Modular fork: move remaining fork code out of upstream files

Goal

Make the fork's customizations modular, so each daily upstream sync has as few conflicts as possible. Fork code lives in fork-owned files and connects to upstream files through small hook points. Upstream code changes only to fix a bug that blocks a fork feature.

The owner approved this as a follow-up to the upstream sync (#11) on 2026-10-12: "Yes, as a follow-up".

Already modular (2026-10-12)

Fork migrations have their own ledger (ForkMigrations.ts, table t3_fork_migrations).
Fork RPCs are a separate group (forkWsRpcs) served from ForkWsRpc.ts.
Issue recovery decisions live in IssueWorkRecovery.ts. Delegated request routing lives in DelegatedRequestRouting.ts, DelegatedRequestWake.ts and packages/shared/src/delegatedRequestVisibility.ts.

To do

Inventory. List every place where fork code is still woven into upstream files: Orchestrator.ts, EventSink.ts, ProjectionStore.ts, ws.ts, ChatView.tsx, Sidebar*.tsx, the provider adapters, contracts, and so on. Record the file, the fork feature and the line count. Use git diff upstream/main...main limited to files upstream owns.
Rank by how often each spot conflicted in the 2026-10-12 merge (132 files) and by size.
Move the largest pieces behind hook points, one feature per change, each with its own review. No behavior changes.
Check each step against the daily sync: conflicts should go down.

Mobile is out of scope.

Plan: maintain the fork as a patch stack (owner asked for a plan, 2026-10-13). For approval before any change.

Today: main is upstream plus about 123 fork commits, synced by merging upstream into it. Fork changes are spread through history and through upstream's big files, so each sync re-fights the same conflicts. git rerere is now on, which helps with repeats.

Target: the fork is upstream/main plus a short ordered list of feature commits, about 15–25, one per fork feature:

issue work and repository orchestration;
Issues page;
subagent question routing;
Don't ask mode;
model roles and usage ranking;
local servers;
sidebar background rows;
fork migrations ledger;
and the rest.

Each daily sync rebases that list onto the new upstream instead of merging.

Why it's cleaner:

A conflict names the one feature it touches.
When upstream ships something equivalent, that commit is dropped.
git range-diff old..new shows exactly how each feature changed between syncs, which suits review.

Steps:

Finish the modular pass in this issue first: fork code in fork-owned files with small hooks. Without that, every commit conflicts constantly.
Build the stack. Starting from upstream/main, re-create each fork feature as one commit (squash its history), in dependency order (migrations/contracts → server → client → UI).
Check the stack. The result must match current main exactly (empty git diff), and the targeted tests pass.
Switch: main becomes the stack, force-pushed once with owner approval. Archive the old history as archive/main-merge-history-<date>.
Daily sync becomes: fetch, git rebase upstream/main with rerere, fix any conflicting commit, range-diff review, then force-push main (owner-approved standing rule needed).

Costs:

Rewritten history on every sync. Worktrees branched from old main must rebase. Each sync needs a force-push, which needs a standing approval from the owner.

Decision points for the owner when step 1 is done:

approve steps 2–4;
standing approval for daily force-pushes of main.

Owner decision (2026-10-13): run the modular pass after the feature batches in #1's plan. New features are built modular from the start.

Inventory and plan (2026-10-14). Read-only research on main cc5a99dcd8 against upstream 570542662a, merge base a01c48f862.

392 upstream-owned files differ: +37,778 / −5,761 lines. Tests and fixtures are 21,251 of those lines.
Replaying the last 3 real upstream syncs gave 8, 129 and 4 conflicting paths, 133 distinct paths in all.
ws.ts looks large, but ignoring whitespace it's +72/−19: the RPC split is already modular. contracts/rpc.ts is already modular too.

Biggest woven spots (conflicts × size; estimated lines that would leave the upstream file):

Orchestrator.ts (+1,948/−299): issue admission, repository lifecycle, app-owned questions, completion delivery, automatic switching. Move to IssueWorkCommands, RepositoryWorkCommands, AppOwnedUserInputCommands, ForkDelegatedCompletion and AutomaticProviderSwitchPolicy. Estimated 1,050–1,400 lines.
ThreadRelationshipsControl.tsx: agent rows, history, columns. Move to useForkAgentRows and ForkAgentRows. 500–600.
ProjectionStore.ts: issue projection, shell enrichment, feature queries. Move to IssueWorkProjection, ForkShellQueries and ForkProjectionQueries. 500–650.
OrchestratorMcpService.ts: managed delegation, worktrees, task status, send policy. 330–430.
Next, in order: Sidebar.tsx, ClaudeAdapterV2.ts, CodexAdapterV2.ts, contracts/orchestrationV2.ts, ClaudeSkills.ts, ChatMarkdown.tsx, contracts/settings.ts, ProviderSessionManager.ts, then EffectWorker.ts and PortScanner.ts.

