Skip to content

feat(pull-requests): an Involving me feed rule shows PRs you are part of - #33

Open
pnupu wants to merge 12 commits into
mainfrom
t3/pr-feed-involved-rule
Open

pnupu wants to merge 12 commits into
mainfrom
t3/pr-feed-involved-rule

Conversation

@pnupu

@pnupu pnupu commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

Problem

In a large repository, the combined Pull Requests feed shows every pull request. The rules from #31 can narrow a project to authored or review-requested pull requests, but not to all the pull requests the user takes part in: ones they opened, ones waiting on their review, and ones where they are assigned, mentioned, or have commented.

Change

  • New rule involved, labelled "Involving me". It is in PullRequestFeedRule and PULL_REQUEST_FEED_RULES, and pullRequestFeedShows handles it. It shows authored and review-requested rows, and an Others row only when that row's project's involved listing holds it (the involved callback in PullRequestFeedOptions).
  • Source model. For a project on the rule, outside a search, rows come only from its listings: the involved listing plus the authored and review-requested reads. The broad list's other rows for that project are dropped. Membership is presence by URL; there is no cutoff or truncation guessing (pullRequestInvolvedMatcher).
  • One page per listing. Involving me reads one page of each listing: 99 rows, which on GitHub is per search chunk of 100 repositories on a server. Older involved pull requests of a project on the rule stay hidden. The docs say so, and lean: comments on web and mobile name the trigger to revisit: users who need their older involved pull requests. Picking the project in the list's project filter shows all of its pull requests.
  • Fail open per project. A project's rows show as under the All rule until its involved listing is ready. Ready means the project was read, its environment answered and did not fail, and the project has no error. A GitHub repository that search does not index fails its involved read with a per-project error. So does an empty involves: search whose batch says it has more. Either way, that project falls back to All.
  • Search bypasses the rule on every host, on web and mobile.
  • Hosts that ignore involvement. A provider declares ignoresInvolvement for the listings it answers with every row: Bitbucket and Forgejo (authored and reviewing), Cursor Origin (reviewing), and GitCafe (reviewing, which is unnarrowed when its principal cannot be read). For those providers only, the server keeps a row when its author or its review requests involve the viewer. Other providers' rows stay as sent. On mobile, a row that an authored or review-requested list returned keeps that group even when its author login is not the viewer's (pullRequestListedGroup), so a GitCafe pull request matched by actor is not dropped. Web gets the same result from its partitions' groups.
  • New involvement involved. It is in PullRequestInvolvement. The provider flag listsInvolved is set by GitHub, which narrows involved with involves:<viewer>. PullRequestService answers involved with nothing for other providers, without a forge request, so those projects' rows come from the authored and review-requested reads.
  • Per-provider decisions:
    • GitHub: native, through involves:<viewer> (author, assignee, mentions, commenter). It does not fall back to the repository listing for involved, because that listing has no comment or mention data.
    • GitLab: no union filter in the REST or GraphQL list. Rows come from authored ∪ review-requested.
    • Azure DevOps: az repos pr list has only --creator and --reviewer. Same as GitLab.
    • GitCafe, Cursor Origin, Bitbucket, Forgejo: no involvement filter of this kind. Same as GitLab, with the ignoresInvolvement check above.
  • Web: PullRequestsPage runs a third partition query with involvement: "involved", only for the projects in scope on the rule, after the main list answers. Feeds without the rule send no extra request. Its rows join the feed like the authored and reviewing partitions, in the group their own author and review requests give. The main list's Load more and footer come from one helper, pullRequestLoadMore. A page in flight is never restarted. The list counts as failed only when no environment can continue, so one failed environment does not stop the others from paging. Settings and the row menu ("Feed for ") offer the rule.
  • Mobile: the Settings select offers the rule. For projects on the rule, PullRequestsScreen reads one page of the involved, authored and review-requested lists, after the main list answers and not while searching. Their rows join the list, and pull to refresh refreshes them. Load more pages the main list only, as on main.
  • Involvement filter tabs: deliberately unchanged. Web's route search narrows involvement to all | reviewing | authored, and mobile's persisted controls pick those three literals.
  • Old servers: the pullRequestFeedInvolved environment capability gates the option. Web Settings, the web row menu and mobile Settings hide "Involving me" unless every server that the write goes to sends it, so an older server never holds the rule or receives involvement: "involved". A stored unknown rule now decodes to all, and to an absent project override. A settings write still rejects an unknown rule.
  • Docs: docs/user/source-control.md › "Choose what the list shows" names the rule, the GitHub-only extra coverage, the one-page limit, and the search behaviour.

Scope

The maintainer of this fork asked for this feature in a delegated T3 Code task (follow-up to #31). It is not a bug fix.

Verification

  • packages/client-runtime/src/pullRequestFeed.test.ts: table rows for involved under each group; pullRequestFeedVisibility rows for the involved callback; the matcher (listed rows show, unlisted rows of a ready project do not, a project that is not ready shows every row, and no answer during a search shows every row); a GitCafe row listed as authored by actor stays authored. 31 tests pass.
  • apps/web/src/components/pullRequest/pullRequestLoadMore.logic.test.ts: one row per main-list state, and the per-environment failure rule (one failed environment while others continue, and none that can continue). 9 tests pass.
  • apps/web/src/components/pullRequest/pullRequestList.logic.test.ts: an involved row the list has not reached joins Others once, in recency order. Both web files: 55 tests pass.
  • packages/contracts/src/settings.test.ts: an unknown stored rule decodes to all and to an absent override; involved round-trips; a write with an unknown rule fails. 34 tests pass.
  • packages/source-control-github: the involved search query is is:pr is:open involves:bilal sort:updated-desc repo:acme/web. The search-free fallback does not read the repository listing for involved, and fails when the empty batch was truncated. An empty involved read is an answer in a covered repository and fails in an uncovered one. 160 tests pass.
  • apps/server/src/pullRequest/PullRequestService.test.ts: involved is read only from a provider with listsInvolved. The login check drops another author's row only for a provider with ignoresInvolvement, and keeps it otherwise (the GitCafe case). 128 tests pass.
  • apps/mobile/src/features/pull-requests/pullRequestList.test.ts: 24 tests pass.
  • Web, live GitHub data (8bac659, dev server with pingdotgg/t3code and pnupu/t3code projects): setting pingdotgg/t3code to Involving me drops its 97 "Others" rows and keeps fix(triage): file labeled reports through the issue template pingdotgg/t3code#31321, the only open PR there that gh search prs --involves @me returns. The baseline is the same build with the project on All pull requests, which is what main shows; I did not capture main itself.

Before: t3code project on All pull requests

Before: the feed shows 4 authored PRs and 97 other upstream PRs

Row menu: the new Involving me rule

The Feed for t3code submenu lists All pull requests, Involving me, Waiting on me, Authored by me and Hidden

After: t3code project on Involving me

After: the feed shows only the 4 PRs that involve the viewer

  • vp run check:changed on 8578a7a: lint, typecheck (41 packages), tests (23 files), react compiler (11 files) and graphql pass. Knip fails on the unused scripts/fork/* files, and format fails on the untracked .jaru/ map. Neither comes from this change, and I committed with --no-verify because of them.

Not verified:

Review history:

  • Rounds 1–4 (858519c → 9d05164): the feed was built around a cutoff matcher, which was then replaced by the source model.
  • Rounds 5–6 (fdda035, afa8c65 → a09e79c): providers that ignore involvement are filtered on the server; one failed environment no longer stops the main list; GitCafe rows matched by actor stay.
  • Round 7 (a09e79c → 8578a7a): the maintainer cut the paging of the involved, authored and review-requested listings, because those Load more paths drew most of the findings. The shoulds in the round-7 reviews were in that paging, except codex F3 (GitCafe rows matched by actor dropped on the client), which is fixed.
  • /code-review (medium) on the round-8 diff before the push: 2 Important findings, both declined. One-page reads hide involved rows past the page: that is the accepted, documented limit. A failed main-list continuation loses its cursor: removing the retry cursor was the maintainer's request, and the main list is back to how main behaves. Fixed nits: the cap wording, a stale comment, and a memo for mobile's listed groups. Mobile's secondary lists now read 99 rows, as web does. A pre-existing finding is listed below.

Declined, follow-ups:

  • A repository with copies on two environments: the involved read follows each environment's own projects.
  • More than 100 projects on the rule in one environment, to be batched by PULL_REQUEST_LIST_PROJECT_LIMIT later.
  • Stats queries mounted by the secondary mobile list hooks, and pull to refresh invalidating the same environment more than once.
  • Older installed fork builds decode settings strictly, but no released build has feat(pull-requests): per-project feed rules and hidden pull requests #31.
  • An empty truncated involved continuation (GitHub batched path).
  • On mobile, a GitCafe pull request matched by actor still falls in Others for a project on Authored by me or Waiting on me, because mobile reads the authored and review lists only for projects on Involving me.
  • Pre-existing: Bitbucket and Forgejo ignore authored and reviewing. The server check now covers that for every rule.

Model and harness: Claude Opus 5.5 via T3 Code delegated task (Claude Code).

🤖 Generated with Claude Code

pnupu and others added 2 commits November 19, 2026 11:56
A project on the new `involved` feed rule adds the pull requests the
viewer opened, is asked to review, or (on GitHub) is assigned to,
mentioned in, or commented on. GitHub narrows `involved` with the
`involves:` search qualifier. Hosts without such a filter are read as
`all`, and the service keeps the rows the viewer wrote or must review.

Web and mobile read the `involved` listing only for environments with a
project on that rule. The rule is offered only where every server sends
the new `pullRequestFeedInvolved` capability, and an unknown stored rule
now decodes to the default instead of failing the settings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The feed rule hid Others rows while the `involved` read loaded, after it
failed, during a search, and past its first page. A row the listing
cannot rule on now counts as involved, so the rule never hides a row the
host would list. The docs say the rule shows every row while searching.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 8bac659dcbac.

