fix(mcp): mark review preflights as writes

[cderinbogaz]

TL;DR

Correct the MCP tool annotations used for OpenAI Plugin review and add the generated submission artifact.

Summary

• Mark review preflights as non-read-only because they persist an expiring confirmation intent
• Add chatgpt-app-submission.json with verified tool justifications and five positive plus three negative test cases

Review focus

• Confirm the preflight annotation accurately reflects the private confirmation-intent write
• Confirm submission copy matches the exposed MCP tool set

Test plan

• Cloud test suite
• Cloud type-check
• Submission artifact coverage validation
• Codex review

[github-actions]

Juror Review

This PR marks review preflight MCP tools as state-changing and adds the ChatGPT app submission manifest with tool annotations and coverage cases.

• Adds non-read-only annotations for both review preflight tools.
• Documents the exposed MCP tool set and its read/write justifications.
• Adds positive and negative submission test cases covering supported workflows and refusal boundaries.

Merge Confidence: 5/5

The changed preflight annotations consistently mark confirmation-intent persistence as non-read-only, and no executable callback or tool behavior is changed.

_{Model votes: GPT-5.6 Luna 5 · DeepSeek V4 Flash 5 → median 5.}

Findings

No publishable findings.

_{Coverage audit: 0/0 raw model findings accounted for → 0 unique findings; none dropped.}

Important Files Changed

Filename	Overview
chatgpt-app-submission.json	Adds the generated ChatGPT app submission metadata, tool justifications, and positive/negative test cases; no concrete concern is evident from the supplied change.
cloud/worker/mcp.ts	Introduces shared non-read-only annotations for the two preflight tools whose handlers persist confirmation intents; no executable flow is altered.

Sequence Diagram

sequenceDiagram
  participant U as User
  participant M as MCP
  participant C as Confirmation store
  U->>M: prepare review or rerun
  M->>C: persist short-lived confirmation intent
  C-->>M: confirmation intent
  M-->>U: preflight details and intent
  U->>M: explicit confirmation
  M-->>U: queue review or rerun

Loading

💸 This review cost $0.03 · 2 models · 2m08s

Model	Harness	Input	Cached	Output	Cost	Source
GPT-5.6 Luna	Codex	17.9k	78.8k	1.8k	$0.01	estimated
DeepSeek V4 Flash	opencode	70.7k	360k	2.2k	$0.02	reported
Total		88.6k	438k	4.0k	$0.03

reported = provider-computed figure returned by the harness.
estimated = token counts × published list price (pricing.json). Long-context tiers applied.
Cached combines cache reads and cache writes; writes bill at a premium, so the first review of a diff costs more than a re-review of it.

Rolling 30d for this repo: $0.03 across 1 PR (avg $0.03/PR)

_{Juror v1.3.3 · reviewed c54afc2 · docs}