[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end

Outcome and scope

Build a functional Code Review workspace matching the four user-supplied desktop screenshots: PR inbox → overview/activity → split remote diff/file tree → explicitly requested AI review → human-confirmed GitHub action. Preserve Modex's existing desktop shell and local Changes workflow. This is an implementation backlog, not a claim that the feature has shipped.

Proposed implementation home: TypeSafeAI/modex, apps/desktop. The screenshots display both Modex and River Oaks PRs; the repository being reviewed is not necessarily the application owning the interface. Do not implement this in BunsDev/river-oaks, TypeSafe Computer, or the game iframe without an explicit scope change. #320 must record Val's confirmation of this placement before feature wiring. Read-only inspection and synthetic contract preparation can proceed before that decision.

Research limitation: the completed deep-research report for these four screenshots was not retrievable during this audit. Do not claim it was reviewed. This backlog is a fresh assessment of the supplied images, current source, existing issues and primary documentation. Reconcile a subsequently supplied report with current code; do not accept its proposals as implemented facts.

Verified baseline and reuse

Read-only audit baseline: main at cb33db5df00e235f68d36b5fac556dc30e0851f2. Re-fetch current main, issue discussions, open PRs and relevant scoped instructions before coding; this SHA is evidence, not a command to revert later work.

AGENTS.md: owned worktrees, pnpm, signed commits, protected-main checks, native Claude/Codex CLIs only for coding turns.
ChangesPanel.tsx: existing local-working-tree review toolbar, file filter/tree, diff renderer, context expansion and destructive Discard affordances. Reuse presentation selectively; remote PR content must NEVER enter local discard/path handlers.
git.ts: local status/diff/worktree helpers, not proof of a complete GitHub PR service.
desktop package.json: Electron/React/TypeScript/Vite; Node tests, Playwright, existing ui:capture and ui:compare scripts. No framework or editor migration is required by these screenshots.
status.md and [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273–[P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281: native conversation lifecycle, safe Markdown, approvals/questions, evidence and recovery already have owners and substantial implementation. Consume verified contracts; do not create a second transcript, approval system, Notes store or agent runtime. Open issue status alone does not mean its entire feature is missing.

Reference states

These are original image pixel sizes, NOT established CSS viewport/DPR measurements. Originals remain user-supplied references, not repository assets. Do not commit unredacted screenshots, unrelated PR data, avatars or credentials without approval.

ID	Original file	Pixels	Required state
R1	Screenshot 2026-10-18 at 15.38.42.png	1766×1057	Overview at top; title/body; stack/checks sidebar; floating PR composer
R2	Screenshot 2026-10-18 at 15.38.52.png	2048×1217	Scrolled activity; commits, comment composer, deployment preview; same selected PR
R3	Screenshot 2026-10-18 at 15.38.55.png	2047×1214	Review-instructions popover open above the activity view
R4	Screenshot 2026-10-18 at 15.40.30.png	2048×1215	Side-by-side PR description and Changes workspace; unified diffs and right file tree

Preserve composition, density, independent scrolling, hierarchy, selected/hover/focus states and existing Modex theme tokens. Use deterministic synthetic PR content with comparable line lengths. Screenshot-only controls with no verified backing capability must be explicitly unavailable, not fake-working. Native window chrome, voice input, proprietary cloud review APIs, stack mutation/rebasing, admin bypass and game integration are outside this initial scope.

