docs: demonstrate server-owned tool approval state - #1951
Conversation
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
markstuart-oai
left a comment
There was a problem hiding this comment.
Reviewed the complete diff at 1a0dbb0. The shared server example keeps snapshots private, binds decisions to stored interruptions and the authenticated owner, validates a complete boolean batch, and consumes the request before awaited restoration or execution. The tests cover tampering, mixed decisions, repeated pauses and replay across concurrent submission, failure and cancellation. The guide accurately separates this single-process demonstration from production authentication, shared storage and recovery. No actionable correctness or structural findings.
Exact-head hosted CI has 16 successful checks and one skipped deployment, including passing Node 22/24.3 tests, coverage, build and Windows checks. Source review only; I did not run local tests, a live model or a deployed approval service.
This pull request updates the human-in-the-loop guide to clarify that deserializing RunState does not authenticate the snapshot or authorize the reviewer.
It adds a CLI client/server simulation that keeps snapshots on the server, checks ownership, validates boolean decisions against stored identifiers, and consumes each approval request before resuming execution. Regression tests cover unauthorized submissions, mutation, concurrent requests, replay, failure, and cancellation.
The guide explains production requirements for authentication, atomic shared storage, sensitive display data, and recovery after uncertain tool side effects. SDK APIs and serialization remain unchanged.