From 1ad5dc958abd2e6ad012263dd2a75392a9458af9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Val=20Alexander=20=F0=9F=91=91?= Date: Fri, 13 Nov 2026 00:16:12 -0600 Subject: [PATCH] docs: reconcile Code Review delivery and real-PR benchmark Record the protected inline-comment delivery, current accepted decisions and new authenticated benchmark while retaining the literal issue, human, provider and activation gates. --- docs/design/code-review.md | 2 +- .../2026-11-01-cr8-acceptance-matrix.md | 6 ++- .../2026-11-13-cr-inline-integration.md | 16 ++++++ .../2026-11-13-cr-real-pr-benchmark.md | 54 +++++++++++++++++++ docs/status.md | 4 +- 5 files changed, 77 insertions(+), 5 deletions(-) create mode 100644 docs/reviews/2026-11-13-cr-real-pr-benchmark.md diff --git a/docs/design/code-review.md b/docs/design/code-review.md index 73527f7d..5ed13f76 100644 --- a/docs/design/code-review.md +++ b/docs/design/code-review.md @@ -6,7 +6,7 @@ screenshots (R1–R4). The work is tracked in [#319](https://github.com/TypeSafe This page is CR-1's contract ([#320](https://github.com/TypeSafeAI/modex/issues/320)). Every later slice works against it. -**Where it stands (2026-11-01).** The core CR-1 through CR-7 slices have landed on `main`, and Code Review is present but hidden in the v0.4 releases, whose built apps require the local switch `modex.codeReview` to be `on`. From v0.4.2 a setting in Settings → General replaces that switch, off by default; Val chose on 2026-11-03 to keep Code Review hidden until its gates pass. Built apps required the switch because Gate A (#319) is not accepted ([v0.4.0 review](../reviews/2026-10-24-v0.4.0-release-review.md)). The candidate also carries the CR-7 sign-out/draft regression and CR-4 through CR-7 accessibility coverage (C76). Remaining engineering work is composer-clearance verification. C58 attachments and C59 effort remain omitted without an approved deviation, alongside the unresolved composer decisions below. Original-reference comparison, human keyboard/VoiceOver, the live CR-6 smoke and the CR-7 canary retain their separate acceptance gates. Each issue's ticked boxes and next action are in [status.md](../status.md#open-issue-checklists). +**Where it stands (2026-11-13).** The core CR-1 through CR-7 slices are on `main`; built apps keep Code Review hidden by default, with Settings → General available from v0.4.2. Val chose on 2026-11-03 to retain that default until the gates pass. The November 4 [decision record](#open-decisions) approves logins only, the built revision-bound local Viewed toggle, removal of C07, no access-mode indicator, and omission of C58 attachments/C59 effort while retaining model selection. Composer-clearance and CR-4 through CR-7 accessibility have fixture coverage; human keyboard/VoiceOver remains open. Val retired the original-image comparison in favour of a real-PR benchmark. A [new ten-PR comparison and live read-only controls check](../reviews/2026-11-13-cr-real-pr-benchmark.md) now covers the fixed build delivered by #520 as `967b703`; it preserves remaining thread/navigation limitations and does not tick issue boxes whose literal text still requires the original comparison. The real CR-6 provider smoke, Val-named disposable-repository CR-7 canary, human acceptance and default activation remain separate gates. Each issue’s ticked boxes and next action are in [status.md](../status.md#open-issue-checklists). CR-1 itself added a pure read contract, synthetic fixtures and this page, with no IPC channel, UI or GitHub request. Gate A then followed in order: diff --git a/docs/reviews/2026-11-01-cr8-acceptance-matrix.md b/docs/reviews/2026-11-01-cr8-acceptance-matrix.md index 04ca5ae0..f1597f14 100644 --- a/docs/reviews/2026-11-01-cr8-acceptance-matrix.md +++ b/docs/reviews/2026-11-01-cr8-acceptance-matrix.md @@ -17,6 +17,8 @@ Columns as in the baseline: **Fixture** (Node tests, or real Electron with synth **Authenticated** (read-only through a real `gh` login), **Live provider / write** (an authorized native-CLI turn, #325; an authorized write on a disposable target, #326), **Human** (a person, keyboard/VoiceOver or references). +**Dated follow-up, 2026-11-13.