The agent authoring loop
Use authority without flooding context
1. Load the compact card
Start with GC_AGENT_CORE_CARD_v0.3. It is generated, token-bounded, parser-checked, and intentionally smaller than the full documentation corpus.
2. Select exactly one task card
Use the task-card set for the current intent. If no task matches, fail closed and retrieve the authoring bundle rather than inventing a workflow.
3. Resolve exact symbols and diagnostics
The generated symbol reference and diagnostic reference are browsing views of the same machine authorities.
4. Plan before mutation
Record expected files, capabilities, obligations, validation commands, rollback, and acceptance evidence. Prefer semantic patches over textual edits when the toolchain supports them.
5. Repair narrowly
A safe repair must preserve policy and obligations, target the exact failing identity, rerun the original command, and abstain when preconditions changed. Never disable a gate to make its diagnostic disappear.
6. Return evidence
Report command vectors, input identities, output hashes, diagnostics, and unresolved gaps. “It passed for me” is not an acceptance artifact.