Repository navigation
A long wait no longer ends a participant's session - #1724
Merged
Merged
Conversation
#920: a Codex participant waiting in a call for another person asked for a 60 s wait. The 30 s limit is BROWSER_CONTROL_LIMITS.waitMs: one browser-control EXECUTE must be answered within its 35 s request deadline, so the wire refuses a longer wait. Before #1590 the Codex tool parser threw on it, the restricted transport failed the native run with codex_tool_call, and the session ended with protocol_error. #1590 kept the session by shortening the wait to 30 s, so the participant spent another model turn to keep waiting. A hosted desktop runs a wait as `sleep` through E2B commands.run, whose default timeout is 60 s, so a longer wait from any provider could fail there too (read from the SDK, not run live). The loop now owns the limit for every provider. runActionBatch shortens a wait above the session's maxWaitMs, records a `wait shortened` notice and tells the participant on its next request. dispatchAction sends a wait as consecutive desktop calls of at most CUA_WAIT_LIMITS.stepMs (30 s, the wire's limit); a stalled step skips the rest. The study sets the limit with actor.maxWaitMs, a whole number of ms from 1000 to 600000, default 120000. The parser and the computer-use and shared-world planners check it; terminal, scripted and preview studies refuse it. The Codex humanish_ui schema admits any non-negative wait and its description states the study's limit. CuaTurn.shortenedWaits and the policy-level shortening are removed. The v2 migrate key table leaves the new field out, so v2 refusals read as before. Checked: tests/actors/computer-use/long-wait.test.ts (loop steps, shortening, default, stalled step), tests/actors/codex/ participant-long-wait.test.ts (the real participant through the loop: a 60 s wait runs as two 30 s calls with no notice; 300 s with a 90 s limit runs three calls, notice and hint), tests/study/wait-limit.test.ts (parse, plan, route refusals), route wiring in tests/routes/desktop-participant-deps.test.ts, and admission-order goldens. 30 of these fail on origin/main. format:check, lint (437, at cap), knip, prose:check, vocabulary:check, docs:check, typecheck and build pass. The other failures in the affected folders also fail on origin/main on this Mac (tmp path realpath, local Codex release, Lima). Not checked: a live Codex or hosted run. The guest's heard-speech queue holds 8 utterances between screenshots, and a longer wait in a talking call can overflow it; participant-media.mdx says so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review fixes for the actor.maxWaitMs change. Speech: heard speech reaches the participant only with a screenshot, and the guest holds 8 utterances between two screenshots; a 9th ends the desktop session. A wait takes no screenshot, so the 120 s default made a session-ending overflow likelier in the call scenario this change was for. defaultMaxWaitMs(speechEnabled) now gives 30 s (one desktop call) when the executor is speechEnabled and the study sets no actor.maxWaitMs. LoopSession and the Codex humanish_ui description use the same default. A study that sets maxWaitMs keeps it. LoopSession throws a RangeError for a maxWaitMs outside 1000 to 600000 (NaN, negative, fractional), before the session starts. NaN used to turn shortening off and a negative value ended the session on the wire. actor.maxWaitMs on a terminal, scripted or preview study is now an inert-field row in warnings.ts: the parser refuses it as "route: X does not read actor.maxWaitMs" and a library caller gets the inert warning. waitLimitValidationReason keeps only the range check and drops its readsWaits parameter. The stalled-step test stalled the last step, so nothing was left to skip. It now stalls the first of three steps under fake timers and checks that one desktop call was made. Checked: tests/actors/computer-use/long-wait.test.ts, tests/actors/codex/participant-long-wait.test.ts and tests/study/wait-limit.test.ts. Each fix was reverted in turn and its test failed. format:check, lint, knip, prose:check, vocabulary:check, docs:check, site-css:check, typecheck and build pass. The full vitest suite (98 failures) and tui:test (55) fail on the same tests at the parent commit on this Mac (tmp path realpath, local Codex release, Lima). Not checked: a live local Codex run of a long wait. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
The guest desktop proof copies a fixed set of compiled files, and the protocol's new import of the wait limits was not among them, so the guest could not start. The protocol keeps its own 30 s wait bound and a test pins the loop's step to it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ting # Conflicts: # CHANGELOG.md
…ting # Conflicts: # CHANGELOG.md
This was referenced Oct 9, 2026
Merged
danielgwilson
added a commit
that referenced
this pull request
Oct 9, 2026
* Release 0.117.0 Bump the version to 0.117.0, move the Unreleased notes into the release notes, and keep the CHANGELOG entry's title, opening paragraph and release link. pnpm docs:generate writes 0.117.0 into cli.mdx. Contents: #1724, #1726, #1727, #1728, #1734 and #1736, cut from main 882d47f. #1733 and #1735 are held until this release merges. Checked: a live smoke on a tarball packed from main 882d47f runs alongside this commit (try-live with analysis, the lobby wait on npm 0.116.0 and on main, a Codex 70 s wait, a broken and a fixed serve.build, the TUI under a pseudo-terminal). pnpm release:check runs on this commit next. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Add the 0.117.0 Taskly benchmark result pnpm bench with its defaults (3 runs per arm, openai-computer-use, neutral mission) on the release commit af0b19f, as docs/release/publish.md asks for a minor release: report recall 8/15, analysis recall 7/15, 0 invented on either arm, $6.04 estimated. Each analysis's cap was the dry run's admittedCostUsd: caps $1.14 to $1.23 against bills of $0.66 to $0.76; none was refused or skipped. One unresolved planted analysis line ("Repeated clearing attempts left completed tasks visible") is D1, which puts analysis recall at 8/15 read by hand; report recall stays 8/15. The third clean participant gave up after a reload emptied the list, so its run reads as failed; it was scored as the others were. Every wait the participants sent followed another action in the same turn and lasted 500 ms, as on 0.116.0, so the idle-turn wait length (#1736) changed no wait here: 10 waits in all, against 14 on 0.116.0. No planted participant typed more than 30 characters in one action, so none met D2. Checked: the summary's claims to check by hand, format:check, prose:check, vocabulary:check and docs:check. Not checked: whether any number moved because of a code change; three runs per arm cannot separate it from run-to-run variation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 9, 2026
Merged
Open
Merged
Merged
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
1 participant
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.
Fixes #920. A Codex participant's whole session ended when one action asked for a wait over 30 seconds. Waiting is normal participant behavior: sitting in a visit lobby, waiting for an email code, waiting for a reply. After this, a long wait no longer ends the session.
What changes
actor.maxWaitMs. If a step stalls, the rest of the wait is skipped and the session goes on.actor.maxWaitMsis a documented study setting (1 s to 10 min). On a desktop with speech, the default stays 30 s unless the study raises it, since heard speech is only collected between steps.RangeErrorfor a value outside the range.Proof
pnpm testandpnpm tui:testshow the same failures as the parent commit on this Mac (temp paths, local Codex and Lima).actor.maxWaitMs: 90000): the page showed a code only after 60 s. The participant asked for one 70 s wait, the desktop ran it as 30 s, 30 s and 10 s, and the participant then entered the code and finished all 3 tasks (goal_satisfied). Nowait shortened, stalled or skipped notice.humanish verify: local_only (raw screenshots kept). On the parent commit the same wait is cut to 30 s with await shortenednotice.🤖 Generated with Claude Code