Test suites: fork tests sit inside upstream test files: thread launch about 4,900 lines, delegated completion about 1,200, Codex adapter about 1,100. They can move into fork-owned test files with no runtime risk.

Batch plan, one feature per batch, each reviewed, no behavior changes:

model roles and ranking
managed issue admission
repository ownership and handoff
issue upkeep effects
routed questions
Don't ask mode
provider switch
background work
provider parity
sidebar sections
sidebar columns
plugin skills
local servers
inline shell runs
Issues page
math
inline copy
host media
idle session release

Keep inline as upstream-style fixes: default-ref fallback without origin/HEAD, native protocol corrections, clipboard and line-break fixes, worker-lane isolation.

Not part of this plan: the patch-stack idea in the earlier comment. It rewrites history and needs separate approval.

Decision (2026-10-14, recommended option; the owner was away in Don't ask mode): start small and measure.

Batch A: move the fork's test cases out of upstream-owned test files into fork-owned *.fork.test.ts files, with no product code changes. In progress on branch t3/24-fork-test-suites.
Batch B: extract issue work from Orchestrator.ts.
Then: compare the conflict count at the next daily upstream sync before doing more batches.

test(server): move fork tests out of ThreadLaunchService.test.ts

test(server): move fork tests out of DelegatedCompletionDelivery.test.ts

test(server): move fork tests out of CodexAdapterV2.test.ts

test(server): move fork tests out of ClaudeAdapterV2.test.ts

test(server): move fork tests out of PortScanner.test.ts

test(server): move fork tests out of EffectWorker.test.ts

Batch A merged: fb7da58, 6 commits, tests only. Fork test cases moved out of six upstream-owned test files into *.fork.test.ts files, with shared fixtures in .testHarness.ts.

File	Fork diff vs upstream, before → after
ThreadLaunchService.test.ts	+5,080/−387 → +14/−465
DelegatedCompletionDelivery.test.ts	+1,252/−67 → +13/−339
CodexAdapterV2.test.ts	+1,177/−48 → +77/−751
ClaudeAdapterV2.test.ts	+854/−73 → +80/−488
PortScanner.test.ts	+458/−2 → +18/−19
EffectWorker.test.ts	+367/−34 → +61/−135

Total: 9,799 → 2,460 changed lines, a 75% cut. The 668 tests are the same, with the same names, and all pass.

Trade-off: shared fixtures had to leave the upstream files, because importing from a .test.ts file would re-register its tests. The remaining diff is mostly those deletions. If upstream edits a moved helper, the conflict now lands in the harness file.

Measurement: a trial merge of today's upstream (614b432) gave the same result before and after: 18 conflicted paths and 33 conflict blocks. These upstream commits touched neither the moved tests nor the moved helpers. The real test is the next daily sync.

Next: batch B, extracting issue work from Orchestrator.ts.

Sync measurement (2026-10-29). main at ffa7e30 against nightly tag v0.1.5-nightly.20261029.3877 (6d450a4): 52 new upstream commits, giving 12 conflicted files and 15 <<<<<<< blocks. Measured with git merge-tree --write-tree --name-only --no-messages.

Sync	Upstream commits	Conflicted files	Blocks
2026-10-27	132	33	68
2026-10-28	159	80	209
2026-10-29 (nightly tag)	52	12	15

Conflicted files: apps/desktop/src/updates/DesktopUpdates.ts, apps/server/src/mcp/toolkits/project/tools.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/scheduledTasks/ScheduledTaskService.ts, apps/web/src/components/ChatMarkdown.tsx, apps/web/src/components/ChatView.tsx, apps/web/src/components/chat/MessagesTimeline.logic.ts, apps/web/src/components/chat/ThreadDetailsPanel.tsx, apps/web/src/components/settings/SettingsPanels.tsx, packages/contracts/src/settings.ts, packages/provider-core/src/server/managedProvider.ts and its test.

T3 Code closeout: done

Issue: #24 Modular fork: move remaining fork code out of upstream files
Thread: thread:delegated-task:command%3Amcp%3Aa45c4156-6756-45dc-bd27-1ea406cde16f%3Adelegate-task%3Aissue-24-orchestrate-20261029
Model: gpt-6.1-sol
Branch: issue-24-modular-fork-move-remaining-fork-code
Completed: 2026-10-30T05:05:50.627Z

Understood. I’ll continue local implementation and Checker review, with delivery held while authentication is unavailable. No pushes, PRs, main fast-forwards, or issue writes.

After pass 2’s correction and review pass, I’ll check gh auth status once. If authentication is still invalid, I’ll report “ready to deliver” with the branch and head SHA. If valid, I’ll resume normal delivery.