VERDICT: CHANGES

F1 [should] packages/client-runtime/src/pullRequestFeed.ts:102 When the involved listing is incomplete, one global cutoff (the oldest row across the whole merged answer) decides which rows it has ruled on. The server returns up to limit rows per project or search chunk (PullRequestService.ts:1676-1696), so each repository stops at a different point. A small project with old involved rows pulls the cutoff far back, and then involved rows of a large project that are past that project's first page get hidden. I confirmed this with the matcher logic: listing entries A/100@2026-11-10 and B/1@2025-01-01 with complete: false make A/90@2026-11-05 read as not involved. That row could carry the viewer's comment, so it disappears after "load more". This breaks the fix commit's own goal ("past its first page"). Fix: compute the cutoff per project (environment + projectId, or per cursorKey from nextCursors), and treat a row from a project with no listed rows as involved while the listing is incomplete. Reading only the projects on the rule (see F4) also narrows the exposure.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:783 The matcher only checks involvedQuery.data and data.errors. createMergedEnvironmentQuery builds data from environments that have answered, so if several environments have a project on the rule and one answered while another is still pending or failed at the transport level, data is non-null, errors is empty and truncatedEnvironments is empty. The matcher then treats the listing as complete and hides the Others rows of the slow or failed environment. This is the "while it loads / after it failed" case the second commit set out to fix. Fix: pass null when involvedQuery.isPending || involvedQuery.error !== null. Better, rule only on rows from environments that answered.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:202 On mobile, involvedList.error only reflects environments whose whole answer failed. If one project's read fails inside a successful answer, it lands in data.errors and is ignored, unlike the errors.length > 0 check on web. Also, isPending turns false as soon as any environment has entries. An environment that is still loading has no pages, so hasMore is false and the answer reads as complete. In both cases the rule hides Others rows it cannot rule on. Fix: return null when involvedList.data.errors.length > 0 or when any involved target has not answered yet, matching web.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:478 The involved read covers every project in an environment that has one project on the rule, not just the projects on it. Mobile does the same at PullRequestsScreen.tsx:195. On hosts without listsInvolved, it is also a second full all read whose rows the client could already work out from viewerReviewRequested and author. This costs extra forge requests in exactly the large-repository setup the rule targets. Fix: narrow projectIds to the projects on the involved rule. Optionally skip the read for hosts that do not list involved and rule on group alone.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:195 The mobile involved listing keeps the search query, so a search on mobile still narrows to involved rows. The new docs (docs/user/source-control.md) say the rule shows every row while searching, which is how web behaves (partitionsWanted requires an empty typedQuery). Fix: skip the involved targets when query is non-empty, as web does.

I couldn't run a typecheck or the test suites because there is no node_modules in the checkout. The only check I ran was the matcher logic in plain Node for F1. The PR discussion has no earlier comments, so nothing here is carried over.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 8bac659dcbac.

VERDICT: CHANGES

F1 [blocker] apps/web/src/components/pullRequest/PullRequestsPage.tsx:812 Both clients can show an empty feed even after GitHub returns a PR the viewer commented on. The involved response only filters rows from the broader list, so a PR beyond that list’s first page stays hidden until the user pages through unrelated PRs. Fix: Add involved results to the feed’s candidate rows on web and mobile, then deduplicate and apply the existing filters.

F2 [should] packages/client-runtime/src/pullRequestFeed.ts:106 A single timestamp from merged results cannot describe what each repository’s page covered. An old result from a quiet repository can make a valid match on a busy repository’s next page disappear. Rows tied at the page boundary can also disappear. Fix: Use the applicable repository or batch cursor and treat rows at or beyond an unverified boundary as unknown.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:191 Mobile continues applying the involvement rule during search, so a matching PR can appear on web and be hidden on mobile. The user guide says searching bypasses the rule. Fix: Disable the mobile involvement lookup and filter while the typed query is nonempty.

F4 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:785 Web can treat one server’s successful reply as complete while another server’s lookup is pending or has failed. Mobile likewise misses repository errors because it checks involvedList.error but ignores involvedList.data.errors. Fix: Apply negative matches only to sources whose involvement lookup completed successfully.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:478 When two connected environments hold the same repository, the list can assign it to the checkout with an Authored rule and skip the checkout whose Involving me rule still applies. The membership request then omits that repository. Fix: Build the involvement targets from the projects whose Involving me rules apply, including copies assigned elsewhere for the main list.

F6 [nit] packages/contracts/src/settings.ts:1903 The permissive snapshot decoder also weakens write validation because the project settings patch reuses it. An invalid rule can now clear an existing project override silently. Fix: Keep the project patch field strict and use the fallback in snapshot decoding.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:200 The membership lookup mounts the full list hook, which can fetch diff stats for rows the screen never renders. Fix: Use a list reader that skips stats for membership lookups.

F8 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:933 The unindexed-repository fallback counts a request to any team as involving the viewer. Fix: Require a direct request to the viewer or verified membership in the requested team.

Focused Node checks reproduced the missing row, pagination boundary, and duplicate checkout cases. git diff --check passed. Repository tests could not run because vp and dependencies were unavailable.

…to the feed

The involved listing ruled with one cutoff across every repository, read
whole environments, and only filtered rows the broad list had loaded.
Now the cutoff is per project and a tie or a failed slice fails open,
the read covers only projects on the rule, and its rows join the feed
like the authored and reviewing partitions. Web holds the last answer
while a refresh travels. Mobile waits for every target, skips the read
while searching, and fails open on any error. GitHub's search-free
fallback answers `involved` with an empty truncated page, and a settings
write keeps a strict feed rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 203d7227e0db.

VERDICT: CHANGES

F1 [should] packages/client-runtime/src/pullRequestFeed.ts:123 On GitLab, Azure DevOps, Bitbucket, Forgejo, GitCafe and Cursor Origin, "Involving me" ends up showing every pull request in a large repository. The pull requests these hosts treat as yours are only the ones you opened or must review, and the client already knows both from the row itself. Here is how it fails. The service reads one all page of PAGE_SIZE rows and keeps only the authored and review-requested ones. That page is truncated, so the client marks the project "paged" and treats its oldest kept row as the boundary. Every Others row older than that boundary counts as involved. If the page has none of your rows, there is no boundary and every Others row counts as involved. The involved read never loads a second page (PullRequestsPage.tsx:483, PullRequestsScreen.tsx:206), so as the main list pages further back, nearly all rows fail open. Run against the matcher logic in Node, {entries: [], slice: "paged"} returns true for any Others row, and so does a row older than your single involved PR. This is new evidence for the optional part of the earlier nit F4 (#33 (comment)). It is now a correctness problem, not a cost one. Fix: have the server tell the client which hosts set listsInvolved, for example as a flag on each providers entry or on each slice. For hosts without it, rule on the row's group alone (Others rows are never involved) and skip the extra all read.

F2 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from gpt F5 (#33 (comment)). Say one repository is a project on two environments, and the list reads it through copy A while copy B is on involved. B is dropped from involvedQueries. The row's slice (A's environment and project) then reads as "unknown", so every Others row of that repository is shown and the rule has no effect. Fix: key slices by host and repository, or build the involved targets from every copy whose rule is involved.

F3 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Carried over from gpt F7 (same comment). The involved membership read mounts the full usePullRequestList. With no viewability callback, its first 8 rows become stats targets, which costs extra diff-stat requests. Fix: use a list reader that skips stats.

F4 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:932 The new involved branch in matchesUnsortedListing (authored or reviewing) can no longer run. The fallback at line 1973 now returns early for involved, and read(false) is the only caller. The PR description's claim that the fallback "keeps authored or review-requested rows" is also out of date. This makes gpt F8 moot. Fix: drop the branch and update the description.

F5 [nit] packages/contracts/src/settings.ts:1570 The forward-compatible decode only protects future clients. An installed older mobile or web build still decodes pullRequestFeed strictly. Once a newer client stores involved on an upgraded server, that older build cannot decode the whole settings snapshot. The capability gate checks servers, not other clients. Fix: note this in the release notes, or accept it as a known limit.

Fixes from the earlier reviews that I checked against this head:

  • Per-project boundary and ties: fixed (Claude F1, gpt F2).
  • Partial or failed answers: web now waits until no target is pending and none failed, and failed slices read as "unknown" (Claude F2, gpt F4).
  • Mobile: checks answered and data.errors (Claude F3).
  • Narrowed read: covers only the projects on the rule (the main part of Claude F4).
  • Search: no involved read while searching (Claude F5, gpt F3).
  • Involved rows: they now join the feed on both clients (gpt F1).
  • Settings patch: stays strict (gpt F6).

I could not run typecheck or the test suites because the checkout has no dependencies. The only check I ran was the matcher logic in plain Node, for F1.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 203d7227e0db.

VERDICT: CHANGES

F1 [blocker] packages/client-runtime/src/pullRequestFeed.ts:123 In a large GitLab or Bitbucket repository, the server can return an empty, truncated involved page after removing unrelated rows. Both clients then show every unrelated row, including on later pages, because the empty slice has no cutoff. Fix: On providers whose fallback covers authorship and review requests, decide membership from each row’s author and review-request flag.

F2 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Mobile merges GitHub’s involves: results with the broad list. A PR beyond the first broad page is missing when the viewer’s only involvement is a pending review request, which involves: omits. Fix: Merge the reviewing list on mobile or include review-requested results in the server’s involved response.