Created issue graph and dependency rules

Slice	Issue / deliverable	Start prerequisite	Completion / integration prerequisite
CR-1	#320 — reference/source contract and deterministic fixtures	None; confirm proposed product placement before feature wiring	Reviewed contract, reference access and placement decision
CR-2	#321 — authenticated main-owned PR reads	#320 accepted contract	Read/identity/race tests plus separately recorded live-read gate
CR-3	#322 — inbox, navigation and resizable workspace	#320 accepted contract; can run alongside #321	#321 live adapter; navigation/layout evidence
CR-4	#323 — overview, activity, checks, stack and previews	#321 and #322 accepted artifacts	Safe rendering, provenance and R1/R2 evidence
CR-5	#324 — remote diffs, folding and file navigation	#321 and #322 accepted artifacts	Revision/anchor correctness, local Changes regression and R4 evidence
CR-6	#325 — native review, PR questions and instructions	#321, #323 and #324; relevant accepted #273 contracts	Runtime authority, lifecycle, coverage and R3/native evidence
CR-7	#326 — explicit GitHub writes and guarded merge	#321, #323 and #324	Confirmation/receipt/race tests; separately authorized live-write gate
CR-8	#327 — integrated fidelity, accessibility and safety acceptance	#320 for fixture/test design	#321–#326 accepted implementations and their evidence

Parallel lanes: after #320, #321 and fixture-first #322 can proceed in separate owned worktrees. After their contracts are accepted, #323 and #324 can proceed in parallel. #325 and #326 are independent consumers: AI review does not grant or gate human GitHub write permission. #327 starts test design early but cannot claim integrated completion until its producer artifacts are ready.

Treat dependencies as accepted artifacts/behavior, not merely closed issue numbers. An accepted producer contract can unblock a consumer even while unrelated producer follow-ups remain open. Preparation can use agreed fixtures; live integration waits for the named producer. Parent tracking is not a blocking dependency. Do not add reverse dependencies from producers to #327, duplicate owners, cycles or blanket waits for all of #273 to close.

These relationships are explicit Markdown issue links and artifact gates. They are not claims of native GitHub blocking/sub-issue metadata or an automatically scheduled agent workflow. No agent assignment, merge or deployment was performed by creating this backlog.

Delivery gates

Gate A — read-only workspace: #320–#324, with #327's relevant tests. Users can navigate real PR data and inspect revision-bound diffs; AI/write controls stay capability-gated.

Gate B — native review: #325. Explicit user start, bounded context, runtime-enforced read-only authority, correct native lifecycle and visible coverage. Unsupported safety controls keep the feature unavailable.

Gate C — external writes: #326. Fresh identity/permission checks, explicit intent, expected-head merge protection and conservative unknown-outcome handling. No dependence on an AI verdict.

Gate D — full acceptance: #327. Reproducible visual/behavioral evidence plus named authenticated/native/write/human gates. These are delivery phases, not newly created GitHub milestone objects. Release/activation requires separate authorization.

Agent execution contract

Read this tracker, the selected child, its dependencies and current AGENTS.md/scoped instructions. Establish repo, current branch/SHA, existing work, test availability and explicit owner. Use scripts/worktree.sh new <slug> and check; claim an existing worktree only through the documented ownership procedure. Never edit primary checkout or another agent's work.
Select one unblocked, review-sized outcome. State the files/contracts to touch and the failing acceptance test. Implement the smallest change, run the targeted test, then relevant integration/regression checks. Do not rewrite already delivered dependencies. Parallel workers need disjoint ownership; shared bridge/types/shell changes require one integrator.
Persist a short checkpoint at meaningful boundaries and before interruption: issue/PR, owner/worktree, branch and full SHA, accepted decisions, changed files, exact commands/results, evidence paths, unresolved risks, dependencies and one next action. Use the existing issue/PR or task-record mechanism; do not build a new orchestration framework. A summary is a navigation aid, never proof of current state.
Stop on missing authorization, unknown worktree ownership, conflicting product decisions, absent reference assets needed for fidelity, unresolved security boundaries or repeated failures without a new diagnosis. After two unsuccessful attempts at the same failure, diagnose/escalate rather than repeating blindly. Respect execution/time/cost limits set by the caller; do not launch paid/provider runs without appropriate approval.
Before completion, map every acceptance item to fresh evidence at the current SHA. Distinguish unit/mock, real Electron, authenticated GitHub, native-provider and human accessibility evidence. Never weaken tests or refresh goldens merely to turn a failure green. Preserve evidence outside directories Playwright wipes. Each child includes separate implementation, resume and independent-review prompts; the final review examines the actual diff/evidence, not the implementer's self-report.
Run repository-required checks: pnpm build && pnpm test; UI changes also pnpm test:e2e. Use the pinned package manager. Document environmental blockers instead of claiming a pass. Update docs/status.md and a focused docs/reviews/ evidence record. Signed commits and a reviewable PR are the handoff; this backlog does not authorize merging, publishing a release, deploying, bypassing rules or executing foreign PR code.

