# GPTsApp (GPTs App)

> GPTsApp provides an evidence-led Codex tool diagnostic called GPTsApp Doctor. It is an independent service, not affiliated with or endorsed by OpenAI.

## Canonical task surfaces

- https://gptsapp.io/diagnose — Diagnose the first supported failing capability layer and receive one bounded next check
- https://gptsapp.io/collect — Build a local sanitized Review Pack and optional screenshot annotations
- https://gptsapp.io/evidence — Inspect the evidence policy, current official source context, and honest review-state counts
- https://gptsapp.io/codex/mcp-not-working — MCP startup, connection, auth, registration, exposure, permission, and execution
- https://gptsapp.io/codex/skill-not-loading — Skill/plugin discovery, parsing, exposure, and activation
- https://gptsapp.io/codex/works-in-cli-not-desktop — Surface-specific exposure comparison
- https://gptsapp.io/codex/broke-after-update — Version-bound regression evidence

## Local tool contract check

- https://gptsapp.io/tools/openai-tool-check — Check one OpenAI function tool against an explicit strict profile, review field-level candidate changes, compare two user-confirmed versions of the same tool and the same argument examples, and export locally. Compare versions reports exact added, removed and changed fields; no structural difference is not a proof of semantic equivalence. No model, server or tool is contacted. Static checks do not prove runtime acceptance; references and unlisted schema features remain explicitly not checked.

## Published reported cases

- https://gptsapp.io/incidents/openai-codex-issue-15508 — Human-reviewed reported case scoped to its listed Codex surface, version, environment, workaround, and verification
- https://gptsapp.io/incidents/openai-codex-issue-30343 — Human-reviewed reported case scoped to its listed Codex surface, version, environment, workaround, and verification

These are human-reviewed reported cases, not OpenAI status declarations, universal compatibility verdicts, fixed-version announcements, or assignments of publisher fault.

## Task contract

- Every result ends with one checkpoint-specific safe check, explicit pass/fail meaning, a stop condition, and a source when the method is documented.
- A structured recheck compares whether the first supported boundary resolved, stayed the same, moved upstream, moved downstream, or remains unknown.
- Attempts form a local troubleshooting timeline in the current browser tab/session.
- A maintainer-ready Markdown handoff identifies the evidence-supported support boundary without assigning fault.

## Evidence and privacy contract

- Raw diagnostic text and optional free-text context are redacted in the browser before submission.
- The Worker performs a second defensive redaction pass on evidence and context.
- Diagnosis is deterministic and explainable; ambiguous evidence returns partial or unknown.
- Diagnosis and recheck are stateless unless the user explicitly saves a private report or exports a local Review Pack.
- Review Packs are generated/imported locally, are versioned and size-bounded, and never include screenshot bytes.
- Scope Links contain only category, surface, OS, recent-change class, checkpoint, rule ID, and route; they contain no evidence, names, versions, notes, attempts, report IDs, or tokens.
- Screenshot preview is local-only; no OCR, upload, retention, or remote image analysis is performed.
- Upstream matches are official method sources or bounded issue candidates from the official repository. Candidates do not increase diagnostic confidence or confirm a current incident.
- No unknown capability, package, URL, or user action is executed.
- Private reports use random URLs, are noindex/no-store, expire after seven days, and can be deleted immediately.
- Deferred candidates, blocked human-reviewed records, and rejected duplicates/corroboration remain noindex and outside the sitemap.