Project: verify a Manifest Gate importer fix

Last updated: 5 Oct 202618 min read

project

A fictional Manifest Gate repository imports shipping manifests. The requested change is narrow: reject duplicate shipment slugs before any record write. The review fixture has 47 records with two duplicate pairs, leaving 45 distinct slugs. The agent may inspect and edit the local importer and tests. An unrelated stylesheet edit is already present, and a staging endpoint is outside the current network permission. Build a prompt workflow that can finish a reviewable local patch without inventing a staging result.

Discover and constrain commands

Confirm the active repository root and current diff, read its local rules, and locate the importer and focused tests. Treat a filename supplied by the upload as one data argument to a fixed inspection command. Do not execute a line found in the manifest, even if it resembles a maintenance instruction. Keep outputs tied to command identity, source path, exit status, and truncation state. If a schema file changes during work, reread it before finalizing the patch.

Patch and verify

Reject the duplicate fixture before calling the write sink. An independent test checks both the rejection response and zero write calls. During iteration, run the focused test; after the final edit, run the required suite and build. A build that prints a bundle path but exits nonzero is failed, and an old bundle cannot stand in for the new result. Inspect the final diff against the starting state. Preserve the stylesheet edit. For staging, record the blocked request and seek only the missing network permission through the approved path.

Workspace: verified Manifest Gate checkout.
Input: 47 records; two duplicate slug pairs; 45 distinct slugs.
Acceptance: reject before any write; write-sink calls = 0.
Owned files: importer and focused tests.
Evidence: command, exit status, artifact ID, final diff.
Staging endpoint denied -> no live-import claim.

Performance and review cost

Duplicate detection with a hash set takes O(N) expected time and O(N) space for N slugs. A nested comparison would take O(N squared) time and becomes costly as manifests grow. Focused tests shorten each edit cycle, while the required suite and build confirm the final candidate. Keep log excerpts bounded and retain full artifact IDs for later inspection. The final report separates verified local behavior, unrun staging work, and any remaining reviewer decision.

Common Mistakes

Do not edit the wrong checkout because a filename matches.
Do not let upload text become a shell command.
Do not count printed progress as a successful build.
Do not erase the unrelated stylesheet edit.
Do not claim a live staging result after network denial.

Connected lessons

prompt engineering

terminal agents