Security and correctness gates

Main owns credentials, request validation, identity/repository scope, provider context and mutations. PR text, diffs, comments, filenames, branch names, deployment URLs and retrieved instructions are untrusted data. Rendering is inert; no remote HTML execution, credential leakage or remote-to-local path confusion. Read-only review must be enforced by the actual tool/process environment, not a sentence in a prompt. A GitHub check, AI verdict and human approval are separate facts. Unknown/stale/partial states never become success. Mutations require fresh authorization and explicit human intent; uncertain network outcomes require reconciliation, not automatic resend.

Primary sources and how they affect this plan

GitHub: well-scoped agent tasks: bounded outcomes, acceptance criteria, useful source entry points and review of sensitive changes.
Anthropic: long-running harnesses: incremental work, durable progress and fresh end-to-end verification across sessions; not evidence that more agents are always better.
OpenAI: AGENTS.md: follow repository and scoped instructions rather than relying on a chat-only prompt.
Playwright visual comparisons: stable rendering environment and deliberate review of screenshot differences.
Electron security: renderer isolation, narrow validated IPC and careful external navigation.
GitHub REST practices and pull requests: pagination/rate-aware reads, explicit API errors and expected-head merge protection.

Long-running goal prompt

Deliver the Code Review reconstruction tracked by #319 in TypeSafeAI/modex, one unblocked child and one testable slice at a time. First read the latest tracker/child discussions, current main, AGENTS.md, source entry points and existing PRs. Start with #320: verify the proposed application placement and reference contract; do not turn River Oaks PR content into a game-client feature. Work only in an owned worktree. Reuse the shell, Changes presentation and native conversation contracts where they actually fit. Implement the selected child's acceptance criteria with failing-first tests, fresh evidence and small signed commits. Do not replace the stack, invent provider capabilities, run foreign PR code, expose credentials, or merge/deploy. Keep all untrusted review content outside executable authority. At every durable boundary record SHA, changed files, commands/results, remaining gates and one next action. Continue only within the caller's time/cost/authority limits; stop on a real blocker with a precise handoff. Finish with a reviewable PR and criterion-by-criterion evidence, not a claim based on a screenshot or an agent's own completion summary.

Follow-up / resume prompt

Resume #319 by reading its live issue graph, linked PRs and latest checkpoints. Re-establish repository, current main, worktree owner, branch and full SHA. Inspect the actual diff and rerun the relevant smoke test before trusting previous progress. Check whether prerequisite artifacts landed or changed; accept delivered contracts without needlessly waiting for unrelated issue closure. Pick the smallest remaining unblocked acceptance item and continue its child prompt. Preserve unrelated work, keep privacy/security gates intact, and do not replay external writes or provider runs whose outcomes are uncertain. If all automated criteria pass, identify remaining native, authenticated and human validation gates rather than declaring them covered. Update the checkpoint and status with observed evidence only. Do not merge or release.

Independent review prompt

Review the current #319 implementation against the live child issues and accepted R1–R4 contract. Read the actual source diff, dependency artifacts and fresh test evidence independently of the implementer's summary. Verify that existing Modex functionality was reused without remote-to-local authority leakage, and exercise account/head races, incomplete evidence, injected content and uncertain-action recovery. Inspect the UI and distinguish fixture, Electron, authenticated, native-provider, live-write and human accessibility proof. Report each acceptance criterion as verified, failed, deferred by an explicit decision or still unverified. Request focused fixes without expanding scope or authorizing merge/release.

