Repository navigation
fix: [SDK-5443] name the structured-question tool and its Codex fallback in skill text - #60
Merged
Merged
Conversation
kalley
approved these changes
Oct 12, 2026
kalley
left a comment
kalley
left a comment
Collaborator
There was a problem hiding this comment.
The §14 rules read clearly, and the session table makes the app/CLI split easy to follow. One thing inline I'd fix before merge.
abdulraqeeb33
approved these changes
Oct 12, 2026
5 of 10 tasks
…ack in skill text The skills asked for "the harness's native structured-question tool" and pointed to safety contract §14 for the name. Codex loads SKILL.md and reads a reference only when the model follows the link, so the tool name was one hop away. In the Codex CLI the questions showed as free text with no reply instruction. Name AskUserQuestion and request_user_input inline at the first ask in each skill. In §14, add the call rules both hosts accept (2 to 3 options, 3 questions per call, recommended option first, no own "Other"), and the Codex Default-mode path: use request_user_input_async when the sync tool is absent, then end the turn with a plain-text line that names the reply words. The Codex app shows the async call as a question card; the CLI shows text. Make the setup platform pick and the verify test-send consent structured questions. Add a README note for Codex users.
…d verify asks as structured The consent question keeps the listed order, with send first, and has no recommended option. The verify message ask is a second question in the same tool call, because a yes/no structured question cannot take a free-text body. The setup dirty-tree and wired-App-ID asks now say structured question, the same as the platform pick.
sherwinski
force-pushed
the
sherwin/sdk-5443
branch
from
October 12, 2026 18:20
b6a8361 to
e977c1b
Compare
A 1-option question breaks the §14 rule of 2 or 3 options, and AskUserQuestion rejects it, which fails the whole batched call with the send consent. Add "I'll type my own message" as the second option.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
One Line Summary
Name
AskUserQuestion/request_user_inputinline in each skill, add the call rules and the Codex Default-mode fallback to safety contract §14, and document the Codex behavior in the README.Linear: SDK-5443
Motivation
Dogfood feedback: in Codex, the plugin's questions show as free text instead of selectable choices, and sometimes the choices appear and then disappear.
The skills asked for "the harness's native structured-question tool" and pointed to §14 for the name. Codex loads
SKILL.mdand reads a reference only when the model follows the link. In the Codex CLI, Default mode lists onlyrequest_user_input_async; the model called it (9 times in the Oct 6 dogfood session), the CLI rendered the call as text bullets, and the follow-up prose named no reply word. Users typedyesto a send/keep question.The Codex app renders the same async call as a question card, so the card already worked there. A selectable prompt in the CLI needs the sync
request_user_input, which Codex lists only in Plan mode, and Plan mode does not edit files. No skill text can change that.Scope
references/safety-contract.md§14: call the tool by name; 2 to 3 options and 3 questions per call; recommended option first with "(Recommended)"; no own "Other"; in Codex Default mode userequest_user_input_asyncwhen the sync tool is absent, then end the turn with a plain-text line that names the reply words; never continue past{"accepted":true}.skills/setup/SKILL.md,skills/credentials/SKILL.md,skills/verify/SKILL.md: the first ask names both tools inline. Setup's platform pick and verify's test-send consent are structured questions.README.md, "Install in Codex" → "After install": the app shows a card, the CLI shows text with a reply word, Plan mode does not edit files.Not in scope: the
.onesignal/dirty-tree gate found during testing, tracked in SDK-5470. No version bump: 1.1.1 is already past the last tag (1.1.0).Testing
Local checks:
package_directory_bundle.py --checkpassed; the portability grep printed nothing;py_compileandjson.toolpassed on the 4 manifests.Setup skill against a Kotlin Android fixture (
SkillDemo3), plugin installed from the local clone throughcodex plugin add onesignal@onesignal, cache diffed against the tree before each run:01a112de-25c4-73a2-8c76-909a9edd2d54request_user_input_asyncat every gate; bullets then prose with no reply word01a12320-814a-7dc3-93c2-93c78a92cd6f01a12321-479d-7261-bba9-f4494430d15501a1235f-dede-7962-8998-b108c9a919f0sendcontinued the run through detection and version pin01a12360-4eb1-7113-870f-609d84f996e201a12368-0736-7602-889a-ada939403150request_user_input; selectable widget; model declined to edit files in Plan modeNot run: the behavior evals and Claude Code. The change touches ask wording only, not the ask points, the gates, or the checkpoint calls.
Affected code
skills/)references/)Checklist