gm-emit — EMIT

EMIT is where intent becomes artifact. The phase exists as a distinct gate because writing without verification produces silent drift between what the agent believes was emitted and what landed on disk.

Pre-Emit

Before writing, debug the planned state. Read the target paths that will be touched — confirm current contents match the assumption the diff is built on. A diff applied to a file the agent has not freshly read is a diff against a stale model. Spool the reads if scope is wide; serial Read calls are acceptable for a small set. Surface mismatches → snake to PLAN.

Sync-Before-Emit

rs-codeinsight and rs-search outputs feeding EMIT must come from a freshly-completed index. No cache serves a result without a digest match against the live filesystem. Default invocation always runs fresh. --read-cache is permitted only when .codeinsight.digest matches exactly; on mismatch, the cache auto-refreshes before the result emits. Emitting from an unverified or partial index is forced closure equivalent to bluffing strength — the agent reads stale output as ground truth and acts on a state that no longer exists.

Write

One Edit or Write per artifact. No multi-file batches that conceal which file failed if one fails. Spool larger payloads through in/nodejs/ when shape demands it.

Post-Emit Verify

After each write, re-read the file from disk and assert the change is present. The Read tool is the post-emit witness. Discrepancy → Fix on Sight: fix at root, re-emit, re-verify. A green Write call is not the witness — the verified disk state is.