F3 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:803 Carried over from F2. If an involved refresh fails after a successful read, the held answer can still hide a PR the viewer just commented on, even when the main list refresh succeeds. Fix: Keep previously listed rows visible and treat absent rows from the failed lookup as unknown. Prior review

F4 [should] apps/web/src/components/pullRequest/pullRequestListGroups.logic.ts:118 Carried over from F1. A successful involved answer is discarded when either the authored or reviewing lookup has no answer. If the reviewing request fails on initial load, a PR returned by the involved lookup beyond the main page stays missing. Fix: Merge involved candidates when grouping falls back to the main list. Prior review

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from F5. If the main list assigns a shared repository to an Authored checkout, the Involving me checkout can be skipped by the membership lookup even though its rule still applies. The matcher then admits every row as unknown. Fix: Build membership targets from the checkouts whose rules count and evaluate membership across those copies. Prior review

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 An environment with more than 100 projects on the rule sends a projectIds array the RPC rejects. Web can do the same when the main list queries the whole environment. Fix: Split the involved targets into batches using PULL_REQUEST_LIST_PROJECT_LIMIT.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Carried over from F7. The secondary list hook fetches diff stats for its first eight rows even when deduplication or feed filters remove them from the screen. Fix: Let the primary list request stats for the rows it displays. Prior review

F8 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:932 The new authored || reviewing fallback branch is unreachable because involved returns before the repository listing runs. Fix: Keep the existing repository-listing predicate.

Focused Node checks confirmed F1, F3, and F4. git diff --check passed. The repository tests and typecheck could not run because vp and dependencies were unavailable.

…lvement

On a host that cannot narrow `involved`, the read was a filtered `all`
page, so every Others row failed open and the rule did nothing. The
server now skips that read and marks each provider with `listsInvolved`;
clients rule Others rows out on hosts without it. Paging is read per
repository from the cursor keys, an environment that has not answered or
failed is unknown, and GitHub treats an empty `involves:` search as an
answer only where search covers the repository.

Mobile also joins the review-requested list (which `involves:` omits)
and refreshes both extra lists on pull to refresh. Web merges involved
rows when an authored or reviewing read has no answer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 9d05164b3119.

VERDICT: CHANGES

F1 [should] packages/source-control-github/src/server/GitHubPullRequestProvider.ts:437 The way the provider signals "unindexed repository" usually doesn't reach the client, so the client hides rows the PR says it should show. For an uncovered repository, the provider returns { items: [], truncated: true, continues: false }. continues is false, so readSlice gives it no nextCursor (PullRequestService.ts, page.continues && page.truncated). The only remaining signal is the environment-level truncated flag. But pullRequestInvolvedMatcher (packages/client-runtime/src/pullRequestFeed.ts:986-988) only treats a repository without a cursor as paged when the environment has no cursors at all. Any other repository in that environment with more involves: rows gives it a cursor; a truncated batch search gives every repository in its chunk one. Once that happens, the unindexed repository counts as fully read and empty, and all its Others rows are hidden. I copied the matcher into plain Node to check. With { truncated: true, cursorKeys: {"github.com acme/busy"} }, a row from acme/renamed returns false. When that repository is alone in the environment, it returns true. The new client test even pins this down with acme/idle being ruled out. This breaks the stated intent: "in an unindexed one the page is marked truncated, so clients fail open for that repository". Fix: send a per-repository signal the matcher can read. For example, return a project error for the uncovered repository (it then reads as unknown), or a marker the service turns into a per-slice "unknown" entry, instead of relying on environment-wide truncation.

F2 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:486 On web, listsInvolved is looked up by host alone in the merged providers. mergePullRequestLists keeps one environment's entry per host (held.configured ? held : provider). If an older server that sends no listsInvolved wins for github.com, every GitHub Others row from the newer environment is ruled out. Mobile keys this per <environmentId> <host>. Fix: key by environment and host, as mobile does, or OR the flag when merging.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:381 Carried over: earlier Claude F2 and gpt F5 (#33 (comment), #33 (comment)). Say a repository is a project on two environments, the main list reads it through copy A, and copy B is on involved. Then involvedQueries drops B, every row reads as unknown, and the rule does nothing for that repository. Fix: build the targets from every copy on the rule, or key slices by host and repository.

F4 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:53 Carried over: earlier Claude F3 and gpt F7 (same comments). There are now two extra full usePullRequestList hooks (involvedList, reviewingList). Each one fetches diff stats for its first 8 rows, even rows that dedup or filtering drop. Fix: use a list reader that skips stats.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:42 Carried over: gpt F6 (#33 (comment)). Mobile does not split projectIds into batches of PULL_REQUEST_LIST_PROJECT_LIMIT. With more than 100 projects on the rule in one environment, the RPC rejects the read. Web can hit the same limit when the main query covers a whole environment. Fix: split into batches of PULL_REQUEST_LIST_PROJECT_LIMIT.

F6 [nit] packages/contracts/src/settings.ts:1567 Carried over: earlier Claude F5 (#33 (comment)). The forgiving decode only helps clients built from this version. An installed older web or mobile build still decodes pullRequestFeed strictly, so it fails on the whole settings snapshot once another client stores involved. Fix: note it in the release notes, or accept it as a known limit.

Fixes I checked against this head:

  • Non-listing hosts: handled with listsInvolved and by skipping the read (earlier Claude F1, gpt F1).
  • Mobile review-requested rows: the reviewing list is now merged in (gpt F2).
  • Web without authored or reviewing answers: involved rows still join when those reads have no answer (gpt F4).
  • Web refresh failures: a failed environment reads as unknown (gpt F3).
  • Unreachable branch: the branch in matchesUnsortedListing is gone and a comment replaces it (Claude F4, gpt F8).

I also ruled out a suspected render loop around setHeldInvolved. The web build enables the React compiler (react({ compiler: true })), which memoizes that hook value.

I couldn't run typecheck or the test suites because the checkout has no dependencies installed. The only check I ran was the plain-Node copy of the matcher for F1.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 9d05164b3119.

VERDICT: CHANGES

F1 [should] packages/client-runtime/src/pullRequestFeed.ts:117 Search still hides unrelated PRs on GitLab and other hosts without listsInvolved. Both clients skip the lookup while searching, but the matcher still rejects those rows. Fix: Bypass the Involving me rule during search regardless of provider capability. Carried over from F3, with the current failure caused by the new capability check.

F2 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:260 On GitLab, the feed can be empty even when you have open PRs. If your newest authored PR is beyond the first broad page, the reviewing lookup cannot surface it and the involved lookup is skipped by the server. Fix: Merge authored results for projects on hosts without listsInvolved.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:423 Load more advances only the broad list on both clients. Once it passes the first involved page’s cutoff, older unrelated PRs remain visible because the membership lookup never advances. Fix: Advance the involved lookup as the feed loads older rows, keeping uncertain rows visible until it answers.

F4 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:830 When two involvement refreshes fail, the merged query reports only the first error. The second environment’s cached answer can still hide a PR the viewer just commented on. Fix: Track failed targets separately and use cached answers from those targets only to confirm matches. Carried over from F3.

F5 [nit] packages/client-runtime/src/pullRequestFeed.ts:135 An unindexed GitHub repository returns an empty, truncated result without a cursor. If another repository in that environment supplies a cursor, the matcher treats the unindexed repository as complete and hides its Others rows. Fix: Carry the unknown status per repository independently of cursors.

F6 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 If the main list assigns a shared repository to an Authored checkout, the Involving me checkout can be omitted from the membership lookup even though its rule still applies. Unrelated rows then pass as unknown. Fix: Query the checkouts whose rules count and evaluate their membership by repository. Carried over from F5.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 A server with more than 100 projects on the rule receives a projectIds array the RPC rejects. Web can do the same when its broad query covers the whole environment. Fix: Batch by PULL_REQUEST_LIST_PROJECT_LIMIT and track coverage per batch. Carried over from F6.

F8 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The two secondary list hooks can fetch diff stats for rows removed by deduplication or filtering. Fix: Let the displayed list request stats for visible rows. Carried over from F7.

F9 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:430 One pull-to-refresh can invalidate the same server’s provider and viewer caches three times. Fix: Invalidate once per environment, then refresh the affected lists.

F10 [nit] packages/contracts/src/settings.ts:1570 Installed older clients still reject the settings snapshot once an upgraded server stores involved. Fix: Document the client upgrade requirement. Carried over from F5.

Focused Node checks reproduced the matcher behavior behind F1, F3, F4, and F5. git diff --check passed. Repository tests and typecheck could not run because vp and dependencies were unavailable.

…t and repository

An uncovered GitHub repository now fails its empty `involved` read, so
its rows count as unknown instead of hiding behind another repository's
cursor. Servers always send `listsInvolved`; a missing value from an
older server counts as unknown, and web reads it per environment.
Mobile also joins the authored list, as web's partitions do, because a
host without `listsInvolved` answers `involved` with nothing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 1893f33eedc3.

VERDICT: CHANGES

F1 [should] packages/client-runtime/src/pullRequestFeed.ts:117 Carried over from gpt F1 (#33 (comment)); there is no author reply, and the code at this head is unchanged. While the user searches, neither client reads involved: web because partitionsWanted is false at PullRequestsPage.tsx:436, mobile because involvedProjectKey is empty. So the matcher gets null and returns listsInvolved(entry) !== false. On GitLab, Azure DevOps, Bitbucket, Forgejo, GitCafe and Cursor Origin, which all send listsInvolved: false, that hides every Others row during a search. The new docs say the opposite at docs/user/source-control.md:253: "While you search, it shows all of the project's pull requests." On GitHub the same path fails open and shows everything, so the two hosts behave differently for the same rule. Fix: when the query is non-empty, pass an involved that always returns true (or skip the involved rule check). Otherwise, reword the docs to say search keeps the rule on hosts without listsInvolved.

