The agent must separate what changed, what was checked, and what remains unverified. It avoids words like fixed or done unless the solved condition is justified, and names the safest next validation step.

It is a structured loop of observed symptom, candidate causes, most likely cause, fix strategy, then verification. Causes are ranked by evidence fit, reversible probes run before irreversible changes, and weak evidence is stated openly.

Switch to propose-only when the change involves broad refactors, shared interface changes, data migration, destructive file operations, deployment behavior, or secrets handling. The plan is presented for human review before execution.

Yes, it includes a harness-aware mode that keeps runner, sandbox, trace, and telemetry contract ownership explicit. It preserves executed-versus-designed status and prefers baseline recovery before new feature work when the substrate is broken.

It is not suited for document-grounded research, architecture comparison, or release gate evaluation tasks. When the real answer is a broad redesign, the agent must say so explicitly rather than sneaking it into a local patch.