The formatting correction is still assigned to the same implementor and will resume this thread automatically. I’ll reference #256 for the existing unused-export hook failure; --no-verify is authorized when that is the only hook failure.

T3 Code closeout: waiting_on_sub_issues

Issue: #24 Modular fork: move remaining fork code out of upstream files
Thread: thread:delegated-task:command%3Amcp%3Aa45c4156-6756-45dc-bd27-1ea406cde16f%3Adelegate-task%3Aissue-24-orchestrate-20261029
Model: gpt-6.1-sol
Branch: issue-24-modular-fork-move-remaining-fork-code
Completed: 2026-10-30T05:36:50.602Z

Pass 2 now includes current origin/main and #260 at head 7d9c2365a0b19075d5279091742c0504117e3bab. Authentication is valid.

The 68 launch/handoff tests, formatting, lint, generated GraphQL check, and affected typechecks pass. The remaining changed-file tests are running; the unused-export scan still reports the 74 findings tracked by #256.

The Checker is reviewing the formatting and integration delta:

t3-thread://v2/df61a5ad-76a1-4901-b919-31ae81c8e67d/thread:delegated-task:command%3Amcp%3Ad3506335-01b7-4a2c-9b60-c53578fcdb05%3Adelegate-task%3Aissue24-pass2-251-check-987ed4c0

I’m waiting on those checks and review, which resume this thread automatically. The refreshed comparison against nightly v0.1.6-nightly.20261030.3904 is 17 files / 23 blocks before and after. Publication remains pending; passes 3–5 are unstarted.

Sync measurement (2026-10-30). main at b539f4b against nightly tag v0.1.6-nightly.20261030.3917 (97e35bcf23): 126 new upstream commits, giving 26 conflicted files and 37 <<<<<<< blocks. Measured with git merge-tree --write-tree --name-only --no-messages.

Sync	Upstream commits	Conflicted files	Blocks
2026-10-27	132	33	68
2026-10-28	159	80	209
2026-10-29 (nightly tag)	52	12	15
2026-10-30 (nightly tag)	126	26	37

Conflicted files:

Server orchestration: Orchestrator.ts, ProjectionStore.ts, ProviderSessionManager.ts, ShellStream.ts, ThreadLaunchService.ts, ThreadManagementService.ts.
Other server files: RpcAuthorization.ts and its test, Sqlite.ts, ClaudeSkills.ts, ProviderRegistry.ts, agentContextBudget.json, serverRuntimeStartup.ts, serverSettings.ts, storageCleanup.ts, terminal/Manager.ts, GitVcsDriverCore.ts, ws.ts.
Web: MessagesTimeline.logic.test.ts, ThreadActivityBanners.tsx, ThreadLineageWorkflowRow.tsx, ThreadRelationshipsControl.tsx, threadActionMenu.logic.ts, useThreadActionMenu.ts.
Packages: clientRpcPermissions.ts, nodeSqliteClient.ts.

Several conflicts are in files the fork changed in the last day: Sqlite.ts, nodeSqliteClient.ts and agentContextBudget.json (#240, #260), and storageCleanup.ts (#238). Modular pass 2, repository orchestration, is still in delivery.

Sync 2026-10-31 (#265, v0.1.7-nightly.20261031.3991, 401 upstream commits including the v0.1.6 release): 61 conflicted files, 99 conflict blocks against main da7da71. Up from 17/23 on 2026-10-30, mostly from the size of the upstream batch (side chats, project pinning, the model/effort picker, sidebar and right-panel changes).

Sync 2026-11-01 (v0.1.7-nightly.20261101.4035, 175 upstream commits): 22 conflicted files, 27 conflict blocks against main ee4c4c8. Down from 61/99 on 2026-10-31 (401 commits) and close to the 17/23 of 2026-10-30.

Sync 2026-11-02 (v0.1.7-nightly.20261102.4080, 176 upstream commits): 31 conflicted files, 54 conflict blocks against main 912112c. Up from 22/27 on 2026-11-01 (175 commits), below 61/99 on 2026-10-31. Largest: Orchestrator.ts, ThreadNotificationCoordinator.tsx and DesktopUpdates.ts with 5 blocks each, then GitVcsDriverCore.ts, ClaudeAdapterV2.ts and FoundationPersistence.test.ts with 3 each. Some of the Claude conflicts come from #276's cherry-picks of upstream pingdotgg#28410 and pingdotgg#29169, which this nightly also contains.

T3 Code closeout: done

Issue: #24 Modular fork: move remaining fork code out of upstream files
Thread: thread:delegated-task:command%3Amcp%3A1178ad5e-4ceb-4452-8960-16ff08db431e%3Adelegate-task%3Aissue-24-owner-20261102
Model: claude-opus-5-5
Branch: issue-24-modular-fork-move-remaining-fork-code
Completed: 2026-11-02T19:11:10.692Z

Both passes now have reviewed pull requests, open and mergeable. Neither pass reduces merge conflicts against today's nightly. Both move fork code out of upstream files. GitHub shows no CI results on either head.