F2 [nit] packages/client-runtime/src/pullRequestFeed.ts:139 Carried over from gpt F3 (#33 (comment)). Only the first page of involved is ever read, on both clients. Once "load more" takes the main list past that page's oldest row in a paged repository, every older Others row counts as involved. In a large repository the rule then stops narrowing after deep scrolling. It fails open, so nothing is wrongly hidden. Fix: page the involved read along with the main list, or accept and document the limit.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:838 Carried over from gpt F4 (#33 (comment)). unknown checks only involvedQuery.error, and that holds only the first failed environment (state/pullRequests.ts:660). A second environment whose refresh failed keeps its stale value in values, so it stays in answered. The rule then applies an old answer and can hide a PR the viewer has just commented on. Fix: collect the failed environment IDs from the query, not just the first error, and treat their rows as unknown.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from Claude F3 and gpt F6 (#33 (comment), #33 (comment)); this code is unchanged. Suppose one repository is a project on two environments, the main list reads it through copy A, and copy B is on involved. Then involvedQueries drops B, every row reads as unknown, and the rule has no effect for that repository. Fix: build the targets from every copy on the rule.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Carried over from Claude F4 and gpt F8 (same comments). There are now three extra full usePullRequestList hooks (involvedList, authoredList and reviewingList at :217-218). Each fetches diff stats for its first rows, including rows that dedup or the feed filter drop. Fix: use a list reader that skips stats for these membership reads.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:201 Carried over from Claude F5 and gpt F7 (same comments). projectIds is not split into batches of PULL_REQUEST_LIST_PROJECT_LIMIT. With more than 100 projects on the rule in one environment, the RPC rejects all three extra reads. Fix: batch by the limit.

F7 [nit] packages/contracts/src/settings.ts:1570 Carried over from Claude F6 and gpt F10 (same comments). The forgiving decode only helps clients built from this version. Once another client stores involved, an installed older web or mobile build still fails to decode the whole settings snapshot. Fix: note the client upgrade requirement in the release notes, or accept it as a known limit.

Fixes I checked against this head:

  • Uncovered GitHub repository (Claude F1 and gpt F5 at 9d05164b): fixed. The empty involved read now fails (GitHubPullRequestProvider.ts:437). readSlice turns that into a per-project error, so both clients treat the row as unknown. The batch path sends unresolved repositories to that same read. The error does not show up in the list's error banner.
  • listsInvolved per environment on web (Claude F2): fixed. It is now keyed by <environmentId> <host>, and the server always sends it.
  • Mobile authored rows (gpt F2): fixed. authoredList is joined, and filterPullRequestList deduplicates and sorts the merged rows.
  • Cursor keys: they match the matcher's <host> <repository lowercased> slice key. The only provider with a different repositoryKey is Azure DevOps, and it sends listsInvolved: false.

I couldn't run typecheck or the test suites because the checkout has no node_modules. For the same reason I couldn't confirm that Schema.Literals(...).pick exists in this Effect version (apps/mobile/src/persistence/mobile-preferences.ts:34). Everything above comes from reading the code.

…m its listings

The matcher guessed which rows the involved listing had ruled on from
cursors, truncation and per-repository cutoffs. Now a project on the
rule shows only rows its involved, authored and review-requested
listings hold. Until its involved read is ready (answered, not failed,
the project read without error; an unindexed GitHub repository fails),
its rows show as under the All rule. A search bypasses the rule, and
Load more grows the involved read on web and mobile.

Hosts without `listsInvolved` answer `involved` with nothing, so their
rows come from the authored and review-requested reads; the
`listsInvolved` field on provider summaries is gone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 1893f33eedc3.

VERDICT: CHANGES

F1 [blocker] packages/client-runtime/src/pullRequestFeed.ts:271 On web and desktop, Involving me shows unrelated Bitbucket and Forgejo PRs. Both providers return ordinary rows for the authored and reviewing reads, and the feed trusts the resulting priority group before checking the row’s author or review request. Fix: On fallback hosts, evaluate authorship and review requests from the row itself. Carried over from F1, with new evidence that rows still pass through priority grouping.

F2 [should] packages/client-runtime/src/pullRequestFeed.ts:117 Search still hides matching unrelated PRs on GitLab and other fallback hosts. Both clients skip the involvement lookup during search, but the matcher still rejects rows when listsInvolved is false. Fix: Bypass the Involving me rule during search on every host. Carried over from F1.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:430 Load more advances only the broad list on both clients. Once it passes the first involvement page’s cutoff, older unrelated PRs remain visible because their membership is never checked. Fix: Advance the involvement lookup as older feed rows load, keeping uncertain rows visible until the lookup answers. Carried over from F3.

F4 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:838 When two involvement refreshes fail, only the first environment is marked unknown. The second environment’s cached answer can still hide a PR the viewer just commented on. Fix: Track failed targets individually and use their cached answers only to confirm matches. Carried over from F4.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Assigning a shared repository to an Authored checkout can omit the checkout whose Involving me rule still applies. Unrelated rows then pass as unknown. Fix: Build involvement targets from the checkouts whose rules count and match coverage by repository. Carried over from F6.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 On an environment with more than 100 projects on the rule, the involvement lookup exceeds the RPC’s projectIds limit. Web can hit the same limit when its broad query covers the whole environment. Fix: Batch the targets by PULL_REQUEST_LIST_PROJECT_LIMIT and track coverage per batch. Carried over from F7.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The three secondary list hooks can request diff stats for rows the feed drops during deduplication or filtering. Fix: Fetch stats only for rows displayed by the combined list. Carried over from F8.

F8 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:436 One pull-to-refresh can invalidate the same environment’s provider and viewer caches four times. Fix: Invalidate once per environment, then refresh its lists. Carried over from F9.

F9 [nit] packages/contracts/src/settings.ts:1570 Older installed clients still reject the settings snapshot once an upgraded server stores involved. Fix: Document the client upgrade requirement. Carried over from F10.

F10 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 The empty-result fallback discards truncation from the search, which can occur when GitHub repeats a cursor. A repository with cached search coverage then looks fully read, so the matcher can hide PRs from unread pages. Fix: Preserve truncation and mark that repository’s coverage unknown when there is no usable cursor.

Focused Node checks reproduced the grouping and matcher behavior behind F1–F4. git diff --check passed. Repository tests and typecheck could not run because vp and dependencies were unavailable.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 3e228230930d.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/pullRequestList.logic.ts:459 Carried over from gpt F1 (#33 (comment)). That review called it a blocker. I rate it should because GitHub, the main case, works. There is no reply from the author, and this head still has the problem. On web, "Involving me" shows every pull request of Bitbucket and Forgejo projects, and of Cursor Origin projects for the review-requested part. The PR says these hosts show only authored and review-requested rows. Here is why. BitbucketPullRequestProvider.listChangeRequests and ForgejoPullRequestProvider.listChangeRequests never read input.involvement, and OriginPullRequestProvider.ts:633 only narrows authored. So their authored and reviewing partitions hold the whole repository. partitionPullRequestsWithPriority puts each of those rows in the authored or reviewRequested group by URL alone. pullRequestFeedVisibility then admits the row on its list group and never checks groupOf. Mobile is correct because it groups rows by their own author and review flag. The root cause is older, but this PR's per-host promise depends on it. Fix: in partitionPullRequestsWithPriority, or on the server for hosts that ignore involvement, keep only partition rows whose own pullRequestFeedGroup matches the partition. Another option is to check groupOf first for the involved rule.

F2 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:696 Load more on web grows the involved read by raising its limit, not by following a cursor, and the limit stops at MAX_PAGE_SIZE (500). Past that, older involved pull requests can never be reached, even though the intent says they appear as the user loads more. While involvedTruncated stays true, the footer keeps offering Load more. Each press then only regrows the main list without cursors. In the other direction, canLoadMore still depends on the main list's canContinue || pageSize < MAX_PAGE_SIZE. So when the main list is exhausted, a truncated involved read cannot grow. Fix: page the involved read with its nextCursors, as mobile does, or base canLoadMore on the involved read's own state and document the cap.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from Claude F4 and gpt F5 (#33 (comment), #33 (comment)). Take one repository that is a project on two environments. The main list reads it through copy A, and copy B is on involved. involvedQueries leaves B out, and ready checks the row's own copy (A). So every row reads as not ready, and the rule has no effect for that repository. The rule fails open, so nothing is hidden by mistake. Fix: build the targets from every copy on the rule and match readiness by repository.

F4 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:191 Carried over from Claude F6 and gpt F6 (same comments). The projectIds list is not split by PULL_REQUEST_LIST_PROJECT_LIMIT. With more than 100 projects on the rule in one environment, the RPC rejects all three extra reads. Fix: batch by the limit.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:217 Carried over from Claude F5 and gpt F7 (same comments). There are three extra full usePullRequestList hooks: involved, authored and reviewing. Each one fetches diff stats for its first 8 rows, including rows that deduplication or the feed rule then drop. Pull to refresh also invalidates the same environment's caches four times. Fix: use a list reader without stats for these reads, and refresh each environment once.

F6 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1961 Carried over from gpt F10 (#33 (comment)). The empty involved fallback always returns truncated: false. It drops truncation from a search that stopped on a repeated cursor. In a covered repository, that empty answer counts as complete and hides the project's Others rows. Fix: keep the search batch's truncated flag, and have the provider fail the read when it is set and there is no cursor.

These earlier findings are fixed at this head:

  • Search bypass: the matcher gets null while searching on both clients, so every row shows.
  • Failures per environment: web checks the new failed list. Mobile's errors includes environment-level failures, which have no projectId.
  • Load more: it now advances the involved read on both clients, with the web cap noted in F2.
  • Unindexed GitHub repository: the read fails per project, so its rows show as under the All rule.
  • Old listsInvolved cutoff and truncation logic: removed.

