[AbirAbbas]

Summary

Four Tasks fixes merged only into workstreams/main (#2446, #2452, #2461, #2464). Its integration pull request, #2306, was closed unmerged after app/v3 was squashed onto dev (#2561), so none of them reached dev. This carries them, one commit each, in their original order.

What changes for a person:

• A task tree is bounded (Tasks: a worker does its own task, and a task tree is bounded in depth, count and spend #2446). A run worker's plandb add, split, task insert and task pivot refuse a title the run already has, a task more than two levels below the run's root, and a run past 20 counted tasks. Every ordinary run has a dollar ceiling for its whole tree: the lower of the per-task limit (crew.task_cap, $5 by default) and what the conversation's limit leaves. Reaching it ends the run on its cost limit and names how to raise it.
• A worker does its own task (Tasks: a proposed run reads only its assignment, workers add only under their own task, run subtasks keep their parent #2452). A task the conversation proposed hands its run only its own assignment. A worker adds tasks only under its own task, five at most. tasks.list keeps a run subtask's parent. Stopping a run leaves nothing reading running.
• Each check has its own checker ceiling (Tasks: each check in a run gets its own checker ceiling #2461). One check spending its ceiling no longer refuses every later check of the run on its first call.
• The older task road has the same bounds (Tasks: the older task road refuses duplicate work and holds a tree spend total #2464). With CODEAF_TASK_BELT=node, duplicate work is refused tree-wide and the tree has a spend total.

The original pull requests hold the full validation contracts.

dev still has each fault

Before each commit, I ran that fix's own tests against the tree without its code change. Where a test needed a symbol the fix adds, I ran the tests that compile without it.

Fix	Tests run before the code change	Result
#2446	TestRunWorkerRefusesDuplicateTitles, TestRunWorkerTreeBoundsRefuseWholeBatches (on dev 80b5861, with only the two limit constants supplied)	fail: duplicate was accepted in all six cases; bound was not explained: exit 0 for count, batch and depth
#2452	TestStopRootCancelsOpenTasksEvenWhenRootAlreadyEnded, TestStoppedRunRejectsALateRootWorkersReview, TestDrainedWorkerReleasesItsClaimWithoutDiscardingRetainedWork, TestRetainedReadyWorkReadsQueuedOnlyAfterItsRunLeaves	all four fail, e.g. stop left open work, late return reopened or rewrote stopped work
#2461	TestCrewFactoryGivesEachCheckItsOwnCeiling, TestCrewFactorySeparatesCheckEstimatesWithASharedDay, TestScopedCheckSpendEndsTheWholeRunAtTheTaskCap	all three fail: check-b call 0: the check stopped at its spend ceiling of $0.12
#2464	TestNodeTree* (without the one test that reads the new checkpoint field)	8 of 10 fail, e.g. duplicate refusal missing "no: this run already has a task named", queued node bought a request

Nothing was dropped: all four are carried whole.

What differs from the originals

• One conflict, in beltRunNotice (internal/session/task_run_belt.go). dev had moved the result words into noticeResultWords; that helper is kept and the per-task limit words are added to its report. Tasks: a proposed run reads only its assignment, workers add only under their own task, run subtasks keep their parent #2452, Tasks: each check in a run gets its own checker ceiling #2461 and Tasks: the older task road refuses duplicate work and holds a tree spend total #2464 applied cleanly on top.
• One manual section is split. With Tasks: the older task road refuses duplicate work and holds a tree spend total #2464's text the section "A task started from the composer carries a cap" reached about 3,300 characters, and TestTheChatManualAnswersTheQuestionsPeopleAsk failed: "how much money can a task spend before it stops and asks me" should reach tasks; it reached [models-and-cost ...]. The tree-limit text now has its own section, "The per-task limit covers the whole task tree", and the composer section points to it. No test was changed.
• The four original change entries are kept under their own numbers, as Carry the open app/v3 pull requests that integrate cleanly onto dev (16 of 22) #2597 did for the pull requests it carried, with one entry for this pull request.

Tests

CI on this head (c7721f5) is green: light gate, touched codeaf, touched session, touched tui3, touched rest, windows native tests and the required check.

Locally on macOS, make pr-ready did not pass as one command. The PR is marked ready on the strength of CI and the comparison with dev below:

Step	Result
make test-quick	pass
make test-packed-manual manual-gates	pass
touched ./internal/plandb	pass
touched ./internal/run	TestDelegateForwarderReachesNote201 fails; the runner reports already failing on the base
touched ./internal/session (8 shards)	5 failures, each reported already failing on the base: TestACommandLookFailsOnAFailingPipeStage, TestWatchTickFailsOnAFailingPipeStage, TestBashBeltTruncationCapFollowsTheWindow, TestBashBeltTruncationWritesHeadTailAndPath, TestTheFixedPrefixStaysUnderItsBudget
touched ./internal/tui3 (8 shards)	pass: Every selected test passed on its first run.
touched ./cmd/codeaf (8 shards)	red, twice: 30 and 28 failing tests, over the runner's cap of 5, so it made no retries or base probes

The cmd/codeaf reds are 21 distinct tests across the two runs, in Factory, memory pause, hosted take-back, doctor and shared tools; none is a task test. Run by name:

• 12 also fail on unmodified dev (ffeeb78) on this Mac: TestAppFactoryPlanFilesAnswer, TestDoReturnsAfterCancelledAccountingAndItsWriterStop, TestFactoryCommandStepRetryAfterOutOfAttempts, TestFactoryMeasuredOnlyPublication, TestFactoryPersonStopInFlightWakeAndPersonChat, TestFactoryRetryActionsDoNotReportStopped, TestFactoryStepRetryWaitsForPreviousLeaseCleanup, TestFactoryWriteUsesJobBaseAcrossAttempts, TestMemoryPauseResumeAfterTrailingProgress, TestMemoryPauseRetainsPendingForResume, TestMemoryPausedJobLifecycle, TestProcessCloseJoinsTheUsageWriter.
• The other 9 pass on this branch when run by name (ok ... 48.649s). They failed only inside the full runs, which shared the machine with other test suites (load average 26 to 39 during the first run).

Each fix's own tests pass on this branch, along with the neighbouring spend-guard, crew, division and quick-task tests.

Changes after review

A pre-merge review and a hand test found defects that the carried commits brought along. They are fixed in six new commits (2ffcc3e to c245451, head now c245451), and two independent verifiers checked the fixes. Each test below failed on the previous head (c7721f5) and passes now.

• A per-task raise did not raise the tree's dollar ceiling.
  • Found: on the run road, the tree ceiling read only the configured crew.task_cap. Raising one task's limit through a proposal, a reroute or a live raise in /crew changed nothing, so a task raised to $8 still stopped at $5. On a live raise, the guard's refusal also still named the old figure.
  • Changed: the ceiling is now the highest of the configured cap, the task's own raised cap and the crew guard's cap. The engine reads it while the run goes (Limits.CostCeiling) and the ceiling only ever rises. SpendGuard.raiseTaskCap updates the figure and the refusal words together. A live raise also lifts the running tree.
  • Tests: TestARaisedProposalRunsUnderItsRaisedTreeCeiling and TestALiveRaiseLiftsTheRunningTreeCeiling (internal/session); TestARaisedCostCeilingIsReadWhileTheRunGoes (internal/run).
• On the node road (CODEAF_TASK_BELT=node), the tree total counted a finished child's spend twice.
  • Found: when a child finishes, its bill is added to its own node and also to the worker that proposed it. The tree total added both, so a child counted twice and a grandchild three times, and the tree hit its limit early.
  • Changed: the fold now marks the dollars it moved. The tree total counts each node's bill minus what was folded into it, so each dollar counts once. The checkpoint gains foldedUsd, so a restored tree also counts each dollar once.
  • Test: TestNodeTreeSpendCountsAPartFoldedIntoItsOwnerOnce (real nest harness, real folds at two levels).
• The per-task limit sentence was appended at the end of the report and went missing from the screen.
  • Found: "this task reached its $N limit · raise it in /crew" was appended as an extra line instead of leading the report. When the guard refused the root it appeared twice. In the hand test, the engine's new tree ceiling ended the run before the guard refused anything, so the routed crew recorded no stop, and the line under the task row and the entry in the Tasks column did not show.
  • Changed: when the per-task limit ends a run, that sentence becomes the run's failure text. It leads the report and the outcome note exactly once. A routed crew records it as its stop, which brings back the line under the row and the Tasks-column entry. No TUI code changed.
  • Tests: TestAPerTaskLimitEndingSaysItsSentenceOnceWhereThePersonLooks (internal/session, through StartTask to settle); TestAPerTaskLimitEndingIsSaidOnceUnderItsRowAndListed (internal/tui3).
• The tasks manual said a proposed task reads the person's message.
  • Found: "What a task is" in internal/manual/chat/tasks.md said every brief carries the person's message and a journal pointer. A task proposed on the run road actually reads only its own assignment.
  • Changed: the paragraph now separates a typed task from a run-road proposal and points to "What a proposed task reads".
  • Test: TestTasksManualSaysAProposedRunReadsOnlyItsAssignment (internal/manual); make test-packed-manual manual-gates passes.
• plandb task pivot counted the tasks it was replacing. This finding was marked "evaluate" in review. It was real and the fix was small, so it is fixed here.
  • Found: a pivot adds the new tasks before it cancels the old ones. The add has to come first, because cancelling the last open child completes the parent. So the five-per-worker limit and the duplicate-title check counted the tasks about to be cancelled, and a worker with three parts could not pivot to three new ones.
  • Changed: the bounded add now takes the list of tasks being replaced and leaves out the ones still open when it counts totals, the worker's allowance and taken titles. A replaced title can be reused. Finished work the pivot keeps still counts.
  • Test: TestWorkerPivotCountsOnlyTheWorkThatSurvivesIt (internal/plandb). A six-part pivot is still refused whole.

Docs moved with the behaviour:

• internal/manual/chat/tasks.md, how-tasks-run.md and models-and-cost.md, plus a new manual test.
• The change entries 2446, 2452 and 2464.
• This PR's entry 2615-tasks-fixes-carried.md, which now lists these fixes.

The wire protocol and the system prompt did not change. The new RunSpec.CostCeiling and Limits.CostCeiling are internal Go fields.

Earlier sections that are now out of date:

• The conflict note under "What differs from the originals" no longer holds. The per-task limit words are no longer appended to noticeResultWords' report. They lead it as the run's failure text.
• The CI results under "Tests" are for the old head, c7721f5.

Left as is: TestAClosedNodesDollarsAreStampedOnTheConversation fails on this Mac the same way on the new head, on c7721f5 and on unmodified dev (ffeeb78): meta.json says the conversation spent 0 once its node closed, want the node's $0.75. This PR does not cause it, so it is not changed here.

• A finished run is not a limit stop (found by the review of the fixes above). The engine ends a run done with Limit cost when its final receipt crosses the tree's ceiling; the limit helpers fired on Limit cost alone, so such a finished task got the per-task limit as its crew's stop: a limit warning on the task sheet, no "run again with a stronger crew", and no completion announced. taskLimitWords now answers nothing for a run the engine ended done, as dev never stopped a crew for one. Proving test: TestAFinishedRunAtItsCeilingIsNoPerTaskLimitStop (failed with the crew stopped on "this task reached its $5 limit · raise it in /crew" before the change, passes after). Commit bd15299.

🤖 Generated with Claude Code