	Pass 4 (#253, worktree placement and cleanup)	Pass 5 (#254, the Issues pane)
Pull request	#283	#285
Head	`9ac293820fc5ddc4acca53d103a01b4556324f04`	`d4f26a54ef13005e320cda649574cc3d09cc4d57`
Base	`main` at `912112ce1c`	`main` at `912112ce1c`
Conflicts before → after	31 files, 54 blocks, 1,298 → 1,290 lines	31 files, 54 blocks, 1,136 → 1,136 lines
Fork code out of upstream files	63 lines in `GitVcsDriverCore.ts`; 28 in `worktreesDirectory.ts`; `git.ts` and its test now match upstream	498 lines, mostly from `rpc.ts`
Checker (GPT-6.1 Sol, xhigh)	Changes needed; all three findings fixed	Approved, no findings

Conflicts are measured against nightly v0.1.7-nightly.20261102.4080. Pass 4's line count includes the conflict-marker lines and pass 5's doesn't, so only the before/after within each pass compares.

Pass 4:

Narrowed design: the 2026-10-30 commits had moved about 970 lines of upstream's own cleanup engine into a fork file. That broke the "upstream code stays in upstream files" rule and grew the conflict in that file from 240 to 883 lines. So I had the worker narrow it: storageCleanup.ts is untouched and only fork code moved.
Checker findings, all fixed:
1. An upstream location check had moved into a fork file. It is back in GitVcsDriverCore.ts.
2. The branch was behind main. It was rebased and force-pushed with a lease, which I authorized for that rebase only.
3. The pull request description understated how much the moved code was restructured. It now says so.
Re-review: I checked the fixes myself rather than running a second Checker review.
vp run check:changed: not rerun after the fix round. Tests beside the changed files and the server and shared typechecks passed.

Pass 5: on today's nightly the Issues pane touches only two conflicts, a one-line selector and one import, and no hook can remove either. The 117 blocks this feature was charged with came from earlier syncs.

Checks: in both passes, tests beside the changed files and the affected typechecks passed. vp run check:changed failed only on known problems, so both commits use --no-verify:

3 typecheck errors in apps/mobile;
72–73 unused exports, tracked in Commit hook fails on every TypeScript change: 74 unused exports on main #256;
lint errors that main's copies of the same files also have.

Left open on #24:

The root merges refactor(worktrees): move fork placement and naming into fork modules #283 and refactor(issues): move Issues pane contracts, tables and wiring into fork modules #285, then closes Modular pass 4/5: worktree placement and cleanup #253, Modular pass 5/5: the Issues pane #254 and Modular fork: move remaining fork code out of upstream files #24.
Commit hook fails on every TypeScript change: 74 unused exports on main #256's unused exports.
Unused worktrees from the Modular pass 4/5: worktree placement and cleanup #253 fix round need removing after merge: the -3 one is confirmed, and the earlier fix round may have left another.
Moving more code no longer cuts sync conflicts on today's nightly. What remains is fork code woven into upstream's own lists and state, where any hook would be as big as the code it moves.

Closing: all five modular passes have landed (#250–#254, via #257, #261, #262, #283 and #285). On today's nightly, more moves no longer cut sync conflicts. What remains is fork code woven into upstream's own lists and state, where a hook would be as big as the code it moves. Each sync still posts its conflict numbers; a new issue can pick up any hotspot that recurs. The unused-export cleanup continues in #256.

Navigation Menu

Repository navigation

Modular fork: move remaining fork code out of upstream files #24

Goal

Already modular (2026-10-12)

To do

23 remaining items

T3 Code closeout: done

T3 Code closeout: waiting_on_sub_issues

T3 Code closeout: done

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Modular fork: move remaining fork code out of upstream files #24

Description

Goal

Already modular (2026-10-12)

To do

Activity

tannerpolley commented on Oct 13, 2026

tannerpolley commented on Oct 13, 2026

tannerpolley commented on Oct 14, 2026

tannerpolley commented on Oct 14, 2026

tannerpolley commented on Oct 14, 2026

23 remaining items

tannerpolley commented on Oct 29, 2026

tannerpolley commented on Oct 30, 2026

T3 Code closeout: done

tannerpolley commented on Oct 30, 2026

T3 Code closeout: waiting_on_sub_issues

tannerpolley commented on Oct 30, 2026

tannerpolley commented on Oct 31, 2026

tannerpolley commented on Nov 1, 2026

tannerpolley commented on Nov 2, 2026

tannerpolley commented on Nov 2, 2026

T3 Code closeout: done

tannerpolley commented on Nov 2, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions