Closed
Description
Fork cannot follow upstream
Plan: The plan did not finish. Run: https://github.com/scuttletoby/t3code/actions/runs/44928132410
The series does not fit the seam list (.github/fork-seams.json):
Fork changes to upstream files must stay within the seam lines in .github/fork-seams.json (list from the head):
- apps/server/src/mcp/OrchestratorMcpService.ts: 5 upstream lines deleted, moved or re-indented, more than its "lost" lines count of 4 (keep upstream's lines; a fork change goes in a file of the fork's own).
- apps/server/src/mcp/OrchestratorMcpService.ts: 5 upstream lines deleted or changed (its text is gone), more than its "lost" gone count of 4 (keep upstream's lines; a fork change goes in a file of the fork's own).
- apps/server/src/mcp/OrchestratorMcpService.ts: "pending" with 35 fork lines, more than its ceiling of 34 (fork lines to move out, not add).
Put the change in a file of the fork's own (a fork setting goes in packages/contracts/src/forkSettings.ts; see docs/fork/updates.md). A new seam line needs its own entry in .github/fork-seams.json, which the owner approves.
The series was left as it is.
| What | State now |
|---|---|
Series branch main |
at e0e01b97ae8f |
Candidate branch fork-follow/candidate |
at 8b669dfc567d |
| Backups of earlier series (tags) | fork-follow/pre-v0.1.11-nightly.20261118.4890-1974264545d5, fork-follow/pre-v0.1.11-nightly.20261119.4954-f49745fab25a |
| Release | none: the plan did not pick a commit |
Metadata
Metadata
Assignees
Labels
No labels
Activity
fork seams: record the OrchestratorMcpService.ts line #660 changed (f…
github-actions commentedon Nov 21, 2026
Fork cannot follow upstream
Plan: The plan did not finish. Run: https://github.com/scuttletoby/t3code/actions/runs/44929539061
Rebasing the series from v0.1.11-nightly.20261119.4954 onto v0.1.12-nightly.20261121.5051 stopped at: ce9e2c4 Auto-update (server): background service updates itself when quiet (#5)
Conflicting files:
What git said:
The series was left as it is. To follow by hand:
git fetch https://github.com/pingdotgg/t3code.git tag v0.1.12-nightly.20261121.5051,git rebase --onto v0.1.12-nightly.20261121.5051 v0.1.11-nightly.20261119.4954 <series branch>, resolve, run the tests, then push the branch with--force-with-lease, and start the follow workflow once (gh workflow run fork-follow.yml --repo scuttletoby/t3code):it releases the series and disables the upstream workflows that push started.
main4549be532b35fork-follow/candidate8b669dfc567dscuttletoby commentedon Nov 21, 2026
Hand rebase onto v0.1.12-nightly.20261121.5051: candidate ready for review
Candidate: branch
fork-follow/candidateat733d372dfc295f5f850d6c5ecfc5b2c22311749c(409 commits on the tag).mainis not moved; nothing released; fork-follow.yml not started.Fork CI: https://github.com/scuttletoby/t3code/actions/runs/44936643721 (fork-tests.yml dispatched on the candidate sha, as in #553).
Seam check:
node .github/scripts/fork-seam-check.cjs --from-upstream v0.1.12-nightly.20261121.5051 HEADpasses.What upstream changed in the conflict files
repositorieslist, and a plain (non-git) project folder skips the branch/base checks, fetches and temporary branches. A created task folder is tracked, andTaskFolderErrorjoins the error type. The conflict was only the imports: both kept (TaskFoldersand the fork'sWorkAdmission).delegate_task: upstream now runs the dispatch andprepareRuninsideEffect.uninterruptibleMask, so a launch whose caller goes away still starts. It also takesparentNodeIdfrom the parent run first. The fork's lines from feat(usage): usage pacing gate on delegate_task (pacing part 4) #73 (pacing note), feat(slots): agent slot gate on delegate_task (off by default) #89 (slot gate), fork: delegate checks behind one fork module (SEAM) #174 (two check seats), fork seams: green from upstream on main; a PR's own list is format-checked (#385) #426 (pacedTarget) and fix(model-guard): drop an inherited agent option; compare Claude levels with T3's model list #660 (comment) were rebuilt on that version: seat 2 and its hold before the mask,forkHold.started/forkHold.releaseinside it,restore(...)only around upstream's wait/read, andEffect.scopedkept.claudeEffectiveQueryPolicyKeygained aworkspaceargument (a changed folder list needs a new live query). Kept with the fork's MCP profile lines (feat(server): Claude threads apply their MCP profile's outside servers #528).Layer.provideMerge(layerTaskFolders)next to the fork'smcpProfileTurnCheckline (Refuse a narrowed MCP profile at turn start on a provider that cannot apply it #679).Fixes after the rebase (folded into the commit that owns each)