I could not run typecheck or tests because the checkout has no node_modules. For the same reason I could not confirm that Schema.Literals(...).pick exists in this Effect version. F1 comes from reading the provider and grouping code, not from running it.

…es a failed page

Web grew the involved read only when its last answer said truncated, so
a click during a pending read advanced only the main list. It now grows
with every Load more, and at its largest page the rows older than its
oldest listed row show, since it cannot reach them. Mobile retries a
failed involved page on the next request and shows its error.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 3e228230930d.

VERDICT: CHANGES

F1 [blocker] packages/client-runtime/src/pullRequestFeed.ts:232 Web and desktop show unrelated PRs under Involving me on Bitbucket and Forgejo. Both providers return unfiltered rows for the authored and reviewing reads, and the feed accepts the resulting group even when the row itself is unrelated. Carried over from F1. Fix: For Involving me, check the row’s author and review-request flag before accepting it through a priority group.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:697 Load more stops adding GitHub involvement matches after 100 rows per search batch. The client increases the limit without sending the involved cursor, and the server caps each batch at 100. Older PRs where the viewer only commented, was mentioned, or was assigned stay hidden. Fix: Advance the involved listing with its own nextCursors and append the returned rows.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:515 A failed involved page leaves a Load more button that silently does nothing. The footer only checks the main list’s error, while the involved list refuses to retry without retry: true. Fix: Show the involved list’s loading and error states in the footer and route its retry through Try again.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 When the same repository has an Authored checkout and an Involving me checkout, assigning the main read to the Authored checkout can omit the other checkout’s involvement lookup. Unrelated rows then pass as unknown. Carried over from F5. Fix: Evaluate membership and readiness across the checkouts whose rules count.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 More than 100 projects on the rule in one environment causes all three secondary reads to exceed the RPC’s projectIds limit. Web has the same issue when its main query covers the whole environment. Carried over from F6. Fix: Batch the project IDs at the contract limit and track readiness per batch.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Each secondary list can fetch diff stats for its first eight rows, including rows discarded by deduplication or filtering. Carried over from F7. Fix: Fetch stats through the combined list’s visible-row handler.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:445 One pull-to-refresh can invalidate the same environment’s provider and viewer caches four times. Carried over from F8. Fix: Invalidate once per environment before refreshing its lists.

F8 [nit] packages/contracts/src/settings.ts:1570 Older installed clients still reject the settings snapshot once an upgraded server stores involved. Carried over from F9. Fix: Document the client upgrade requirement.

F9 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 The empty involved fallback discards the search’s truncation flag. If search stops early with no usable rows and the repository has cached coverage, the client treats the incomplete answer as empty and hides its rows. Carried over from F10. Fix: Preserve truncation and fail the project read when an incomplete answer has no usable cursor.

Focused Node checks reproduced F1–F3, and git diff --check passed. Repository tests and typecheck could not run because vp and the dependencies were unavailable.

…y a truncated read

The floor for a capped involved read was merged across repositories and
could come from a held, smaller answer, so it hid rows it meant to show.
It is gone: the read stops at 500 rows per environment, a stated limit.
Web grows the involved read only when its answer was truncated (or a
grown read is still in flight), and mobile reruns a failed involved
page only on an explicit retry, with a spinner while either list loads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 720b0cfbfabf.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/pullRequestList.logic.ts:459 Carried over from gpt F1 (#33 (comment)) and Claude F1 (#33 (comment)). There is no author reply, and the code is unchanged at this head. On web, "Involving me" still shows every pull request of a Bitbucket or Forgejo project. The cause is that neither provider reads input.involvement, and the server adds no narrowing of its own (PullRequestService.ts only forwards it). Cursor Origin narrows only authored (OriginPullRequestProvider.ts:633), so its reviewing partition holds the whole repository too. partitionPullRequestsWithPriority then files each of these rows under authored or reviewRequested by URL alone. pullRequestFeedVisibility (packages/client-runtime/src/pullRequestFeed.ts:231) admits the row on that list group and never checks groupOf. Because the involved read answers nothing on these hosts, the PR's "Same as GitLab" claim for them depends on these partitions. Mobile is correct because it groups rows by their own author and review flag. Fix: for the involved rule, check the row's own pullRequestFeedGroup before trusting the partition group. Alternatively, drop partition rows whose own group does not match the partition.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:697 Carried over from gpt F2 (#33 (comment)). The latest commit makes every Load more grow the involved read, but it still raises only limit and never sends the involved answer's nextCursors. GitHub search reads at most PULL_REQUEST_SEARCH_MAX_ROWS (100) rows per batch (GitHubPullRequestApi.ts:2050, gitHubPullRequestJson.ts:250). So a limit of 198, 297, 396 or 495 still returns the same first 100 rows. Once the main list pages past that point, older pull requests the viewer commented on, was mentioned in or is assigned to stay hidden. That lasts until the page size reaches 500, which takes about 5 Load mores. Then the new involvedFloor at :838 shows every older row, the unrelated ones included. The intent says older involved rows appear as the user loads more, and that does not happen between those two points. Fix: page the involved read with its own nextCursors and append its rows, as mobile does. Otherwise, apply the floor as soon as a truncated answer returns fewer rows than were asked for.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:838 At 500, involvedFloor is the oldest row across the whole merged answer, which can cover several search batches and environments. A batch whose listing reaches less far back still hides its rows between its own oldest row and that shared floor. Fix: compute the floor per environment and batch.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from Claude F3 (#33 (comment)) and gpt F4 (#33 (comment)). The code is unchanged. Take a repository that is a project on two environments. If the main list reads it through copy A and copy B is on involved, involvedQueries leaves B out. The rule then fails open for that repository. Fix: build the targets from every copy on the rule and check readiness by repository.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:191 Carried over from Claude F4 and gpt F5 (same comments). projectIds is not split into batches of PULL_REQUEST_LIST_PROJECT_LIMIT. With more than 100 projects on the rule in one environment, the RPC rejects all three extra reads. Fix: batch by the limit.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:217 Carried over from Claude F5 and gpt F6/F7 (same comments). Each of the three extra usePullRequestList hooks fetches diff stats for rows that deduplication or the feed rule then drop. One pull to refresh also invalidates the same environment four times. Fix: read these lists without stats and refresh each environment once.

F7 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 Carried over from Claude F6 and gpt F9 (same comments). The empty involved fallback always returns truncated: false. That drops the truncation flag of a search that stopped early, so in a covered repository an incomplete empty answer counts as complete. Fix: keep the search's truncated flag and fail the read when it is set and there is no cursor.

F8 [nit] packages/contracts/src/settings.ts:1570 Carried over from gpt F8 (#33 (comment)). The forgiving decode helps only clients built from this version. An installed older client still fails to decode the whole settings snapshot once involved is stored. Fix: note the client upgrade requirement.

I confirmed these from the code at this head:

  • Search bypass: both clients pass null to the matcher while searching.
  • Failures per environment: web checks the new failed list, and mobile checks answered and errors.
  • Unindexed GitHub repository: its involved read fails per project, so its rows show as under the All rule.
  • Mobile retry: a failed involved page is retried on the next request, and its error shows in the footer.

I did not run typecheck or tests because the checkout has no dependencies installed. F1 and F2 come from reading the provider, search and grouping code, not from running them.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 720b0cfbfabf.

VERDICT: CHANGES

F1 [blocker] packages/client-runtime/src/pullRequestFeed.ts:232 On web and desktop, Involving me admits unrelated Bitbucket and Forgejo PRs because their unfiltered partition responses label them as authored or review-requested. Cursor Origin has the same problem for review requests. Fix: For fallback providers, verify the row’s author and review-request flag before admitting it. Carried over from F1.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:698 Load more raises the limit without sending a cursor, so GitHub’s batched search rereads the same first 100 rows. At the new cap, one old result from another repository can still cause an unread match to be hidden even after the broad list reaches it. Fix: Follow the involved listing’s own cursors, append its rows, and track any remaining coverage gaps per repository. Carried over from F2.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:245 Load more never advances the authored or reviewing lists. On GitLab, for example, the 31st authored PR can sit beyond many pages of unrelated activity, forcing the user to page through the broad list to reach it. Web also keeps these partitions at their initial limit. Fix: Advance the authored and reviewing partitions that supply Involving me rows when those partitions have more results.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 If an Authored checkout owns the main read for a shared repository, the Involving me checkout can be omitted from the lookup. The selected copy then fails open and shows unrelated PRs. Fix: Read the copies whose rules count and evaluate readiness by repository. Carried over from F4.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 More than 100 projects on the rule in one environment makes the server reject all three secondary reads. Web can hit the same limit when its main target covers the whole environment. Fix: Split project IDs at the RPC limit and track readiness per batch. Carried over from F5.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The three secondary hooks can fetch diff stats for rows the combined feed discards, and pull to refresh can invalidate the same environment four times. Fix: Fetch stats through the visible list and invalidate each environment once per refresh. Carried over from F6 and F7.

F7 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:831 When a grown read spans two environments, the first reply replaces the entire held answer. Rows found only by the slower environment’s involved listing disappear until it replies. Fix: Retain the previous answer for each environment until its replacement arrives.

F8 [nit] packages/contracts/src/settings.ts:1570 Older installed clients reject the settings snapshot after an upgraded client saves involved. Fix: Document the client upgrade requirement. Carried over from F8.

F9 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 The empty involvement fallback discards the search’s truncation flag. If search stops early without usable rows and coverage is cached, the project looks fully read and its rows can be hidden. Fix: Preserve truncation and fail the project read when an incomplete answer has no usable cursor. Carried over from F9.

Focused Node checks reproduced F1 and F2. git diff --check passed. Repository tests and typechecking could not run because vp and workspace dependencies were unavailable.

…r was complete

The involved read grew only on a truncated answer, so it could stall
while the main list grew, and a failed read kept the button growing it.
It now grows with each Load more unless its latest answer is complete,
Load more stays offered while only the involved read can grow, and a
complete main list is not read again for it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of fdda035a27e6.

VERDICT: CHANGES

F1 [should] packages/client-runtime/src/pullRequestFeed.ts:232 Carried over from gpt F1 (#33 (comment)) and Claude F1 (#33 (comment)). There is no author reply, and the code is the same at this head. On web, Involving me still shows every PR of a Bitbucket or Forgejo project. Neither provider reads input.involvement; grep finds it only in their tests. PullRequestService only forwards the value. Cursor Origin narrows only authored (OriginPullRequestProvider.ts:633). As a result, these hosts' authored and reviewing partitions hold the whole repository. partitionPullRequestsWithPriority files each row by URL alone, and pullRequestFeedVisibility admits it on that group without checking groupOf. The PR says these hosts are "Same as GitLab", which depends on those partitions being narrowed, so the claim does not hold. Fix: for the involved rule (or for any partition from a host that ignores involvement), accept a row only when its own pullRequestFeedGroup matches the partition.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:696 Carried over from gpt F2 and Claude F2 (same comments), with new evidence against the new lean: comment at :839. The read is said to stop at 500 rows per environment, but for GitHub it stops at 100 rows per search chunk:

  • Every GitHub project goes through listChangeRequestsAcross (PullRequestService.ts:1654-1668), even a single-repository chunk.
  • searchPullRequests reads Math.min(input.limit + 1, PULL_REQUEST_SEARCH_MAX_ROWS) rows, with the cap at 100 (GitHubPullRequestApi.ts:2049).
  • Web never sends the involved read's nextCursors.

So each Load more raises the limit from 99 to 198, then 297, 396, 495 and 500, and every time it fetches the same first 100 rows with truncated: true. Load more stays offered through involvedTruncated for those five useless reads. Once the main list pages past the 100th involved row, PRs where the viewer only commented, was mentioned or is assigned stay hidden. This is the main GitHub case, and the PR says older involved rows appear as the user loads more. Fix: page the involved read with its own nextCursors and append its rows, as mobile does. At minimum, correct the stated limit and stop offering growth that cannot return more rows.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:648 loadingMore checks only listQuery.isPending. When only the involved read grows (the main list is complete), the web footer shows no spinner. The footer also disappears while the read travels, because involvedQuery.data is null so involvedTruncated is false, then comes back when it answers. The description says the spinner covers either list, but only mobile does this. Fix: include the involved read's pending state in loadingMore and the footer condition.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 Carried over from Claude F4 and gpt F4 (same comments). The code is unchanged. Take a repository that is a project on two environments. If the main list reads copy A and copy B is on the rule, involvedQueries leaves B out, so the rule fails open for that repository. Fix: build the targets from every copy on the rule and check readiness by repository.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:191 Carried over from Claude F5 and gpt F5 (same comments). projectIds is not split by PULL_REQUEST_LIST_PROJECT_LIMIT (100). With more than 100 projects on the rule in one environment, all three extra reads fail schema validation. Fix: batch by the limit.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:217 Carried over from Claude F6 and gpt F6 (same comments). Each of the three extra usePullRequestList hooks fetches diff stats for rows that deduplication or the feed then drops. One pull to refresh also invalidates the same environment four times. Fix: read these lists without stats and invalidate each environment once.

F7 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 Carried over from Claude F7 and gpt F9 (same comments). The empty involved fallback always returns truncated: false. That drops the truncation flag of a search that stopped early, so in a covered repository an incomplete empty answer counts as complete. Fix: keep the batch's truncated flag, and fail the read when it is set and there is no cursor.

F8 [nit] packages/contracts/src/settings.ts:1570 Carried over from Claude F8 and gpt F8 (same comments). The forgiving decode helps only clients built from this version. An older installed client still fails to decode the settings snapshot once involved is stored. Fix: note the client upgrade requirement.

These are fixed at this head:

  • Removed floor: the cross-environment involvedFloor from the previous Claude F3 is gone.
  • Mobile retry: a failed involved page now reruns only through Try again, and the footer shows its error and spinner.
  • Main list: a complete main list is no longer read again when only the involved read can grow.

I confirmed F1 and F2 by reading the provider, service and search code. I did not run typecheck or tests, because the checkout has no dependencies installed.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of fdda035a27e6.

VERDICT: CHANGES

F1 [blocker] packages/client-runtime/src/pullRequestFeed.ts:232 Involving me shows unrelated PRs on web and desktop for Bitbucket and Forgejo because their unfiltered partition responses label those rows as authored or review-requested. Cursor Origin has the same issue with its reviewing partition. Fix: Check the row’s author and review-request flag before admitting it through a partition from these providers. Carried over from F1.

F2 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:700 Load more raises the limit without sending the involved listing’s cursor. GitHub’s batched search then returns the same first 100 matches, leaving older PRs where the viewer only commented, was mentioned, or was assigned hidden. Fix: Follow the involved listing’s nextCursors and append the results. Carried over from F2.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:245 Load more leaves the authored and reviewing partitions on page one on both clients. On GitLab, reaching the 31st authored PR on mobile can require paging through many pages of unrelated activity. Fix: Advance the authored and reviewing partitions that supply Involving me rows, and include them in the footer’s pagination state. Carried over from F3.

F4 [nit] apps/web/src/components/pullRequest/pullRequestListGroups.logic.ts:118 If the authored request fails before its first answer, web also drops the successful reviewing results. A review-requested PR beyond the broad page then stays missing. Fix: Merge each available partition independently.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:452 If an Authored checkout owns the main read for a shared repository, the Involving me checkout can be left out of the lookup. Every unrelated row then passes as unknown. Fix: Query the checkouts whose rules apply and match readiness by repository. Carried over from F4.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 An environment with more than 100 projects on the rule exceeds the RPC’s projectIds limit for all three secondary reads. Web can hit the same limit when the broad read covers a whole environment. Fix: Batch at the contract limit and track readiness per batch. Carried over from F5.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The three secondary hooks can fetch diff stats for rows the feed discards, and pull to refresh can invalidate the same environment four times. Fix: Fetch stats through the combined list’s visible-row handler and invalidate each environment once per refresh. Carried over from F6.

F8 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:834 The first environment to answer a larger read replaces the entire held answer. Rows supplied by a slower environment’s involved listing disappear until it replies. Fix: Retain each environment’s previous answer until its replacement arrives. Carried over from F7.

F9 [nit] packages/contracts/src/settings.ts:1570 Older installed clients can reject the whole settings snapshot after an upgraded client saves involved. Fix: Document the client upgrade requirement. Carried over from F8.

F10 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1968 The empty involved fallback discards the search’s truncation flag. If search stops early without usable rows and coverage is cached, the project appears fully read and its rows can be hidden. Fix: Preserve truncation and fail the project read when an incomplete answer has no usable cursor. Carried over from F9.

F11 [nit] packages/client-runtime/src/pullRequestFeed.ts:80 The involvement === "involved" exception and its test cover a list tab that neither client exposes. Fix: Drop the exception and its test row while keeping involved as an RPC input for the feed rule.

Focused Node checks reproduced F1, F2, and F4. git diff --check passed. Repository tests and typechecking could not run because vp and workspace dependencies were unavailable.

Web's involved read grew only its limit, so GitHub's 100-row search
chunk returned the same rows on every Load more. A small paged-list hook
now follows each listing's own `nextCursors` and adds its pages up; for
projects on the rule it pages the involved, authored and review-requested
reads together. One `pullRequestLoadMore` helper turns the main and
involved read states into what Load more grows and what the footer
shows, so a page in flight is never restarted and a failed page retries.

Hosts that ignore `authored` and `reviewing` (Bitbucket, Forgejo, Cursor
Origin) sent every row into those partitions; the server now keeps only
rows whose author or review requests involve the viewer outside GitHub.
GitHub's empty `involved` fallback fails when its batch was truncated.
Mobile advances the authored and review lists with Load more too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of afa8c65e6efe.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/PullRequestsPage.tsx:742 The main list stops paging whenever any environment's read is failing. If listQuery.error is set, the main read counts as "failed", and Load more then only calls listQuery.refresh() (:755) and never moves to nextCursors. A persistent failure keeps the error set: for example, the server rejects a whole environment when none of its hosts has a viewer (missing or unauthenticated gh, PullRequestService.ts:1411-1431). In that case Load more just re-reads the current page on every click, and the healthy environments never get past their first page. Before this PR, Load more kept paging the environments that answered. This affects every feed, not only Involving me. Fix: refresh only the failed targets, and still move the environments that answered forward with their cursors. Count the read as "failed" only when no environment can continue.

F2 [should] apps/web/src/components/pullRequest/usePagedHostList.ts:96 The new paged listings have the same problem. state becomes "failed" when any target environment failed, and advance() (:102) then only refreshes and never applies held.nextCursors. A persistent failure in one environment therefore keeps the involved listing on page one for every other environment. For a project on the rule, that listing is the only source of Others rows, so older PRs where the viewer commented, was mentioned or was assigned never appear. The footer also keeps offering a Load more that does nothing new. Fix: on advance, page the environments that have cursors and retry only the environments that failed.

