Verification
Checked 2026-09-26. Six synthetic migrations and eight generated-skill response probes passed. This demonstrates portable instruction behavior in Work Mode, not live GPT parity or native execution in every destination.
Backup-first extension, 2026-10-01
Section titled “Backup-first extension, 2026-10-01”The private-backup helper and current packaging pass 37 synthetic tests in both normal and optimized Python, schema and skill validation, strict plugin checks, a clean local install, the 38-page docs build, and the original 11-check docs gate. Current validation record
These checks do not rerun the historical model evaluations below. The revised workflow still needs a hand-test with an owned GPT and the selected Web account; use the hand-test checklist.
Behavioral evidence
Section titled “Behavioral evidence”Each migration used a fresh agent given the source configuration and migration skill. Another fresh agent received only the resulting skill and task inputs. The grading requirements were withheld from both runners. An independent reviewer compared the retained artifacts and responses with those requirements.
| Case | Demonstrated result |
|---|---|
| Instruction-only meeting brief | Preserved headings, word/question limits, explicit owners, and request for missing notes. |
| Missing knowledge | Copied the available policy accurately, cited it, and kept the missing fee unknown. |
| Actions | Preserved the operation/auth/error contract and returned an unsent draft without inventing a connector or reservation. |
| Conflicting instructions | Preserved both incompatible formats and asked which controls before producing a summary. |
| Strict JSON | Returned exact valid and invalid objects with the required keys, values, and JSON types. |
| Unavailable image capability | Preserved PNG requirements, labeled the text fallback, and made no image or Mac-installation claim. |
No conversion failure was found in these six cases. Missing integrations and unresolved source conflicts correctly remain partial or blocked. The run exposed a verification gap in the earlier guidance: conversion inspection and execution of the generated workflow needed separate evidence. The revised instructions require expected requirements and retained observed outputs.
Two reruns using the revised guidance completed their own separate execution checks: three meeting-brief cases and eight strict-JSON cases, all passing. Those additional cases include missing input, invalid types, decimals, zero, negative values, and integer arithmetic beyond JavaScript’s safe integer range. Their evidence is retained alongside the baseline, for eight conversions and 19 observed responses in total. This small sample is not a reliability rate.
The evaluation set and run procedure include synthetic fixtures, generated files, response bytes, hashes, and the independent assessment. The replay checker checks recorded evidence; it does not call a model. Baseline reports predate the separate response runs, so their then-unrun notes are retained as history. The September 17 single migration trial did not execute its generated workflow; the new suite adds that evidence.
Installation and structural checks
Section titled “Installation and structural checks”The official skill validator passes for all five canonical skills. Claude Code
2.1.283 strict marketplace and plugin-manifest validation pass with no warnings
or errors. Codex 0.154.0-alpha.3 has no plugin validate command; its read-only
app-server plugin/read loads all five skills from the repository marketplace.
That loader result is distinct from automatic activation or a personal install.
Skills CLI 1.7.0 installs only migrate-cgpt into a clean disposable Codex project
from the local candidate checkout. The installed files match the canonical
source byte-for-byte. This checks project packaging, not remote-default-branch
availability or installation on the user’s Mac or ChatGPT account.
Documentation gate
Section titled “Documentation gate”The documented Bun setup, three documentation tests, and Astro build pass.
The build emits 37 pages and all three llms*.txt files with the changed
migration content. The original docs-spec checker passes 11 checks with zero
warnings or failures. It was recovered through its documented installation
layout, not replaced with a substitute. See apps/docs/README.md for setup and
the evaluation validation record for the recorded gate result and generic setup.
No documentation deployment was performed. Existing build notices about deprecated Markdown processor options and the missing custom 404 entry are nonfatal and outside this migration change.
Limits
Section titled “Limits”One run per case is a smoke evaluation, not a reliability estimate. The runner’s exact model and effort were not exposed by spawn results. Some probes shared an agent context and were instructed to act independently. Explicit invocation does not test automatic trigger selection, including nearby unrelated requests.
Live Actions, authentication, 401/409/timeout handling, original-GPT before/after parity, hosted migration, native ChatGPT/Mac execution, and production image generation remain untested. Error rules were inspected in the generated Action contract; only the unavailable-integration fallback was executed. A supplied OpenAPI extract is not a working service or a full OpenAPI conformance test.