** The [real-PR benchmark](2026-11-13-cr-real-pr-benchmark.md) now records ten public PR comparisons on source `f9bef70`, whose complete tree matches protected main `967b703` delivered through #520. Real instructions/Escape focus restoration, pointer/keyboard resizing and nine narrow-window controls also passed. The [protected delivery receipt](2026-11-13-cr-inline-integration.md#actual-protected-delivery) keeps full source and merge-group counts separate. The dated baseline rows below retain their original evidence; this follow-up neither replaces historical failures nor accepts human, real-provider/write, original-reference wording or default activation gates. + ## Environments added since the baseline | ID | What ran | SHA | Result | @@ -45,7 +47,7 @@ Status values: *built*, *unavailable*, *deviation Dn*, *needs decision N*, *unch inventory left a control out. "unverified" in a Fixture cell means the control is in the code but no spec was found that asserts it; it is not a pass. Since 2026-11-04 no status is *needs decision N*. A Human cell that reads "open: reference comparison" names a comparison Val retired on 2026-11-04; it stays open until the -benchmark on real pull requests is recorded and the issue text that names the comparison is met or changed. +benchmark on real pull requests is recorded and the issue text that names the comparison is met or changed. The [2026-11-13 benchmark](2026-11-13-cr-real-pr-benchmark.md) supplies new authenticated evidence; literal issue-text reconciliation remains open. | ID | Control | Owner | Status | Fixture | Authenticated | Live provider / write | Human | | --- | --- | --- | --- | --- | --- | --- | --- | @@ -184,7 +186,7 @@ and C07 is removed. | AC | Where it stands on `0a8c513` | | --- | --- | -| AC1 R1–R4 in real Electron | All four reached through real controls over production reads with synthetic `gh` (`F-1101` journey). The original references are gone. On 2026-11-04 Val retired the comparison in favour of a benchmark on real pull requests against GitHub's pages; that benchmark has not run, and this does not tick a box whose text still requires the comparison. | +| AC1 R1–R4 in real Electron | All four reached through real controls over production reads with synthetic `gh` (`F-1101` journey). The original references are gone. On 2026-11-04 Val retired the comparison in favour of a benchmark on real pull requests against GitHub's pages; the [new ten-PR benchmark](2026-11-13-cr-real-pr-benchmark.md) and real instructions/narrow controls are now recorded on the tree delivered as `967b703`. This does not tick a box whose literal text still requires the original comparison. | | AC2 every visible control has behaviour or an explicit unavailable state | Partial: C07 is removed and C65 is built (decisions 6 and 5). C58 attachment and C59 reasoning effort are approved omissions (Val, 2026-11-04). The other controls are mapped above; seven built controls also lack an asserting spec (follow-ups below). | | AC3 isolation, injection, incomplete data, at-most-once | Met at fixture: read regressions, native denial on 2.1.316 (`N`), agent stop/unknown/restart, write at-most-once and reconciliation (`F-1101`). **Open:** the live smoke (#325) and live canary (#326). | | AC4 local Changes, navigation, conversation and approvals intact | `S-1101`: unit green; the local E2E had one non-Code-Review failure that passed alone, so it is not counted as green. The hosted required check passed on `c1f74c7`. | diff --git a/docs/reviews/2026-11-13-cr-inline-integration.md b/docs/reviews/2026-11-13-cr-inline-integration.md index 05336ca6..18f8253a 100644 --- a/docs/reviews/2026-11-13-cr-inline-integration.md +++ b/docs/reviews/2026-11-13-cr-inline-integration.md @@ -1,5 +1,7 @@ # Code Review line-comment integration +**Delivery follow-up:** #520 is on protected main as signed squash `967b703768d040d64456c7d118548b97a3bae35b`; see [actual protected delivery](#actual-protected-delivery). Pending sentences in the dated source-development record below describe their original checkpoints. + This owned integration starts from protected main `927b443414befe5f4300bb3f3d39af4666032609` and merges ready PR #513 source `c599dd7a9fbaeb4d6946666d8859e892e6847aca`, preserving its history and authors. The original owner branch and worktree are unchanged. Git independently reports only a `docs/status.md` conflict; this branch retains the full current-main file and applies narrow, evidence-backed status additions instead of reviving closed #378 or discarding delivered LM/CF/browser facts. The feature reads the first 100 REST diff-line comments into Activity, preserves line/file target provenance, labels file/line/replies and unavailable current-line positions, and reports partial, malformed or capped reads. Comment text remains untrusted Markdown. Resolved-thread state, placement beside changed lines and later comment pages are not implemented by this slice and remain part of their accepted parent requirements. No default activation or new live-write authority is implied. @@ -43,3 +45,17 @@ Frozen signed runtime/test source `57d2773e755259a2e618e40f62e179bd8f5d6d8c`, tr The seventeen desktop skips are four live-provider cases, four local-model native cases, three Open Knowledge cases and six installed-CLI Code Review cases. They are not acceptance passes. The six Code Review cases were then separately qualified on the same unchanged inputs: four ordinary passes and two intentional expected-failure controls in 22.1 seconds, using installed Claude Code 2.1.323, Electron 44.4.5, Chromium 152.0.7977.130 and embedded Node 24.21.0 on macOS 27.2.1 arm64. The actual CLI consumed hostile PR content, allowed selected evidence and refused outside-secret/symlink reads, shell, write, fetch, MCP and the placeholder; forbidden effect markers remained absent. Three native starts were recorded; root independently observed their three PIDs absent and their three provider ports refusing connections. Fifty native output files and all three owned fixture trees were independently byte-compared. Incoming GH_TOKEN was already absent, so this is not a supplied-token native scrub challenge. The provider and GitHub payloads were scripted, with no paid inference or live product write; cleanup observations cover sampled recorded children, not every possible descendant. One fresh whole-branch review and every ruling are retained, with no Critical/Important finding and no deferred Minor. The first coalesced review correction and the subsequent public-schema correction remain separately identified above. No second feature review was requested; the later CI-only correction received its own scoped review. The complete signed recovery restores into an empty bare repository, passes full fsck and reproduces all thirteen authored files byte-for-byte; original #513 source history and authorship remain ancestors. The final qualification-record commit changes documentation only; every runtime/test/dependency input stays byte-exact to the qualified source. Source-hosted CI, actual protected delivery and merged-tree equivalence remain pending. This is a reviewable implementation candidate, not whole-parent, human/live-provider/live-write, release or default-activation acceptance. + +## Actual protected delivery + +#520 merged on 2026-11-13 at 05:53:08 UTC as signed squash `967b703768d040d64456c7d118548b97a3bae35b`, parent `927b443414befe5f4300bb3f3d39af4666032609`, tree `6a241c26d089ac4d7daf3e4b30b23a28257010c5`. GitHub signature verification passed. The complete tree equals final qualified source `f9bef700390eb300d40424699193b2731d00e36c`; original #513 source remains in the integration/recovery history, while main records the squash. #513 was closed as superseded with [its reconciliation receipt](https://github.com/TypeSafeAI/modex/pull/513#issuecomment-6845436906), preserving its owner’s local branch and checkout. + +[Actual protected run 43370489247](https://github.com/TypeSafeAI/modex/actions/runs/43370489247/job/131704833478) checked out that exact squash and passed 2,587 units with 36 skips, 533 desktop cases with fourteen opt-in skips in 32.4 minutes, eight site cases in 23.0 seconds and four web cases in 4.2 seconds. No unit failure/cancellation or retry annotation was recorded. The 4,826-byte memory artifact has SHA-256 `ad0aa17380fe8db1c39c76fabba97c5cc7f6256e8b36de1f460023d0fad5a931`; the 9,133,856-byte E2E artifact has SHA-256 `946a9f46aedbb2079a74337262f109829667c337b9656ae4619fc6cbb819e234`. Both match API digests and pass ZIP integrity. Merge-group iPhone and coverage-upload jobs skipped as configured; separately triggered post-merge checks are not this protected run. + +Corrected source [run 43366524777](https://github.com/TypeSafeAI/modex/actions/runs/43366524777) passed all twelve checks. Its actual preview `8aaa11eb9468a2b23ff3529e86eb61382fb4d561`, parents `927b443` and `f9bef70`, has the same independently reproduced tree. Source macOS qualification passed 2,587 units with 36 skips, 533 desktop cases with fourteen opt-in skips in 25.5 minutes, eight site cases and four web cases. The complete iPhone results contain 65 native-unit passes, 43 ordinary UI passes and one intentional expected-failure control, with zero unexpected failures/skips. All four source artifacts passed API-digest and ZIP checks, including the 87,207,851-byte iPhone artifact with SHA-256 `6fcbf42641cace781e11a3f52aa8141ded3feb11e06a9ac84f7a62a2acadc544`. + +The first iPhone job in run `43360620861` hit its confirmed 30-minute whole-job maximum after 65 native-unit passes and 39 UI-case pass lines including one intentional expected failure; another case was running and its paired result bundle was incomplete. The separately reviewed correction raised only the whole iPhone job limit to 45 minutes. Individual waits, assertions, cases and retries stayed unchanged. Two workflow regression tests passed; the CI-only review found no severity findings. The later successful job used a different hosted runner, so the budget change alone is not proved to have caused the pass. The first cancellation and its artifact remain retained in the [published first-failure receipt](https://github.com/TypeSafeAI/modex/pull/520#issuecomment-6843893746). + +The [new real-PR benchmark](2026-11-13-cr-real-pr-benchmark.md) adds authenticated rendered evidence on the source-identical tree; it does not turn scripted-provider, simulator, sampled cleanup or automated keyboard results into real-provider/write, physical-device or human acceptance. No parent issue or default activation is accepted by this slice. + +The session-owned checkout was released and removed only through the guarded helper. Its new recovery bundle has SHA-256 `457c2fc561ce3588579bea811ae1cd693cd3945ca3fccb1d8bfa2d99d67000ad`; an independent empty-bare restore passed full fsck, verified all six own SSH signatures and reproduced all fourteen authored files byte-for-byte. All 38 owned probe files retain their original tree outside the removed checkout and independent byte-verified copies. Other owners’ worktrees and branches remain preserved. diff --git a/docs/reviews/2026-11-13-cr-real-pr-benchmark.md b/docs/reviews/2026-11-13-cr-real-pr-benchmark.md new file mode 100644 index 00000000..e5416e48 --- /dev/null +++ b/docs/reviews/2026-11-13-cr-real-pr-benchmark.md @@ -0,0 +1,54 @@ +# Code Review real-PR benchmark, 2026-11-13 + +Val’s November 4 [decision record](../design/code-review.md#open-decisions) replaced comparison with the deleted original screenshots with a benchmark on real PRs against GitHub’s pages. This record supplies a new ten-PR sample after the first usability fixes and #520’s review-comment integration. It preserves literal issue boxes that still require the original comparison. Human keyboard/VoiceOver, native real-provider review, a Val-named disposable-repository write canary and default activation remain separate gates. + +## Source and method + +The benchmark used signed source `f9bef700390eb300d40424699193b2731d00e36c`, tree `6a241c26d089ac4d7daf3e4b30b23a28257010c5`. That complete tree equals actual protected main `967b703768d040d64456c7d118548b97a3bae35b`, delivered by #520. All 852 qualified runtime/test/dependency inputs and 804 compiled files remained unchanged through the runs. [Source and actual protected qualification](2026-11-13-cr-inline-integration.md#actual-protected-delivery) are distinct from this read-only UI benchmark. + +Real Electron, preload, IPC and main-owned read services used BunsDev’s own `gh` login in new private Modex homes. No read handler or remote response was replaced. Test-only bootstrap instrumentation redirected each window’s `show` to `showInactive`; every sampled window remained unfocused. A temporary CLI wrapper admitted only GETs to this repository, its scoped search and account identity. Broad default inbox searches were deliberately refused; their visible refusal cards are excluded from product findings. GitHub’s public overview and Files pages loaded in a separate headless browser without an authenticated browser profile. No provider review, product GitHub write, installed-profile activation or release was requested. + +Each comparison selected the canonical PR link through the real input, checked its title/file count/head, navigated Changes and split views, loaded its file list and selected a file. Bracket navigation and F6 were exercised; the selected head was checked again independently afterward. GitHub page titles matched the same PR titles, and captures of both applications were retained privately. These are new-sample, inactive-window observations; the timing stops at visible overview metadata and is neither a paired historical speedup nor whole-app readiness. + +| PR | State during comparison | Files loaded | File pages | Visible metadata, ms | +| --- | --- | --- | --- | --- | +| #514 | merged | 58 | 1 | 823 | +| #506 | merged | 19 | 1 | 814 | +| #498 | merged | 151 | 2 | 1,445 | +| #497 | merged | 9 | 1 | 825 | +| #490 | merged | 31 | 1 | 1,349 | +| #475 | open | 23 | 1 | 1,324 | +| #519 | draft | 7 | 1 | 1,339 | +| #513 | open | 11 | 1 | 1,309 | +| #511 | draft | 19 | 1 | 803 | +| #520 | open | 14 | 1 | 799 | + +The ten observations range from 799 to 1,445 ms, with a 1,067 ms median. PR states above are historical observations: #520 subsequently merged, and #513 was closed as superseded. This benchmark neither merged #475 nor changed the draft/default-activation gates on #519/#515/#511. + +## Additional observed controls + +Activity on #514 displayed its real comment and reply. #506 displayed three real review comments, including two with an unavailable current line. Comments filtering retained them. A separate compiled-service control matched all five records’ path/subject/location/reply fields and body hashes against independent GitHub GET responses; it also verified the zero-review-comment case on #520. No raw comment bodies were published in fixtures or this record. + +An additional #506 control passed the instructions dialog’s local scope, initial textbox focus and Escape restoration to its trigger at 1440×900 and 900×600. No instructions were saved and no review was started. The PR divider moved from 512 to 544 by keyboard and to 524 by pointer drag. At 900×600 the split collapsed to one pane, both tabs remained usable, and pin/copy/Open in GitHub/Review/instructions/Merge/split/inbox-menu/search controls stayed inside the viewport. These are automated read-only interactions, not a person’s keyboard or screen-reader acceptance. + +## First failures and continuation + +The first packet completed #514 and #506, then stopped because the probe expected all 151 files on #498 without clicking the actual “Load next 100 files” control. Its continuation loaded both pages, then stopped at a different probe assumption: `]` was expected to select the grouped tree’s second row. The actual shortcut follows displayed diff/API order. On #498, the next diff after `.github/workflows/platform-cli.yml` is `README.md`, while the grouped tree’s second row is `apps/desktop/e2e/fixtures/local-draft-selection.ts`. + +The final continuation inspected the actual diff-section path order and verified its selected tree path, completing the remaining eight PRs without replaying the two completed comparisons. The additional controls packet passed separately. Both diagnosed failures, incomplete results and their closed-process receipts remain retained; no production repair, automatic framework retry or all-green first attempt is claimed for these probe corrections. + +## Ranked remaining friction + +1. Review-thread coverage: the reader now shows the five observed comments, reply and unavailable-line labels. Resolved-thread state, inline placement and comments beyond the first 100 remain outside its coverage. REST-only authority does not authorize a new GraphQL read path. +2. Reading width: reviewed 1440×900 split captures divide the available area among inbox, overview, code and tree, causing long code to wrap into short lines. Tested resizing and one-pane Changes are working alternatives. This is observed usability friction, not a demonstrated acceptance failure. +3. Navigation order: hierarchical tree order differs from bracket navigation’s displayed diff/API order on #498. This is an observed distinction, not evidence that selection failed after the corrected oracle checked the actual order. + +Closed-PR inferred-stack unavailability is documented behavior and is not classified as a transport defect. Historical check completeness does not imply current merge authority. Representative visual assessment covered real comments, split layout, GitHub Files, a large PR, a second overview, GitHub’s open-PR page, both instructions dimensions and narrow Changes; it did not compare every retained capture or substitute new images for the original references. + +## Retained evidence and acceptance limits + +The [published criterion receipt](https://github.com/TypeSafeAI/modex/issues/327#issuecomment-6845313006) has an exact independent body read-back. Private closeout evidence under `2026-11-11/modex-closeout` includes `cr-ten-real-pr-benchmark-combined-audit.json`, a byte/SHA-256 manifest for 55 captures across the two comparison packets and controls packet, `cr-inline-f9-real-read-audit.json`, complete per-packet requests/results, and the original/corrected probe drivers. The capture set includes the partial first attempt. Each sampled Electron PID was absent after app closure; this is sampled process evidence, not exhaustive descendant proof. + +Reproduction of these authenticated runs requires the retained private probe drivers and an owned checkout with the same built inputs. The corrected driver derives the app path from its working directory, uses `cr-real-read-inactive-bootstrap.cjs` in that checkout’s ignored `.probes` directory and writes a new private output directory. Its explicit PR list contains the eight remaining cases; the additional controls driver selects #506. Inspect those lists and the fixed private output location before a new sample. This record does not package a portable benchmark runner or claim that its private captures are public repository assets. Repository fixture regressions remain reproducible through the documented pinned-pnpm commands, with authenticated, provider/write and human evidence kept separate. + +No issue checkbox changes in this documentation slice. Original-reference wording reconciliation, real native-provider review, the live-write canary, human accessibility and default activation remain unaccepted. The narrower rendered/read evidence above does not close #319 or #322–#327. diff --git a/docs/status.md b/docs/status.md index 37bc4e66..f88d8931 100644 --- a/docs/status.md +++ b/docs/status.md @@ -147,7 +147,7 @@ before relying on a tick. ## In flight -- **Code Review line comments integration** (#513, #520). A new owned branch integrates ready source `c599dd7` from protected main `927b443`, retaining source authorship/history and every unrelated current-main status fact. Activity loads the first 100 REST review comments with distinct file/line targets, path/line/reply labels and explicit partial/capped states. Resolved-thread state, inline placement and further comment pages remain separate requirements. Signed runtime `57d2773` passes build/typecheck, 2,587 units with 36 skips, 530 desktop cases with seventeen opt-in skips, eight site cases and four web cases. Separate installed Claude Code 2.1.323 qualification passes four ordinary checks and two intentional fault controls with a scripted provider. The public-schema follow-up preserves valid location metadata on file targets. The first hosted iPhone job hit its 30-minute job maximum after 65 native-unit passes and 39 UI-case passes, including one expected-failure control; its full UI suite remains unverified. The original required macOS job passed 2,587 units, 533 desktop cases with fourteen opt-in skips, eight site cases and four web cases; the cancelled iPhone suite prevents a whole-source-green claim. The separately reviewed correction allows 45 minutes for that job while preserving all test deadlines and assertions. Corrected-head source-hosted qualification and protected delivery remain pending. [Integration receipt](reviews/2026-11-13-cr-inline-integration.md). +- **Code Review line comments integration** (#513, #520). Protected #520 delivered signed squash `967b703768d040d64456c7d118548b97a3bae35b`, tree `6a241c26d089ac4d7daf3e4b30b23a28257010c5`, from qualified source `f9bef70`; #513 is closed as superseded and its owner’s checkout/branch remain preserved. Activity loads the first 100 REST review comments with truthful file/line/reply/unavailable-location labels and partial/capped states. The actual protected checkout passed 2,587 units with 36 skips, 533 desktop cases with fourteen opt-in skips, eight site cases and four web cases, with no retry annotations; both artifacts passed digest/ZIP checks. Corrected source CI passed all twelve checks, including 65 native iPhone units and 43 ordinary UI cases plus one intentional expected-failure control. The first iPhone job’s 30-minute cancellation remains retained; the reviewed whole-job correction to 45 minutes changed no test assertions, individual waits or retries. Hosted simulator, scripted-provider installed CLI and real authenticated-read evidence remain distinct from physical phone, human, live-provider/write and activation acceptance. Resolved-thread state, inline placement and further comment pages remain separate requirements. [Integration and delivery receipt](reviews/2026-11-13-cr-inline-integration.md); [new real-PR benchmark](reviews/2026-11-13-cr-real-pr-benchmark.md). - **Local-model refresh preservation** (#381). Protected #503 delivered signed `558b54f8`, exact reviewed tree `3ba66a4`. Its queue passed 2,478 units with 36 skips, 522 desktop E2E cases with eight skips and seven site cases. Canonical cases preserve two thread drafts, staged image bytes, search/edit focus and selected text through delayed metadata, and keep an old revision-bound target after endpoint replacement. The full S2/AC3 UI/race audit and conservative S5 option/pinning guards are reconciled at five of 13 boxes; qualification/provenance/admission/human requirements remain open. [Delivery and limits](reviews/2026-11-12-local-model-refresh-preservation.md#actual-protected-delivery). - **CR-8 integrated native boundary fixture** (#327). Protected #517 delivered the boundary harness; fresh integrated Claude Code 2.1.323 qualification on the exact #518 tree passed four ordinary checks and two intentional expected-failure controls. The actual Electron/main/installed-CLI run consumes hostile PR text, permits selected evidence and refuses outside-secret/symlink reads, shell, write, fetch, MCP and the reserved placeholder. All 853 tracked inputs match protected main `927b443` byte-for-byte. Incoming GH_TOKEN was already absent, so this is not a supplied-token native scrub challenge. Full adversarial/race integration is now accepted at this named fixture/native scope; live-provider and human gates remain open. [Delivery audit and limits](reviews/2026-11-13-cr8-adversarial-delivery.md). @@ -172,7 +172,7 @@ before relying on a tick. - **Companion command recovery (COMP-CMD-2).** #497 is on main as `85143ee`; its tree exactly matches reviewed `d3d2eee`. It keeps timed-out command reads in the picker with explicit retry, retained search/drafts and keyboard Done actions. The required protected queue check passed 505 desktop cases (four skips) and seven site cases; iPhone was skipped for that merge-group event. The independent three-cycle context regression passes without #496's reset, so #496 is closed as superseded; its branch and checkout are preserved. Two fresh-install closeout integration suites each passed 44 executions (43 ordinary passes and one intentional expected failure, zero actual failures/skips), with source hashes unchanged. Human and physical-device acceptance on #276/#281 remains open. [Current integration qualification](reviews/2026-11-11-native-closeout-delivery.md); [historical producer review](reviews/2026-11-08-companion-recovery-pr.md). - **v0.4.2 release.** #492 (`4297235`) prepared it with Code Review hidden by default. Its tag exists, and the GitHub release remains a draft; it is not published. [Review](reviews/2026-11-03-v0.4.2-release-review.md). -- **Code Review acceptance** (#319). Every slice's code is on `main` (design and history: [design page](design/code-review.md), [CR-8 matrix](reviews/2026-11-01-cr8-acceptance-matrix.md)). Built apps hide it by default (#492); Settings → General can show it. #490 added that setting, C76 accessibility coverage, performance budgets as gates and the CR-7 sign-out and CF-3 question-card regressions; it also carries #481, #484, #485 and #491. No live provider run, live GitHub write, VoiceOver pass or reference comparison has happened. Those are listed in [Val's one session](reviews/2026-11-02-v0.5.0-val-session.md), which predates the 2026-11-04 answers and still asks for them. On 2026-11-04 Val answered decisions 1, 5, 6 and 7 ([record](design/code-review.md#open-decisions)): logins only, a local revision-bound Viewed toggle, the new-tab control removed, no access indicator; C58 and C59 effort are approved omissions; the reference comparison is retired for a benchmark on real pull requests; real reads and one CR-6 smoke may run on this repository. The Viewed toggle is built and C07 is removed. A benchmark on ten real pull requests of this repository ranked the friction; its first fixes are in: a pull request opens in about two seconds instead of four, comment HTML reads as text, inbox rows show the pull request number, and F6 and `[` / `]` move between areas and files. Review comments made on diff lines are still not shown. The benchmark has not been re-run on the fixed build. +- **Code Review acceptance** (#319). Every slice's code is on `main` (design and history: [design page](design/code-review.md), [CR-8 matrix](reviews/2026-11-01-cr8-acceptance-matrix.md)). Built apps hide it by default (#492); Settings → General can show it. #490 added that setting, C76 accessibility coverage, performance budgets as gates and the CR-7 sign-out and CF-3 question-card regressions; it also carries #481, #484, #485 and #491. No live provider run, live GitHub write, VoiceOver pass or reference comparison has happened. Those are listed in [Val's one session](reviews/2026-11-02-v0.5.0-val-session.md), which predates the 2026-11-04 answers and still asks for them. On 2026-11-04 Val answered decisions 1, 5, 6 and 7 ([record](design/code-review.md#open-decisions)): logins only, a local revision-bound Viewed toggle, the new-tab control removed, no access indicator; C58 and C59 effort are approved omissions; the reference comparison is retired for a benchmark on real pull requests; real reads and one CR-6 smoke may run on this repository. The Viewed toggle is built and C07 is removed. A benchmark on ten real pull requests of this repository ranked the friction; its first fixes are in: a pull request opens in about two seconds instead of four, comment HTML reads as text, inbox rows show the pull request number, and F6 and `[` / `]` move between areas and files. #520 now shows the first 100 review comments in Activity. A [new comparison on ten public PRs against GitHub’s pages](reviews/2026-11-13-cr-real-pr-benchmark.md) exercised the fixed build, with real reply/unavailable-line labels, two-page file loading, instructions, resizing and narrow controls. Its new sample and inactive-window timings are not a controlled historical speedup; literal original-reference boxes and human/live gates remain unaccepted. - **Conversation fidelity** ([tracker #273](https://github.com/TypeSafeAI/modex/issues/273)). CF-3, CF-7 and CF-8 stay open; their code is on `main`. The #489 question-focus repair and Approve/Deny/Stop/withdrawal draft coverage are integrated by #495. The human gates are under [Open acceptance gates](#open-acceptance-gates). - **Protected delivery diagnostics** (#498). Earlier #497 queue shutdown failures and #498's hosted iOS 26.4.1 post-relaunch TLS stall remain unexplained. The closeout candidate adds bounded test-only native stack sampling and DEBUG-only opt-in TLS callback phases, preserving close outcomes, certificate APIs, exact pins, redirects and deadlines. Fresh build/typecheck, 2,412 units (35 skips), 508 desktop E2E cases (11 explicit skips), seven site cases, 65 native units and the fresh 44-execution traced paired suite passed. Actual iOS logs show five callbacks across two app processes; macOS security controls preserve matching/mismatched-pin and redirect behavior, with default/Release logging silent. The additional canonical export regression passed within eight token-record cases after compilation. The published `be5a4c8` head subsequently passed required macOS, iPhone and Linux/Windows checks; those checks do not validate the later private-workspace trust repair. These remain candidate results, not a TLS/shutdown repair or protected delivery. [Current evidence and limits](reviews/2026-11-11-native-closeout-delivery.md); [shutdown observer](reviews/2026-11-11-electron-shutdown-diagnostics.md). - **Private local workspace continuation** (#379/#380, #498). A network-denied real Codex 0.167.0/Ollama 0.40.16 read reproduced refusal on the next turn after the child persisted its own workspace trust mark. The strict repair admits only that one main-bound private workspace entry, keeps other settings and managed/project overrides refused, and checks captured realpaths before setup and after configuration reads. These spot checks do not establish race-free filesystem identity. All 45 focused tests, root build/typechecks and full units (2,420 passes, 35 skips) passed on the frozen candidate; full E2E passed with 508 desktop passes, 11 explicit skips and seven site passes, with no failures or observed retries. The unchanged 500 ms cold-open gate measured 73.2 ms; all four affected real-model Electron lifecycle cases separately passed with zero skips or retries. Native continuation and completed compaction were observed, but independent session audit found two repeated reads after compaction, so action-safe compaction and production admission remain unqualified. [New qualification and limits](reviews/2026-11-11-native-private-workspace-trust.md).