F3 [nit] apps/web/src/components/pullRequest/pullRequestList.logic.ts:467 Rows that only the paged involvedAuthored and involvedReviewing listings bring are added to the Others group, because the Authored and Review requested groups are built only from the first-page partitions. groupOf lets them through the rule, but the viewer's own PRs and PRs waiting on their review appear under Others. Fix: add the authored and reviewing listings' rows to the matching groups, or group the extra rows by pullRequestFeedGroup.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:871 One failed Load more page of the involved listing adds its environment to involvedListing.failed. Every project on the rule in that environment then fails open and shows all its rows, even though page one already answered. Fix: keep projects ready based on the pages that already answered, and use failed only before the first answer.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:458 Carried over from Claude F4 (#33 (comment)); the code is unchanged. Take a repository that is a project in two environments. If the main read covers copy A and copy B is the one on the rule, involvedQueries leaves B out, and the rule fails open for that repository. Fix: build the targets from every copy on the rule and check readiness by repository.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 Carried over from Claude F5 (same comment). projectIds is not split at the RPC's 100-project limit. With more than 100 projects on the rule in one environment, all three secondary reads are rejected. Web has the same issue when its main query covers a whole environment. Fix: split the project IDs into batches at the limit.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:217 Carried over from Claude F6 (same comment). The three secondary usePullRequestList hooks fetch diff stats for rows that deduplication or the rule then drop. One pull to refresh also invalidates the same environment four times. Fix: read these lists without stats and invalidate each environment once.

F8 [nit] packages/contracts/src/settings.ts:1570 Carried over from Claude F8 (same comment). The forgiving decode helps only clients built from this version. An older installed client still fails to decode the settings snapshot once involved is stored. Fix: document that clients must be upgraded.

F9 [nit] packages/client-runtime/src/pullRequestFeed.ts:80 Carried over from gpt F11 (#33 (comment)). The involvement === "involved" exception and its test cover a list tab that neither client offers: web's INVOLVEMENT_TABS does not include it, and mobile's preferences restrict the choice to all, reviewing and authored. Fix: drop the exception and its test row.

These earlier findings are fixed at this head:

  • Bitbucket, Forgejo and Cursor Origin partitions (Claude F1 and gpt F1 at fdda035): the server's new keepsInvolvement check keeps an authored or reviewing row only when its author or review requests involve the viewer. It uses the same comparisons as toEntry.
  • Load more not paging the involved read on web (Claude F2 and gpt F2): usePagedHostList now follows each listing's own nextCursors.
  • No spinner while only the involved read loads (Claude F3): the footer now spins through pullRequestLoadMore.
  • GitHub's empty involved answer dropping truncation (Claude F7 and gpt F10): a truncated empty answer now fails the read.
  • Authored and review lists not paging (gpt F3): mobile advances them with Load more, and web pages them for projects on the rule.
  • First environment's answer replacing the held answer (gpt F8): a first page is now taken only after every environment has answered.

I found F1 and F2 by reading the code (createMergedEnvironmentQuery, mergePullRequestLists and the server's no-viewer failure path); I did not run anything to reproduce them. I did not run typecheck or tests, because the checkout has no dependencies installed.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of afa8c65e6efe.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/usePagedHostList.ts:84 A live refresh or completed turn can erase older GitHub matches the user already loaded. The refresh raises the initial limit, but GitHub still returns at most 100 rows per search batch. The hook replaces all accumulated pages with that first batch. Fix: Keep the loaded pages visible while the refresh follows cursors through the previously loaded range.

F2 [should] apps/web/src/components/pullRequest/usePagedHostList.ts:66 Paging a healthy repository clears earlier errors for every repository in the same environment. An unindexed GitHub repository then starts hiding its PRs even though its involvement lookup never succeeded. Fix: Replace errors only for repositories the new page actually read.

F3 [should] apps/web/src/components/pullRequest/usePagedHostList.ts:96 A repository-level error on a continuation is returned inside a successful RPC response, so the pager can mark the listing complete and remove Load more after a transient failure. Fix: Keep the failed repository’s request cursor and make that page retryable.

F4 [should] apps/server/src/pullRequest/PullRequestService.ts:1453 The new login check drops GitCafe PRs that the provider correctly matched by actor ID when the viewer’s linked GitHub login differs from their GitCafe handle. These PRs disappear from Involving me and the existing web Authored by me feed. Fix: Preserve provider-verified authorship and use login matching for providers that return unfiltered rows.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:458 When an Authored checkout owns the main read for a shared repository, the checkout set to Involving me can be left out of the lookup. Unrelated PRs then pass as unknown. Fix: Query the checkouts whose rules apply and match readiness by repository. Carried over from F5.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 More than 100 projects on the rule in one environment can make all three lookups exceed the RPC’s project ID limit. Web has the same issue when its main query covers the whole environment. Fix: Split requests at PULL_REQUEST_LIST_PROJECT_LIMIT. Carried over from F6.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The three secondary hooks can fetch diff stats for rows the feed discards, and pull-to-refresh can invalidate the same environment four times. Fix: Fetch stats through the visible list and invalidate each environment once per refresh. Carried over from F7.

F8 [nit] packages/contracts/src/settings.ts:1570 Older installed clients can reject the whole settings snapshot after an upgraded client saves involved. Fix: Document the client upgrade requirement. Carried over from F9.

F9 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1963 The new guard catches an empty truncated initial repository read, but continuation reads bypass it. An empty truncated continuation has no row from which to build a cursor, so web can silently end the listing. Fix: Treat empty truncated involved responses as retryable failures on both the batched and single-repository paths. Carried over from F10.

F10 [nit] packages/client-runtime/src/pullRequestFeed.ts:80 The involvement === "involved" exception and its test cover a list tab neither client exposes. Fix: Remove the exception and its test case while keeping involved as an RPC input. Carried over from F11.

Focused Node checks reproduced F1–F4, and git diff --check passed. Repository tests and typechecking were unavailable because the checkout lacked vp and workspace dependencies.

…n its own

One failed environment marked the whole main list or paged listing as
failed, so Load more only refreshed and healthy environments never left
page one. The main list now fails only when no environment can continue,
and the paged listings advance environments with cursors while retrying
only the failed ones. Paged pages merge through one tested helper: a
first page replaces rows in its span and keeps older loaded rows with the
deeper cursor, a continuation clears only the errors of projects it read,
and a continuation that failed for a passing reason offers its cursor
again from the server.

The login check now runs only for providers that declare they ignore an
involvement (Bitbucket, Forgejo, Cursor Origin for reviewing, GitCafe
for unnarrowed reviewing), so rows a provider matched another way stay.
Rows only the paged authored and review listings bring join their own
group, and the unused `involved` tab case is gone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of a09e79c9aa88.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/pullRequestLoadMore.logic.ts:97 When an environment holds several projects with their own cursors, a refresh drops pages the user already loaded, and Load more never brings them back. mergePagedListing takes the oldest row of the whole environment's refreshed first page as the boundary (oldest is keyed by environmentId). It then keeps only held rows older than that boundary. Meanwhile deeper (:124) keeps each repository's deeper cursor. Take a busy project X and a quiet project Y with old rows. X's loaded page-2 rows are newer than Y's oldest row, so a live refresh or a finished turn removes them. X's cursor still points past those rows, so later pages skip them. I reproduced this with Node on a copy of the helper. After one Load more, the rows were [x1,x2,y1,x3,x4] with cursor x-c2. After refreshing the first page, they were [x1,x2,y1] and the cursor was still x-c2. This breaks the stated rule that "a refresh never shrinks what you loaded". It hits the authored and review-requested listings on hosts that read each project on its own (GitLab, Azure DevOps and others). It also hits GitHub projects that end up in different search groups or chunks. The only test uses one project per environment. Fix: work out the boundary, and whether a page has more, for each project's cursor key (projectId/repository) rather than per environment. Then keep a held row whenever its own project's deeper cursor is kept.

F2 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:861 Carried over from Claude F4 (#33 (comment)). If one Load more page of the involved listing fails at the RPC level, involvedListing.failed includes that environment. Every project on the rule in that environment then fails open, even though its first page already answered. Fix: base readiness on the pages that answered, and use failed only before the first answer.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:454 Carried over from Claude F5 (#33 (comment)) and gpt F5 (#33 (comment)). The code is unchanged. Take a repository that is a project in two environments. If the main read covers copy A and copy B is the one on the rule, involvedQueries leaves B out, so that repository fails open. Fix: build the targets from every copy on the rule and check readiness by repository.

F4 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:195 Carried over from Claude F6 and gpt F6 (same comments). projectIds is not split at the RPC's 100-project limit, so with more than 100 projects on the rule in one environment, all three extra reads are rejected. Fix: split the IDs into batches at PULL_REQUEST_LIST_PROJECT_LIMIT.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 Carried over from Claude F7 and gpt F7 (same comments). The three extra usePullRequestList hooks fetch diff stats for rows that the dedupe step or the rule then drops. One pull to refresh also invalidates the same environment four times. Fix: read these lists without stats, and invalidate each environment once.

F6 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1963 Carried over from gpt F9 (#33 (comment)). The empty-and-truncated guard skips continuations (input.cursor !== undefined). An empty truncated involved continuation therefore gives nextListCursor no rows to build a cursor from, and the listing ends silently. Fix: treat an empty truncated involved page as a retryable failure on continuations too.

F7 [nit] packages/contracts/src/settings.ts:1570 Carried over from Claude F8 and gpt F8 (same comments). The forgiving decode only helps clients built from this version. An older installed client still rejects the settings snapshot once involved is stored. Fix: document that clients need to be upgraded.

These earlier findings are fixed at this head:

  • Main list stops paging after one failed environment (Claude F1 at afa8c65): mainReadState now reports failed only when no environment can continue.
  • Paged listings stop after one failed environment (Claude F2): nextPagedSend advances environments that have cursors and retries only the failed ones.
  • Extra authored and review rows filed under Others (Claude F3): they now join the group their own author and review requests give them.
  • Refresh erasing loaded rows (gpt F1): fixed for a single project per environment. F1 above covers the case with several projects.
  • Errors cleared for every repository in an environment (gpt F2): a page now clears only the errors of projects it brought rows for.
  • Failed continuation ending the listing (gpt F3): retriesFrom offers the cursor again.
  • GitCafe rows dropped by the login check (gpt F4): the check now runs only where a provider declares ignoresInvolvement, and GitCafe's narrowed review answer adds the viewer.
  • Unused involved tab case (gpt F10): removed.