Typecheck and tests found four things the merge alone did not show:
layerRuntimeCoreDependenciesBase'spipeplus the fork's two make 21, and Effect'spipetakes at most 20 (TS2554). The fork line now carries upstream's next one, wired exactly as before:Layer.provideMerge(ForkLayers.mcpProfileTurnCheck.pipe(Layer.provideMerge(layerTaskFolders))), // Fork: MCP profiles. In fork-seams.json: the server.tsaddedtext for that line updated, itswhyexplains it, and itslostcount goes from 6 to 7 (lines and gone): upstream'sLayer.provideMerge(layerTaskFolders),is the one line lost. This raises a count, so it needs your OK. The other route, splitting thepipewith a fork).pipe(line, makes the formatter re-indent about 35 upstream lines, so I did not take it.claudeEffectiveQueryPolicyKeyto be a pure function (same input, same key). The fork gave overrides with no worked-out plan a new key on every call. They now get one fixed keyunplanned, which still never matches a planned turn's key (undefinedfor all, ormcpNUL ...). In production the adapter always plans before it computes the key, so only tests reach this. The fork's ownplanKeyFortest was updated to match. Also in feat(server): Claude threads apply their MCP profile's outside servers #528: the fork's fake Claude session gains upstream's three new methods (mcpAuthenticate,mcpSubmitOAuthCallbackUrl,reconnectMcpServer).mcpApps/McpSignIn.tsreads a live session throughProviderSessionManagerV2and never starts one, so it is added to the known users inworkAdmission.startPaths.test.ts.pullRequestWatchSlot.test.tsnow provideslayerNoTaskFoldersandModelManifest.layerTest, which upstream's runtime layer now needs (as in upstream's own runtimeLayer test).Plus one new commit on top: fork seams: exclude upstream's new
.repos/effect/.../test/utils/forkRegistration.tsfrom fork files. Upstream vendors that Effect test helper, and its name matches the fork-file pattern. It adds one!exclude, changes no count, and is a separate commit because no seam commit owns it (the same as #554 last time).Range-diff (
v0.1.11-nightly.20261119.4954..4549be53vsv0.1.12-nightly.20261121.5051..733d372dfc)git merge-tree): the tree differs only by the conflict resolutions, the fixes above and the seams exclusion. No upstream line is undone except the one server.ts line in fix 1.0.1.11 to 0.1.12
No fork code, test or release script assumes 0.1.11. The only hits are made-up versions in test fixtures (fork-follow.test.sh synthetic tags, fork-release.test.sh, fork-kit go-live and agent tests). Upstream's own package.json files still say 0.1.11 at the v0.1.12 tag.
Local checks (under the capacity gate, one at a time)
vp run -r typecheck): 43/43 ok. Lint: ok (warnings only, in upstream files).StatementSync.stat, which this machine's Node 24 lacks. The repo pins Node 26.13.1, which CI uses, so the run above decides it.Conflict resolution diff (conflicted merge to final, per file)
Fixes after the rebase (first rebase result ebe4045551 to candidate)
Note: GitHub shows the escape backslash-u-0000 (a NUL character) in two diff lines above as
\^@; the code itself has the normal\u+0000escape.fork seams: record the OrchestratorMcpService.ts line #660 changed (f…
scuttletoby commentedon Nov 21, 2026
Review of the rebased series
Candidate:
fork-follow/candidateat733d372dfcon upstreamv0.1.12-nightly.20261121.5051. Old series:v0.1.11-nightly.20261119.4954..4549be532b.Verdict: NOT SAFE to move main yet. The rebase itself is clean, but the candidate's own fork CI (run 44936643721) is red: 3 fork tests fail. Old main passed the same suite (Fork tests run 44929515497). All 3 need small fixes, and each one comes from an upstream change in this range that a fork test or fork setting did not follow. The local run in the report did not include these test files.
What I checked myself
git range-diff: 386 identical, 22 changed, 1 new. This matches the report.OrchestratorMcpService.ts(the same fork lines, only re-indented inside upstream's newuninterruptibleMask) andserver.ts(fix 1).fork-seams.json, adds only the server.ts entry and the one!.repos/effect/...forkRegistration.tsexclude.startedandreleaserun inside it.restorewraps only upstream's wait or read.Effect.scopedis kept. Correct.BLOCKING (candidate CI red)
apps/server/src/policy/ManagedPolicy.test.ts:71(upstream's new test) fails onhandoff.idleTokens(job "Test Server 8").listManageableSettings(apps/server/src/policy/managedSettings.ts:250) goes down into every settings struct. ThenmakeManagedSettingsrefuses any path inside a setting that decodes when set tonull(managedSettings.ts:342-354).handoff(packages/contracts/src/forkSettings.ts:114) reads any bad value,nullincluded, as off. Upstream's check therefore treats it as a setting that can be empty.t3 policy settings(apps/server/src/cli/policy.ts:166) listshandoff.idleTokensand the other handoff fields, and a policy file that uses any of them is refused.mcpCarriedProfiles,modelPolicy,playtestandreviewRules. They are very likely hit the same way: the test stops at its first failure, andhandoffsorts first.managedSettings.tsthat reads a fork-owned set, or a fork-file helper, and not a change to upstream's test.apps/server/src/fork/remoteReplyRead.test.ts:296"an older version's reply on its last page is read by the tail probe" (job "Test Server 5").afterPosition: -1for the first page (apps/server/src/peer/RemoteDelegation.ts:641). Upstream's ownRemoteDelegation.test.tschangedtoBeUndefined()totoBe(-1).olderPagedAnswer(remoteReplyRead.test.ts:143) still matches onlyafterPosition === undefined. It therefore answers{unexpectedPage: -1}, and the task fails.-1as the first page. The fix(remote-delegation): a reply from a linked environment on another version still completes its task #540 code is unaffected, since it passes the input through unchanged.fork-kit/model-guard/tests/guard.test.py:1958test_each_built_in_driver_reads_its_effort_where_the_guard_looks(OpenCodeDriver) (job "Fork kit tests 1").OpenCodeRuntime.openCodeVariant(packages/provider-opencode/src/server/OpenCodeRuntime.ts:396). Both adapters now call that helper.variant, soDRIVER_EFFORT_IDSstays right.readsatOpenCodeRuntime.ts, and check that both adapters callopenCodeVariant(.MINOR
t3:default(packages/shared/src/model.ts:288), meaning "send no variant", is not known to the fork's model policy or model guard.modelPolicyPickEfforts(packages/contracts/src/modelPolicy.ts:178) would read an OpenCode pick withvariant: "t3:default"as the effortt3:default.enforce, it refuses that pick unless a row listst3:default, and Settings may mark it as outside the policy. It fails closed, so it is not a safety problem.t3:defaultas no effort. This can come after the rebase.refs/heads/fork-follow/candidate, not on main withref=, so no fork-ci-result is recorded. After the fixes, re-run it as the error says (from main withref=fork-follow/candidate).The three points asked about
apps/server/src/server.ts:767), allowance 6 → 7: OK.X.pipe(provideMerge(T), provideMerge(TF))andX.pipe(provideMerge(T.pipe(provideMerge(TF))))give the same layer graph. TF is still built first, is provided to T and to everything above, and is not given T's wrapped registry. Layers memoize by reference, so it is built once.lost6 → 7 infork-seams.jsonis the only widening.handoffThreadsSkipIdleCompaction) must stay closer thanlayerOrchestrationApplication, or the workers get upstream's own orchestrator. So the two fork steps cannot be merged into one.).pipe(split re-indents upstream lines.planKeyFor"unplanned": OK.mcpProfiles(...)on the same overrides object right before the key (ClaudeAdapterV2.ts:13128-13134).PLAN_KEYS.setruns only after a plan succeeds, and a refusal ends the turn before the key line. Sounplannedis not reachable there.all) or starts withmcpNUL.queryContextRef, so there is no old state with counter keys after a restart.unplannedwere ever reached (tests only), two such turns on the same native thread with equal servers, tools and workspace could reuse one process. Neither turn has a profile, so nothing is widened.OrchestratorMcpService.slotGate.test.ts:638): not weakened. It follows a real upstream change.prepareRuninsideEffect.uninterruptibleMask. Interrupting the caller therefore waits for the 60 s fake dispatch, and the old test'sFiber.interrupt+ immediatedelegatehung.afterWorkspace), before the mask. That is now the only point where a cancel can still release through the scope finalizer. The failed-dispatch and no-task-event tests cover the same finalizer.fork seams: record the OrchestratorMcpService.ts line #660 changed (f…
fork seams: record the OrchestratorMcpService.ts line #660 changed (f…