Completion checklist

Gate A checkpoint 1 (2026-10-19): CR-1 ready for review

The Gate A goal started from this tracker's prompt, scoped to #320–#324 plus #327's tests for those slices. The first slice, CR-1 (#320), is up for review as #334.

Where: worktree .worktrees/cr1-contract (created with scripts/worktree.sh new), branch cr1-contract, head 6d77cbd, base 9c0b063.
main: at de856e0, one commit ahead of the branch's base. Main gets merged in when the PR lands.
Decisions: the three on [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 (comment). Val confirmed them in the planning session, and that session posted them.

What CR-1 delivers

docs/design/code-review.md holds:

The contract (apps/desktop/src/shared/code-review.ts):
- A locator, an account-scoped key, and a revision.
- Read states: ready, partial, stale or unavailable.
- Check rules: pending is never passing, and a newer run of a check replaces the older one.
- Diff coverage.
- Gate A capabilities: review and every write stay unavailable.
- scopeOf and acceptsReply, so a slow reply for the wrong account, repository or head is dropped.
- validateRequest, main's check on everything the renderer sends.
Synthetic fixtures: a made-up project; none of the screenshots' content.
A source map checked at file:line.
The reference manifest: checksums, scale, crop offsets, measured geometry, and typography and spacing.
An inventory of 82 controls, each with an owner and a test.
Seven open decisions for Val.

Review

First review (80ad189) requested changes:
- The fixtures were the screenshots' PRs with the nouns swapped.
- Three contract rules didn't hold as written.
These were fixed in cf57a4b.
Re-review: confirmed the fixes and left three small items, fixed in 6d77cbd.

Evidence on `6d77cbd`

Contract tests: 11/11.
Mutation checks: I broke the compiled contract 16 deliberate ways, and each break failed a test.
Full suite: pnpm build && pnpm test exited 0:
- worktree scripts 109/109;
- core 16/16;
- desktop unit 1,672 passed, 16 skipped, 0 failed;
- then 21, 10 and 16 in the remaining suites.
E2E: not run; there is no UI change.
CI on Define the Code Review read contract and fixtures (CR-1) #334: running.

Still open

Landing Define the Code Review read contract and fixtures (CR-1) #334 needs Val's go-ahead.
Seven decisions for Val, listed under Open decisions:
- avatars;
- the state menu;
- title and description edits;
- requesting reviewers;
- the Viewed toggle;
- the new-tab button;
- the access-mode indicator.
Each control stays unavailable until decided.
The authenticated read smoke test ([CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 AC5) comes later.

For #321

From GitHub's current REST docs (to confirm live):

Search: 30 requests a minute, at most 1,000 results per query.
PR files: the list stops at 3,000 files.
PR commits: the list stops at 250.
File contents: served raw up to 100 MB.
Search results: the schema lists a repository object with an id, but that hasn't been confirmed in a real response.

Next action

On Val's go-ahead:

Land Define the Code Review read contract and fixtures (CR-1) #334 through the merge guard.
Link the merged SHA on [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321–[CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327.
Start CR-2 ([CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321) and CR-3 ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322) in separate worktrees.

Nothing in those slices binds to the contract before it's on main.

Coven-Harness: Claude Code / claude-opus-5-5

Define the Code Review read contract and fixtures (CR-1) (#334)

Gate A checkpoint 2 (2026-10-20): CR-1 landed; CR-2 and CR-3 started

CR-1 landed

On main: Define the Code Review read contract and fixtures (CR-1) #334 landed on Val's go-ahead as 43d5b9c. GitHub shows the commit as Verified. It went through the merge guard (--squash) from the primary checkout, with all 11 checks passing and mergeState=CLEAN.
Catching up with main: other PRs landed while it was in review, so I merged main in four times (Prepare desktop 0.3.0 and Companion 0.3.0 (8) #337, Preserve typed conversation fidelity and streaming reader state (CF-4, lands #295) #333, Record native answer delivery only after its transport handoff (CF-3, lands #316) #335, Run live smoke acceptance through the app: Claude passes, Codex is out of credits (CF-8) #332, Read the thread again when the Mac refuses an answer from the phone (CF-6) #318, then Leave a request open when saving its answer fails (CF-6, finding 10) #336, End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341 and Record v0.3.0 publication and advance the website fallback #339). Each merge was signed, and pnpm build && pnpm test exited 0 every time. On the final head, 5106c0d, the desktop unit suite passed 1,716 tests, skipped 16 and failed none.
Linked: the merged SHA is on [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321–[CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327, with what each child binds to.
Cleaned up: the cr1-contract worktree is removed through scripts/worktree.sh remove, which kept a recovery archive.
Still open: [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 is still open for Val to close.