I ran F1's reproduction with Node 22 type stripping on a copy of pullRequestLoadMore.logic.ts. I did not run the repository's tests or typecheck, because the checkout has no dependencies installed.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of a09e79c9aa88.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/pullRequestLoadMore.logic.ts:97 A refresh can discard loaded GitHub matches from a busy repository when a quiet repository in the same environment returns an older first-page row. The merge drops the busy repository’s rows using the quiet repository’s timestamp, then keeps the busy repository’s deeper cursor, so Load more skips the dropped rows. Fix: Reconcile refreshed rows and cursors using each repository’s actual first-page coverage. Carried over from F1, with the gap reproduced in the new merge helper.

F2 [should] apps/web/src/components/pullRequest/usePagedHostList.ts:83 A transient repository error on a continuation comes back in a successful RPC response with the same cursor. When that repository is the only target left, Load more selects the same cached query and never refreshes it because query.failed is empty. Fix: Explicitly refresh continuations that return a retry cursor with a project error, and avoid caching those errors as successful pages. Carried over from F3, with the failure now in the client retry path.

F3 [should] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:312 GitCafe can correctly return a synced PR in the authored listing by actor ID, but mobile drops it when the linked GitHub login differs from the GitCafe handle. Web does the same for rows found only on later authored pages. Fix: Preserve provider-confirmed membership from the authored and reviewing listings through deduplication and feed filtering. The server-side fix from F4 preserves the row, but the clients still discard it.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:454 When an Authored checkout owns the main read for a shared repository, a second checkout set to Involving me can be left out of the lookup. The rule then admits unrelated rows because membership is unknown. Fix: Query the checkouts whose rules count and resolve membership by repository. Carried over from F5.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:203 More than 100 projects on the rule in one environment exceeds the RPC’s project ID limit for all three secondary reads. Web can hit the same limit when the broad read covers a whole environment. Fix: Batch requests at PULL_REQUEST_LIST_PROJECT_LIMIT. Carried over from F6.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:206 The three secondary hooks can fetch diff stats for rows the feed discards, and pull-to-refresh can invalidate the same environment four times. Fix: Fetch stats for visible rows and invalidate each environment once per refresh. Carried over from F7.

F7 [nit] packages/contracts/src/settings.ts:1570 An older installed client can reject the entire settings snapshot after an upgraded client saves involved. Fix: Document the client upgrade requirement. Carried over from F8.

F8 [nit] packages/source-control-github/src/server/GitHubPullRequestApi.ts:1963 An empty, truncated GitHub continuation bypasses the new guard and can end pagination because the server has no row from which to build a cursor. Fix: Treat empty, truncated involved continuations as retryable failures on both the batched and single-repository paths. Carried over from F9.

Focused Node checks reproduced the state behind F1–F3. git diff --check passed. Repository tests and typechecking were unavailable because vp and workspace dependencies were absent.

The Load more paging of the involved, authored and review-requested
listings is removed on web and mobile, with its hook, its merge helpers
and the server's retry cursor. Involving me reads one page of each (99
rows, on GitHub per search chunk of 100 repositories), a limit the docs
and a `lean:` comment now state. The source model, per-project fail-open,
search bypass and the `ignoresInvolvement` check stay, and the main
list's Load more keeps the per-environment failure handling.

A row an authored or review-requested listing returned now keeps that
group on mobile, so a GitCafe pull request matched by actor is not
dropped when its GitHub login differs from the GitCafe handle.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

claude-opus-5-5 responding on behalf of Ilkka

Automatic review of 8578a7a6ca67.

VERDICT: APPROVE

F1 [nit] apps/web/src/components/pullRequest/pullRequestLoadMore.logic.ts:807 mainReadState and pullRequestLoadMore came in to support listing paging, which a later commit in this PR removed, and they change the main list's Load more in every feed. When any environment's read failed (listQuery.error) and no environment has a cursor, the state is "failed" whenever pageSize < MAX_PAGE_SIZE, so Load more only calls listQuery.refresh() (PullRequestsPage.tsx:708). It no longer grows the page. Before this PR, Load more grew the page there, which also retried the failed environment. Now a persistent failure in one environment stops every healthy environment that has no cursor from ever growing past its current page. Involving me does not need this change. Fix: drop pullRequestLoadMore.logic.ts and its test, and restore the base canLoadMore/loadMore/footer logic. If a retry is wanted, keep growing the page whenever canGrow.

F2 [nit] packages/source-control-gitcafe/src/server/GitCafePullRequestProvider.ts:799 GitCafe declares only reviewing in ignoresInvolvement. But when /auth/principal cannot be read for a non-auth reason, authored is unnarrowed too: no authors parameter is set (:831-838), and filters.author is absent. The service then keeps every row of the authored listing. Mobile's new pullRequestListedGroup files each of those rows as authored (PullRequestsScreen.tsx:325), and web does the same through its authored partition. So Involving me shows up to 99 of the project's pull requests, all labelled as authored by the viewer. Declaring authored would wrongly drop rows GitCafe matched by actor when the principal is known. Fix: when the principal is unavailable, narrow authored by login inside the GitCafe provider, as the service already does for reviewing.

F3 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:453 Carried over from Claude F3 (#33 (comment)); the logic is unchanged. Take a repository that is a project in two environments. If the main read covers copy A and copy B is the one on the rule, involvedQueries leaves B out. ready is then false for A's rows, and the repository falls back to All. Mobile behaves the same way, because ready also keys on the row's own environment and project. Fix: build the targets from every copy on the rule, and check readiness by repository.

F4 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:213 Carried over from Claude F4 (same comment). projectIds is not split at PULL_REQUEST_LIST_PROJECT_LIMIT. With more than 100 projects on the rule in one environment, all three extra reads are rejected. Web hits the same limit when an environment query has no projectIds. Fix: split the IDs into batches at the limit.

F5 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:220 Carried over from Claude F5 (same comment). The three extra usePullRequestList hooks each fetch diff stats for their first visible rows, including rows that dedupe or the rule then drop. One pull to refresh (:461-464) also invalidates the same environment four times. Fix: read these lists without stats, and invalidate each environment once.

F6 [nit] packages/contracts/src/settings.ts:1570 Carried over from Claude F7 (same comment). The forgiving decode only helps clients built from this version. An older installed client still rejects the whole settings snapshot once involved is stored. Fix: document that clients need to be upgraded.

Earlier findings that no longer apply at this head:

  • Paging merge, retry and continuation (Claude F1, F2 and F6, and gpt F1, F2 and F8 at a09e79c): Involving me now reads one page of each listing and never sends cursors, so this code is gone.
  • GitCafe rows dropped on mobile (gpt F3 at a09e79c): fixed by pullRequestListedGroup, which keeps a row's group when the authored or review-requested listing returned it.

I read the code paths without running them. Repository tests and typecheck were not run, because the checkout has no installed dependencies. I also could not confirm that PullRequestInvolvement.pick(...) (mobile-preferences.ts:34) exists in the installed Effect version.

@pnupu

pnupu commented Nov 19, 2026

Copy link
Copy Markdown
Owner Author

Note

gpt-6.2-sol responding on behalf of Ilkka

Automatic review of 8578a7a6ca67.

VERDICT: CHANGES

F1 [should] apps/web/src/components/pullRequest/pullRequestListGroups.logic.ts:118 Web discards a successful review-requested listing while the authored listing is pending or has failed, and vice versa. A PR returned only by the successful listing disappears from Involving me. Fix: Merge each available listing independently.

F2 [should] packages/client-runtime/src/pullRequestFeed.ts:95 Both clients admit authored and review-requested rows from the broad list even when those rows are absent from all three listings. Load more can therefore reveal older PRs that the documented one-page limit says stay hidden. Fix: For ready projects, require URL membership in the union of the three listings.

F3 [nit] packages/source-control-gitcafe/src/server/GitCafePullRequestProvider.ts:825 A transient principal lookup failure can return an unfiltered authored page. The service passes that page through, and both clients treat unrelated PRs from it as authored. Fix: Filter by login when the actor ID is unavailable and preserve actor matching when it is available.

F4 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:708 Load more still stalls when one environment fails and a healthy environment reports truncation without a cursor, as GitHub’s repository fallback can. Each click retries the same limit. Fix: Grow the healthy environment’s request and leave the failure to the Retry action. Carried over from F1.

F5 [nit] apps/web/src/components/pullRequest/PullRequestsPage.tsx:453 When the broad list assigns a shared repository to a checkout with an Authored rule, it can omit the checkout with an Involving me rule from the lookup. The shared repository then shows unrelated PRs. Fix: Query the checkouts whose rules apply and match readiness by repository. Carried over from F4.

F6 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:210 More than 100 projects on the rule causes the RPC to reject all three mobile lookups. Web can hit the same limit when the broad query covers a whole environment. Fix: Batch project IDs at the RPC limit and track readiness per batch. Carried over from F5.

F7 [nit] apps/mobile/src/features/pull-requests/PullRequestsScreen.tsx:220 The secondary hooks can fetch diff stats for discarded rows, and pull-to-refresh can invalidate the same environment four times. Fix: Fetch stats for visible rows and invalidate each environment once per refresh. Carried over from F6.

F8 [nit] packages/contracts/src/settings.ts:1570 An older installed client can reject the entire settings snapshot once an upgraded client saves involved. Fix: Document the client upgrade requirement. Carried over from F7.

Focused Node checks reproduced F1–F4. git diff --check passed. Repository tests and typechecking were unavailable because vp and workspace dependencies were missing.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant