Repository navigation
Description
Canonical tester-army/e2e downstream review & promotion queue
Wave: 2026-10-17. Owner approval: user's direct request for a massive code-review wave in e2e; this does not grant approval to merge PRs or bypass maintainer policy.
Origin: https://github.com/tester-army/e2e/pulls (37 open PRs at scan time). Coordination: this one issue for this repository; reuse it rather than creating per-PR factory issues. Cross-repo index: https://github.com/kvnloo/oss-factory/issues/14 .
Permissions blocker: attempt to submit a COMMENT review on tester-army/e2e#1067 at 241ba3b4a9d548f15e4bb450d9608b885f663122 returned HTTP 403 Resource not accessible by integration. No upstream reviews were posted by this wave. Do not pretend any comments exist on origin; prefer local gh/OMP credentials or fix app permissions, then refresh heads/threads.
Current promotion candidates (drafts; exact heads)
- Fork PR #13 ↔ origin #805, head
4932a5761ddbf75ea566eb705d787cbb6a02e0e1: classify Playwright detached-then-dispatched as non-retryable. Test uses actual call-log newline escapes. Await CI. - Fork PR #14 ↔ origin #1151, head
9614f8278718692c5ca35834560073f8aa5eb601: bounded digest snapshot and overflow cleanup; isolated smoke checks passed, await CI. - Fork PR #15 ↔ origin #958, head
ef5a825999ff84b595bde460bd57f23b701f2b51: abort-safe file chooser; await CI. - Fork PR #16 ↔ origin #1209, head
9d8b0ff12f099fa353e1ab035dd3eb12659875d4: maintainer-requested lifecycle comments, worker reuse revocation, and mobile recording reset docs. Await CI and reserved maintainer policy decisions. - Factory registration draft → canonical registry (not upstream).
Origin reviews/merges by this wave: none; review connector returns HTTP 403. Hermes can cherry-pick targeted commits into existing contributor branches only with appropriate authorization. Do not auto-approve or merge.
Source-grounded review questions for origin PR e2e-inc#1067
Current target: tester-army/e2e#1067, draft, author szdziedzic, inspected head 241ba3b4a9d548f15e4bb450d9608b885f663122. Official Expo REST contract: https://docs.expo.dev/preview/simulators/reference/rest-api/ (published October 16, 2026).
Verified from official documentation: REST POST /v2/device-run-sessions, GET /:id, POST /:id/stop, lowercase status values, and a 200 JSON data session response for stop are documented. Do not repeat the prior speculation that the endpoint might not exist or that stop should accept an empty 2xx body. That prior claim is superseded.
Two narrower unresolved implementation-contract questions:
- Authentication: Official REST documentation requires
Authorization: Bearer <EXPO_TOKEN>on every request, butpackages/eas/src/client.tsstill sends onlyexpo-sessionfor the documentedeas loginfallback. The new fake-server tests accept it. Ask whether this endpoint supports an undocumented session-secret mode; if not, the existing no-token login workflow needs explicit migration handling or a supported bridge. - Client attribution: PR sends
client: '@e2e-dev/eas@version'in create payload, yet official request fields don't listclientand documentation says unlisted fields are ignored. Ask how server-side client attribution is registered/verified; don't claim it works from stubbed tests alone.
These are CHECK (source-contract conflicts needing maintainer/server confirmation), not proven live failures. Do not create a competing PR or expense-generating session without authorization. Proposed one-comment review:
I compared the REST migration with Expo's published Simulators REST reference. It confirms the endpoints and stop response shape. Two compatibility questions remain: the reference requires Bearer token auth, while the provider continues to support
eas loginwith anexpo-sessionheader; does the REST API accept that undocumented mode? Also, the create payload'sclientfield isn't in the documented request contract (unlisted fields are ignored); is it supported for client attribution? The test stub currently accepts both. No live Expo account was used in this review.
High-value existing reviewer requests (do not repost others' feedback)
These are triage pointers, credited to the existing discussions/maintainers, not findings originating in this wave:
| PR | Current review target / action |
|---|---|
| #805 | Maintainer @okwasniewski flagged potential duplicate form submit after Playwright detach log then elementHandle.press(): DETACHED_PATTERN checked before post-dispatch evidence. Resolve with targeted error-log fixture and prove no retry after a possibly committed Enter. |
| #1151 | Maintainer reverified thinking signature fix; remaining cleanup: revert no-op identity-preserving shrink helper and consider bounded hashed snapshots vs ~20MB stored message copies. Respect maintainer guidance. |
| #1165 | Secret-redaction bug reproduction and fixes credited to @okwasniewski/@Anmol202005; last maintainer review asks move stored-text logic into a module, revision-based refresh, remove WeakMap side-channel if possible. Treat security-sensitive specifics privately under SECURITY.md. |
| #1209 | Maintainer and bot already reviewed reused mobile lease; last maintainer requests restoring deleted JSDoc, simplified retirement semantics, recording reset behavior. Bot flagged recovered VideoSegment discarded. Check latest author response before replying. |
| #1206 | Large PR-impact filter. Maintainer @okwasniewski explicitly left human design decisions: throwing custom filter fails run vs fail open; AI_GATEWAY_API_KEY on CI testbed and PR-run scope; no automated resolution. |
| #1200 | Maintainer requested separate testmu web package, remove optional dual-peer setup and unused routing options; live provider verification outstanding. |
| #1177 | Awaiting maintainer keepResults API design and archive lifecycle redesign; no new superficial review. |
| #1154 | Author and maintainer agreed current patch does not fix issue e2e-inc#1144 under defaults; wait for API/budget direction rather than approving. |
| #677 | Maintainer asked split email/maildev core vs agentmail and refresh branch against main; avoid re-asking. |
| #648 | Large service lifecycle refactor; maintainer already fixed port lease release ordering at 4ab12fd9; verify races after changes before additional comments. |
| #960 | ACP integration: Cursor adapter requires live verification and manual initial npm publishing/trusted publishing. |
| #1214 | Refusal taxonomy: active maintainer review already covers typed failure, malformed provider output and tool-loop path; avoid duplicate. |
| #1056 | Active maintainer re-review covers serial retry verdict mutations and timing; focus on unanswered head-specific issues only. |
Post-repair fast-track (refresh CI, exact head, and unresolved threads first)
Maintainer or contributor notes report fixes for #996, #1048, #1046, #1049, #1050, #1072, #1191, #1207, #871, #900, #1037, #1041, #1044, and #961. Not approvals; review discussion and CI must be refreshed before any comment. Cross-PR conflict to track: e2e-inc#1041 vs e2e-inc#1046 docs; e2e-inc#1049 vs e2e-inc#1037 LocatorExpression.
CR wave rules and next steps
- Refresh all 37 open PRs and changed heads, group by risk (security/correctness > lifecycle > API > performance > docs) and avoid duplicating maintainer or bot comments. Credit the original reporter and author.
- For each candidate, inspect the patch and current maintainer review. Reproduce a defect with a minimal test before posting an assertion; do not submit speculative security issues publicly.
- Submit one grouped COMMENT review per head only for fresh, independent actionable findings; never APPROVE or REQUEST_CHANGES automatically; stop if maintainer flags review noise.
- Record each posted review permalink and author response in this issue. On 403 do not retry the integration repeatedly; use approved local GH/OMP authentication for posting.
- Prioritize the two specific
#1067REST authentication/attribution questions,#805duplicate-submit safety, and code/CI repair on existing branches over another new PR.
OMP handoff if GitHub connector still denies origin writes
Run an evidence-first code-review wave against tester-army/e2e. Read AGENTS.md, CONTRIBUTING.md, SECURITY.md, and the canonical kvnloo/e2e downstream review queue (this issue). First check current author, SHA, and existing review threads for all 37 open PRs. Prioritize #1067 (documented REST bearer auth versus expo-session, and undocumented client attribution; confirm official Expo REST reference before posting), #805 (Playwright detached-then-press post-dispatch retry can submit twice), #1209 (recovered recording segment), and #1151/#1165 (maintainer-requested cleanup). Do not repeat existing findings, and do not claim speculative issues reproduced. Use local authorized gh credentials to publish one consolidated, focused COMMENT review only when newly verified and nonduplicate. NEVER auto-approve/merge. If origin rejects writes, update the same downstream issue with exact blockers and tests; do not create duplicate queue issues. Credit original authors, maintainers, bot reporters and testers. Report PR links, SHA, tests performed, review URLs, and unresolved risks.
No project-local tests ran in the ChatGPT environment: the review uses GitHub's live PR diff/timeline; upstream origin review submission was blocked by GitHub integration scope.
Activity
kvnloo commentedon Oct 17, 2026
Dependency/CR wave 2 — 2026-10-17
This is a single update to the existing canonical queue; do not open another per-repo promotion issue. Read AGENTS.md, CONTRIBUTING.md, and the earlier review/403 evidence above before posting upstream.
Completed, head-specific work
5bfc777e44f448d2d0f521feb2cb4db6588f4b40:a04bf1d: all four changesets are patch under the repository's 0.x policy.d1c5685: exclude horizontal-only web scrollers from up/down decision candidates. The public engine direction grammar supports left/right too, but this decision loop currently offers only up/down; a horizontal-only target can never progress.5bfc777: regression fixture asserts a horizontally overflowing element does not advertise vertical scrollability.Live dependency map
Important compatibility gate: fork PR #5 still targets historic
mirror/decision-775-a1c14b75, while originmainatff3bf3917f2b5bb0f81e5bc5f4fdaa24d023afd0uses the substantially refactored e2e-inc#965 decision model (controlsis aSet, the dispatch layer uses discriminated actions). Do not retarget the old PR and assume it will merge. Keep the lab as evidence, and port a small semantic-signal + scroll-selection slice to current main after checking maintainer appetite.Priority / promotion policy
infinite-scroll,virtualized-table,sticky-chromefrom the existing benchmarks against main before an upstream PR. If red, fix the first root-cause failure in a small commit, not a broad rewrite.Review publication gate
The earlier wave's COMMENT review to origin returned 403 Resource not accessible by integration. This wave did not retry origin writes or post review noise. Use locally authorized
gh/OMP to post a single grouped, evidence-backed COMMENT review per current head if useful and within maintainer preferences. Keep all links, failure evidence, and attribution here.kvnloo commentedon Oct 17, 2026
Code-review execution receipt (Oct 17, 2026)
Three isolated downstream draft PRs were authored against exact upstream PR heads, with original author/reviewer credit. Nothing was posted or merged to origin; GitHub reviews remain blocked by integration 403.
212194de285737a9ea8bcf1a6edad0e066059c7a9614f8278718692c5ca35834560073f8aa5eb601dc03fcddb9029396fe929b6927cd80222f11a183Important correction of the earlier e2e-inc#1067 note (supersedes its broad API-shape concern): The official Expo Simulators REST API reference, updated October 16, 2026, confirms
POST /v2/device-run-sessions,GET /v2/device-run-sessions/:id,POST /:id/stop, JSON{data: ...}, lowercase statuses, and a 200 JSON session object for stop. The prior suggestion that an empty-body stop response might be expected is refuted. Do not post that old suggestion. Source: https://docs.expo.dev/preview/simulators/reference/rest-api/ (Create / Get / Stop).New independently grounded e2e-inc#1067 review points (source-only, no live credentials):
Authorization: Bearer <EXPO_TOKEN>; a request with no Authorization header gets400 UNAUTHORIZED_ERROR. Existingpackages/eas/src/credentials.tsandclient.tsstill promise and send anexpo-sessionheader for aneas loginsession-secret fallback. If only the documented contract is supported, this path fails; the test stub accepts it instead. Ask whether Expo privately supports session-secret auth for these REST endpoints; otherwise preserve behavior with a supported auth route or document/validate a token-only migration before release.client: '@e2e-dev/eas@version'on session create. Expo's REST reference enumerates request fields and explicitly says unlisted fields are ignored;clientis not listed. Thus the promised EAS client attribution needs an authoritative server extension or a different supported field/endpoint before it can be claimed. Ask for confirmation rather than assuming a production bug.e2e-inc#1067 current inspected head:
241ba3b4a9d548f15e4bb450d9608b885f663122. Previous generic "unknown endpoint/shape" warning is SUPERSEDED by official evidence. The two narrower concerns are CHECK/HOLD for clarification, not a reproduced outage.Hermes promotion commands / review policy
Read each linked fork PR and its exact target upstream head; validate CI and latest review threads again. Do NOT open duplicate origin PRs. For #13/#14/#15, if green and no reviewer objection, promote individual commits onto the authors' existing upstream PR branches only with their permissions/consent, or leave a single polite review linking the precise patch. Do not push to someone else's branches without authorization.
ghon the user's local Hermes credentials may bypass the connector's 403; stop and report if it does not. Never APPROVE or merge upstream. Credit original authors and issue discoverers. On any changed origin SHA, rebase/recreate the patch first and rerun focused tests.Execution limits: no full pnpm monorepo install, Chromium or live EAS session in ChatGPT sandbox (no network). No claims of CI success or live Expo REST verification.
kvnloo commentedon Oct 17, 2026
Exact-head refresh: downstream PR #13 is now
4932a5761ddbf75ea566eb705d787cbb6a02e0e1, superseding previous212194d…. A follow-up commit corrected seven Playwright log newline escape sequences in its unit regression. Origin target head remainsdc80df2424294f3514ef8a1c1f6e2ee04db1b119. Spec/benchmark/mobile CI triggered for the new head but not yet green. Do not cherry-pick the stale test commit without the correction. Registry draft: oss-factory#189, cross-repo index #14 updated.kvnloo commentedon Oct 17, 2026
Hermes promotion handoff — target branch lock
Original user-authorized objective: complete a substantive e2e code-review wave and promote validated, nonduplicate fixes to existing upstream threads. This issue is the canonical queue; do not create competing origin PRs.
4932a5761ddbf75ea566eb705d787cbb6a02e0e1dc80df2424294f3514ef8a1c1f6e2ee04db1b119,tester-army/e2e:oskar/cache-relocation-ladder9614f8278718692c5ca35834560073f8aa5eb6014c095a6592ed49e956fee54e069f58c0bff9cc9d,tester-army/e2e:piotrek/eng-984-agent-anthropic-claude-5x-rejects-replayed-thinking-blockdc03fcddb9029396fe929b6927cd80222f11a183d310846848068b2e82a96d192b6a826f5a563447,kuishou68/e2e:fix/mcp-upload-filechooserEach fork draft has an immutable base ref (
review/805-base,review/1151-base,review/958-base) pointing at the named upstream SHA. Inspect the fork PR diff, not all fork changes against its default branch. CI is not verified green in this receipt.One-pass Hermes procedure
pnpm check/ relevant package unit suite; for fix(agent): keep the request prefix a Claude 5.x thinking block is bound to e2e-inc/e2e#1151 validate an actual Anthropic thinking run and non-Anthropic baseline if model access exists. For engine/agent performance changes, use the required baseline-vs-branch benchmark.#13: replay false retries after a detached-then-press attempt. Parent review by okwasniewski is already recorded; do not duplicate. Do not use stale source head212194d: the fixed newline regression is in4932a57.#14: smaller, bounded digest for model-message shape; no full process-RSS claim until measured; preserve TheTuso's original work and attribution. Review memory and CPU tradeoffs.#15: abort guard cannot allow deferredsetFiles; preserve kuishou68's original filechooser feature; Cubic already surfaced the issue.szdziedzic/e2e:szdziedzic/eas-rest-api,241ba3b4…): use the corrected, official Expo REST-contract questions in this queue. Don't post the earlier claim about a nonexistent endpoint or empty stop response. No live EAS billing without explicit approval.Current integration cannot post origin reviews: HTTP 403 on
tester-army/e2e#1067. GitHub origin PR e2e-inc#805 and e2e-inc#958 currently reportmergeable:false; verify actual conflicts on fresh heads before claiming ready. No completed upstream reviews/merges yet.kvnloo commentedon Oct 17, 2026
Next CR wave: targeted maintainer repair for e2e-inc#1209
New downstream draft PR: kvnloo/e2e#16, exact source head
9d8b0ff12f099fa353e1ab035dd3eb12659875d4, based directly on tester-army/e2e#1209 headefa739b1889028853ec44ead5929bdecd711f246. Origin code author knowbody; maintainer feedback and architectural direction from okwasniewski.Changes: restore all deleted source-lifetime JSDoc from commit 37f5681 (including shutdown semantics, retained device lease, pool finish and TestMu), factor retired-worker
revokeReuse()without restarting the kill deadline, factorcanKeepSessionwithout changing its predicate, and document the native recording reset's session/claim behavior next todevice.closeAppand the mobile changeset. 8 modified files, 42 changed lines.Grounding:
scheduler.test.tsalready covers retained reuse permission then finalshutdown, including force termination of a live retired worker. Re-read those tests and the actualTargetWorkerprotocol; this is source-level reasoning only, not an executed unit test.Current CI, exact SHA: fork PRs #13, #14, #15, #16 all show actions queued/pending, with no green verdict yet. Do not claim verified full checks or promote the patches as ready. Connector upstream write access remains unavailable (prior HTTP 403); user will promote from Hermes after CI/review gates.
Still unresolved in e2e-inc#1209, and not quietly changed by #16: the maintainer's two design decisions about reusing a failed worker's session and preserving the lease after recording-stop failure; bot's concern that a late recovered VideoSegment is lost on previous-attempt finalization. A new artifact/report pathway needs explicit supported contract and tests before patching. Reviewer already identified it, so avoid reposting as a new independent finding.
Hermes next: fetch these four fork draft diffs against their named immutable
review/*-basebranches (not forkmain), runpnpm check+ focused unit/integration and required comparative benchmark where applicable, evaluate newest CI/reviewer threads and origin exact heads, then ask/apply author-controlled minimal updates to the existing origin PRs. Respect reviewer requests and contributor attribution. No auto-approve, force push or origin merge.kvnloo commentedon Oct 17, 2026
CR wave continuation: tighten existing drafts
PR #15 guard cancelled file uploads advanced to exact head
ef5a825999ff84b595bde460bd57f23b701f2b51(previous headdc03fcdd...superseded). New tiny test executes the actual productionisFileInputelement.evaluate predicate with an associated<label>and fake HTMLInputElement instead of a unit stub that always returns true; asserts directsetInputFilesand no chooser/click. This addresses Cubic's pre-existing label test-coverage gap without creating another origin PR. Not yet run in repository Vitest or native browser.PR #14 signed-thinking message snapshots source remains
9614f8278718692c5ca35834560073f8aa5eb601. Isolated Node 22 smoke of the matching message-hashing code in this sandbox: 5/5 assertions pass; fixed 64-character digest; model tool-result change, provider metadata exclusion, in-place binary edit and cloned messages checked. 200 x 2MiB hashing runs: 277 ms wall (~1.385ms per snapshot) on this sandbox CPU. This is not a full repo Vitest result or an A/B runtime performance claim.All four fork drafts #13, #14, #15, #16 remain CHECK pending full
pnpm check, focused integration tests and applicable branch-vs-main benchmarks. Latest action statuses seen: queued/pending, no successful verdict. Do not promote without tests, reviewer check, current SHA and ownership authorization. Existing upstream reviewers and authors retain credit. Origin reviews posted: none (403 integration).