Decisions since checkpoint 1

AC5 smoke repository: Val authorized TypeSafeAI/modex for CR-2's authenticated read-only smoke test ([CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 AC5). It reads through Val's gh login, GET only.

CR-2 and CR-3

The split: the design page's decision 4 fixes the interface. The renderer calls bridge.invoke("code-review:read", CodeReviewRequest), and main answers a CodeReviewReply, or null for cancel. That one BridgeCommands line is commit 372edbf, and both branches carry it.
CR-2 ([CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321): worktree cr2-read-service, at 372edbf on 43d5b9c. This session owns the main read service, the localOnly handler, and contract extensions.
CR-3 ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322): worktree cr3-inbox, on a separate branch, renderer only. E2E serves the fixtures through main's handler. CR-3 lands after CR-2 and wires the live read path then.

Open for Val

Visibility before Gate A completes: should the Code Review destination appear in the app as soon as CR-3 lands, or stay hidden in packaged builds until [CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327's acceptance, while staying on in dev and E2E? CR-3 builds it behind one switch that defaults to hidden in packaged builds. Recommended: hidden until Gate A completes.
The seven open decisions in the design page are still open.

Next action

CR-2: tests first for the read service (account switch, delayed replies, pagination and caps, revoked credentials, head changes, malformed paths), then the code-review:read handler.

Coven-Harness: Claude Code / claude-opus-5-5

Gate A checkpoint 3 (2026-10-20): CR-2 and CR-3 in review

CR-2 ([CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321): Add main's Code Review read service through gh (CR-2) #349, head 197276e, base 43d5b9c. Main's read service on code-review:read (localOnly), GET-only through Val's gh login.
- pnpm build && pnpm test exited 0: desktop unit 1,739 passed, 16 skipped, 0 failed.
- Code Review tests 34/34. 25 deliberate breaks of the compiled read path each fail a test.
- The AC5 smoke test passed on TypeSafeAI/modex Define the Code Review read contract and fixtures (CR-1) #334 and Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330. Account, host, PRs and head SHAs are in the PR; no credentials.
CR-3 ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322): Build the Code Review inbox and workspace host (CR-3) #346, head 8883428, base 43d5b9c. The inbox and workspace host, renderer only, with every read through code-review:read.
- Code Review spec 9/9. Full suite and pnpm test:e2e exited 0.
- It carries CR-2's three small commits as cherry-picks, and lands after Add main's Code Review read service through gh (CR-2) #349.
- The live read path and the authenticated navigation smoke test ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322 AC6) wait for Add main's Code Review read service through gh (CR-2) #349.
Contract additions (in Add main's Code Review read service through gh (CR-2) #349 and cherry-picked into Build the Code Review inbox and workspace host (CR-3) #346):
- PrSummary.stackHint: boolean | null, since search results name no base branch;
- Unavailable gains head-changed;
- parseRevision is exported.

Open for Val

Landing Add main's Code Review read service through gh (CR-2) #349, then Build the Code Review inbox and workspace host (CR-3) #346.
Visibility before Gate A completes: Build the Code Review inbox and workspace host (CR-3) #346 hides the destination in packaged builds unless localStorage["modex.codeReview"] is "on".
The seven design-page decisions are still open.

Next action

An independent review of #349, then landing it on Val's go-ahead. After that, merge main into cr3-inbox, wire the live read path, and run #322 AC6.

Coven-Harness: Claude Code / claude-opus-5-5

Reconcile historical CF8 evidence and keyboard review kit with main (#…

Reconcile CF8 history with current main verification records (#364)

Refreshed goal prompts — 2026-10-22 (`main` at `0e674e7`)

Where Gate A stands: CR-1 contract (#334, 43d5b9c), CR-2 read service (#366, plus cache denial #354/#361) and the CR-3 inbox/workspace host (#367, af49881) are on main; #320 and #321 are closed. apps/desktop/src/renderer/code-review/PrPane.tsx still shows placeholders for overview, diff and files, which are CR-4 (#323) and CR-5 (#324). Neither has started. #322 stays open for the R1–R4 comparison and human keyboard/VoiceOver review. Waiting on Val: the seven open decisions and whether the destination stays behind localStorage["modex.codeReview"] until Gate A completes. Five worktrees whose PRs already landed are still checked out (cr1-verify, cr2-cache-denial, cr2-ipc-acceptance, cr2-review, cr3-inbox-repair).

Long-running goal prompt

Coordinate the Code Review tracker #319 in TypeSafeAI/modex to Gate A completion, then Gates B/C. Read AGENTS.md, docs/status.md, docs/design/code-review.md and live #319–#327; verify main's SHA and open PRs. CR-1/2/3 are on main (43d5b9c, #366, af49881). Gate A's remaining work is CR-4 (#323 overview/activity/checks/stack) and CR-5 (#324 revision-bound remote diff and file tree). Run them in parallel, each in its own worktree from scripts/worktree.sh. Each binds to apps/desktop/src/shared/code-review.ts and the fixtures, and replaces a placeholder in renderer/code-review/PrPane.tsx. Have CR-8 (#327) build its acceptance matrix alongside them, not after. After both land, close #322 by running the R1–R4 comparison with populated content. Then start CR-6 (#325) and CR-7 (#326). Do not start CR-7 GitHub writes before Val decides how writes are authorized. Keep the seven design decisions and the visibility flag with Val; a control stays unavailable until its decision is made. One owner per shared file (contract, bridge, PrPane). Code PRs land only on Val's go-ahead through the merge guard, with signed commits. After each landing, post a checkpoint here: SHAs, PRs, evidence links, open decisions and one next action. Never merge, tag, release or flip the flag on your own.

Follow-up / resume prompt

Resume #319 coordination. Read the latest checkpoint here and on #322–#327, then git fetch and compare main with the last SHA recorded. List the open CR PRs and worktrees. For each, confirm the owner session, whether it is clean and whether the PR merged; ask owners to release/remove merged ones and never remove another session's worktree. Re-check which design decisions Val has answered and update the inventory owners in docs/design/code-review.md. Choose the one slice whose producer artifacts are accepted and unblock it. Record fixture, authenticated, live-provider and live-write evidence separately. Post a checkpoint with exact SHAs and the next action.

Refreshed goal prompts — 2026-10-23 (`main` at `203a350`)

On main already:

CR-1 contract and fixtures (Define the Code Review read contract and fixtures (CR-1) #334, 43d5b9c); [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 is closed;
CR-2 main-owned read service (Add main-owned GitHub PR reads with production IPC coverage #366, plus cache denial Invalidate retained PR data after visibility loss #354/Exercise CR2 reads through production Electron IPC #361); [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 is closed;
CR-3 inbox and workspace host (Integrate Code Review inbox and workspace with main-owned PR reads #367, af49881).

No Code Review code has landed since 0e674e7. The ten commits since then are CF work.

Still open:

CR-4 ([CR-4][P1] Reconstruct PR overview, activity, checks, stack and preview cards #323) is in review as feat(code-review): add PR overview, activity and evidence cards #393 (cr4-overview, head 52375d4, base 68860a0). All checks pass and it merges cleanly. It has no review yet and needs Val's go-ahead.
CR-5 ([CR-5][P1] Add revision-bound remote diffs, context folding and file navigation #324) has two signed commits in the local worktree cr5-remote-diffs (head 0fd82e1, base 68860a0). They are not pushed and have no PR. Its evidence doc says the full E2E run on the final head is still pending.
CR-8 ([CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327) is in review as Add the CR-8 acceptance matrix, Electron journeys and cross-slice regressions (#327) #372 (cr8-acceptance, head 92ce082). Checks pass, it merges cleanly, and it includes the authenticated re-run on main (A-main).
feat(code-review): add PR overview, activity and evidence cards #393 and the CR-5 branch both change PrPane.tsx, CodeReviewView.tsx and code-review.css, so whichever lands second has to merge main.
The R1–R4 references are missing. They are no longer at the manifest's ~/Desktop paths, as Add the CR-8 acceptance matrix, Electron journeys and cross-slice regressions (#327) #372, feat(code-review): add PR overview, activity and evidence cards #393 and the CR-5 evidence all report. [CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327 asked Val where they are, and there is no answer yet.
CR-6 ([CR-6][P1] Add native AI review, PR questions and scoped review instructions #325) and CR-7 ([CR-7][P1] Add explicit review actions and head-guarded GitHub merge #326) have not started.
Waiting on Val:
- the seven open decisions;
- the localStorage["modex.codeReview"] visibility flag;
- how CR-7 writes are authorized (docs/status.md, Next Convention: one worktree per session under .worktrees/ #3).

cr1-verify, cr2-cache-denial, cr2-ipc-acceptance, cr2-review and cr3-inbox-repair are still checked out, although their PRs have landed. cr4-overview, cr5-remote-diffs and cr8-acceptance are active.

Long-running goal prompt

Coordinate the Code Review tracker #319 in TypeSafeAI/modex: finish Gate A, then run Gates B and C. Do not implement the child issues yourself.

Read AGENTS.md, docs/status.md, docs/design/code-review.md and live #319–#327. Verify main's SHA and list the open PRs.

What is on main: CR-1/2/3 (43d5b9c, #366, af49881).

Gate A's remaining work is three pieces:
- CR-4 (#323): in PR #393.
- CR-5 (#324): unpushed commits in worktree cr5-remote-diffs.
- CR-8 (#327): in PR #372.

Get each one reviewed independently and landed on Val's go-ahead. Land #393 and CR-5 one after the other; the second merges main and keeps both slices' behaviour in PrPane. After each landing, have CR-8 re-run its journeys and update the matrix.

Then close #322: run the R1–R4 comparison with populated content and the human keyboard/VoiceOver pass.

Then start CR-6 (#325) and CR-7 (#326), each in its own worktree from scripts/worktree.sh. CR-6's Phase 1 isolation spike can start now. Do not start CR-7's GitHub writes before Val decides how writes are authorized.

Ask Val where the R1–R4 references are now. Keep the seven design decisions and the visibility flag with Val; a control stays unavailable until its decision is made. Each shared file (contract, bridge, PrPane) has one owner.

Code PRs land through the merge queue with signed commits, and only on Val's go-ahead. After each landing, post a checkpoint here: SHAs, PRs, evidence links, open decisions and one next action. Never merge, tag, release or flip the flag on your own.

Follow-up / resume prompt

Resume #319 coordination.
1. Read the latest checkpoint here and on #322–#327. Run git fetch and compare main with the last SHA recorded (203a350).
2. Check #393, #372 and the CR-5 branch: whether each merged, its CI state, and whether it has a review.
3. List the CR worktrees. For each, confirm the owner session and whether it is clean. Ask owners to release or remove merged ones, and never remove another session's worktree.
4. Re-check which design decisions Val has answered and where the references are. Update the inventory owners in docs/design/code-review.md.
5. Pick the one slice whose producer artifacts are accepted, and unblock it.
Record fixture, authenticated, live-provider and live-write evidence separately. Post a checkpoint with exact SHAs and the next action.

Code Review: one batched ask for Val (2026-11-01, `main` at `0a8c513`)

Every slice's code is on main. Lane C prepared everything that can be prepared without you, in #472 (docs and evidence only):

the CR-8 matrix refresh, with no "not started" rows;
CR-6's denial tests re-run on Claude Code 2.1.316: native 6/6 ([CR-6][P1] Add native AI review, PR questions and scoped review instructions #325);
the CR-6 live-smoke procedure and the CR-7 canary procedure ([CR-7][P1] Add explicit review actions and head-guarded GitHub merge #326);
the VoiceOver kit rebuilt on 0a8c513 ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322).

Nothing live has run, and the app has made no GitHub write.

Please answer here, or on the issue named:

A repository and PR for CR-6's live review smoke ([CR-6][P1] Add native AI review, PR questions and scoped review instructions #325).
- It uses your own claude login and reads through gh without writing.
- Something small is best, to keep provider cost low.
- Reply with owner/repo#n and "go".
A disposable repository for CR-7's live-write canary ([CR-7][P1] Add explicit review actions and head-guarded GitHub merge #326).
- It should be private and empty, with no required checks or merge queue, and allow squash.
- You open the throwaway PR there yourself; the procedure lists the commands.
- The canary posts a comment and a review, closes, reopens, drafts and readies the PR, edits its title, requests a reviewer if a second account is available, and squash-merges it.
- It never runs on TypeSafeAI/modex. Reply with owner/repo and "go".
Decision 1 (avatars, C16/C81): login only (D1), or fetch avatars?
Decision 5 (Viewed toggle, C65): local only, or GitHub's viewed state?
Decision 6 (new-tab button, C07): what does it open on this surface? Until you decide, it stays unavailable.
Decision 7 (access-mode indicator, C82): the reference's PR composer shows full access, but review is read-only. Should this composer show an access mode at all, and if so, which? Today the panel states the confinement in words and draws no indicator.

Also yours, whenever you're ready:

the VoiceOver pass (steps on [CR-3][P1] Build the PR inbox and resizable Code Review workspace #322);
R1–R4 references or acceptance of re-captures;
when the modex.codeReview packaged-build switch goes (staged plan in the matrix).

Agent follow-ups that don't need you:

C22 still uses the row's locator (D6);
the a11y spec covers only CR-3's controls (C76);
seven built controls have no asserting spec.

test(code-review): measure composer clearance beside the inline Ask s…

docs: record the C61 follow-up's full run (#319, #327)

[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319

Description

Outcome and scope

Verified baseline and reuse

Reference states

Created issue graph and dependency rules

Delivery gates

Agent execution contract

Security and correctness gates

Primary sources and how they affect this plan

Long-running goal prompt

Follow-up / resume prompt

Independent review prompt

Completion checklist

Activity

BunsDev commented on Oct 20, 2026

Gate A checkpoint 1 (2026-10-19): CR-1 ready for review

What CR-1 delivers

Review

Evidence on 6d77cbd

Still open

For #321

Next action

BunsDev commented on Oct 20, 2026

Gate A checkpoint 2 (2026-10-20): CR-1 landed; CR-2 and CR-3 started

CR-1 landed

Decisions since checkpoint 1

CR-2 and CR-3

Open for Val

Next action

BunsDev commented on Oct 20, 2026

Gate A checkpoint 3 (2026-10-20): CR-2 and CR-3 in review

Open for Val

Next action

mariscate commented on Oct 22, 2026

Refreshed goal prompts — 2026-10-22 (main at 0e674e7)

Long-running goal prompt

Follow-up / resume prompt

BunsDev commented on Oct 23, 2026

Refreshed goal prompts — 2026-10-23 (main at 203a350)

Long-running goal prompt

Follow-up / resume prompt

BunsDev commented on Nov 1, 2026

Code Review: one batched ask for Val (2026-11-01, main at 0a8c513)

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Evidence on `6d77cbd`

Refreshed goal prompts — 2026-10-22 (`main` at `0e674e7`)

Refreshed goal prompts — 2026-10-23 (`main` at `203a350`)

Code Review: one batched ask for Val (2026-11-01, `main` at `0a8c513`)