scuttletoby commentedon Nov 21, 2026
Candidate fix round
fork-follow/candidatemoved from733d372dfcto7910eaab87. Each fix is folded into the fork commit that owns the code or test. Nothing else in the series changed.1. Blocking: ManagedPolicy.test.ts
handoff.idleTokens. Fixed in fork code, folded intoa4b3b42476→e610aa9e8f"MCP profile carry ... (#640) (#648)". That is the first commit where all the affected settings exist.apps/server/src/policy/forkManagedSettings.tslists them: handoff, modelPolicy, playtest, reviewRules, mcpCarriedProfiles, mcpProfile, mcpProfiles, mcpSubagentProfile, mcpThreadProfiles. These are exactly the fork settings that readnullas off or a default.listManageableSettingsgets one seam line. It stops at these settings, so it listshandoffbut nothandoff.idleTokens. The seam line is recorded in fork-seams.json.null. They keep their per-field paths, so the pc-agents kit'sagentSlots.maxPerMachinelock still works.handofformodelPolicyas a whole, a user change to any field is refused as Locked. For handoff that covers mode, source and command, instructions, sizeTokens, idleTokens and threads. For modelPolicy it covers turning it off and changing its rows.handoff.idleTokens,modelPolicy.mode,playtest.enabledorreviewRules.presetis refused with the plain reason "... is inside a setting that can be empty; manage "handoff" as a whole".nulland is missing from the list.2. Blocking: remoteReplyRead.test.ts "tail probe". This was a test-data fix only, folded into
c8c0374210→f004647fa9(#540). Upstream now asks for the first page withafterPosition: -1, so the fake reader now checks for-1instead ofundefined. The fork code,remoteReplyRead.ts, never readsafterPosition; it passes the input through unchanged. So the fork code needs no change.3. Blocking: model-guard guard.test.py OpenCode case. Folded into
a091bc54de→1a20b16d83(#629, #638). It changes the test, the kit job's sparse-checkout lists and one comment; the guard code is unchanged.openCodeVariantinpackages/provider-opencode/src/server/OpenCodeRuntime.ts. The test pins that helper reading"variant".server/adapter.tsandserver/v2/adapter.ts) calling the helper.4. Minor: OpenCode
t3:default. Fixed inpackages/contracts/src/modelPolicy.ts, folded into0bf94f9c0e→5fba2cbc92(#731, #747).t3:defaultnow reads as "no effort option". Upstream sends no variant for it, which is the same as leaving the option out. So a row without efforts allows it, and a row with efforts still refuses it, because the pick names no effort. It never reads as a higher effort.DEFAULT_VARIANT_CHOICE_ID.t3:defaultas an effort. That is fail-closed, so I did not change it.5. Checks. The seam check from upstream (
fork-seam-check.cjs --from-upstream v0.1.12-nightly.20261121.5051 HEAD) passes. Local results on Node 26.13.1:vp run typecheck(all workspaces, then apps/server again after a lint fix) andvp lint: pass.fork-ci-tests.cjs): the workflows part and the fork-ci/fleet-walls part pass.__pycache__folder that the local fleet-walls run left behind. Neither file differs from733d372dfc, and both pass on CI.6. CI on
7910eaab87.ref=fork-follow/candidate, as the review asked: run 44942368885 failed. Every test job passed exceptFork kit tests 1, which still cannot findOpenCodeRuntime.ts. A dispatch from main uses main's own fork-tests.yml, and main's sparse lists do not have the new path yet. So this job cannot pass from main until main moves. After that, a run on main uses the new lists.range-diff
733d372dfc...7910eaab87: 148 commits are compared. 144 are=and 4 are!, and the 4 changed ones are exactly the commits listed above. The final tree equals the old tip plus the four fixes.github-actions commentedon Nov 21, 2026
Fork cannot follow upstream
Plan: Release the series on v0.1.12-nightly.20261121.5051 as it is. Run: https://github.com/scuttletoby/t3code/actions/runs/44944687121
A job failed without writing a reason (results: guard success, plan success, release failure, advance skipped); see the run.
main4460c8502297fork-follow/candidate7910eaab87a57910eaab87a5github-actions commentedon Nov 21, 2026
Fork cannot follow upstream
Plan: Release the series on v0.1.12-nightly.20261121.5051 as it is. Run: https://github.com/scuttletoby/t3code/actions/runs/44944687121
This is attempt 2 of the run, but the guard passed in attempt 1: "Re-run failed jobs" and "Re-run job" keep the guard's earlier result without checking again, so the jobs that get FORK_PUSH_TOKEN were skipped. Use "Re-run all jobs", or start a new run (
gh workflow run fork-follow.yml --repo scuttletoby/t3code).main4460c8502297fork-follow/candidate7910eaab87a57910eaab87a5github-actions commentedon Nov 21, 2026
Fork follows upstream
Plan: Release the series on v0.1.12-nightly.20261121.5051 as it is. Run: https://github.com/scuttletoby/t3code/actions/runs/44946638590
main58d5012fae6afork-follow/candidate7910eaab87a5d4dccd3fa003