Repository navigation
Finalization: confirm the owning agent is idle before retiring its worktree #230
Description
What: Finalization lets the merging agent retire a task worktree while the authoring agent thread is still active in it.
Why: "Retire at merge" names the merger as the remover. The activity check it implies, processes whose cwd is inside the worktree, misses an agent thread that is mid-turn or about to resume.
So what: The finalization procedure should require confirming that the owning writer is idle, or handing retirement to the owner, before removing another agent's workspace. An AgentsMD maintainer agent makes the change.
Outcome
A merging agent never removes a workspace another agent thread is still using.
Evidence
On 2026-10-18 in toolboxmd/chromeria, the #437 planner (a Claude Opus thread) merged PR #437 at 17:06:12Z. In the same step it removed /home/luke/dev/toolboxmd/chromeria-worktrees/222-terminal-runs and the fix/222-terminal-runs branch. The worktree was clean, at the merged head, and had no processes with their cwd inside it.
The authoring Codex thread had just received a new user directive (integrate, merge, stage) and was running integration preflight in that worktree. It found .git and its sources gone mid-check and raised a concurrent-delivery alarm. Nothing was lost, since the work was already squash-merged. Still, two agents both believed they owned delivery.
Proposed change
In skills/operations/references/finalization.md, under Resources: before removing a workspace whose writer is another agent or thread, confirm that writer is idle and has no queued or pending work (for example, the host's thread or task status), or hand the retirement to that owner. A process-cwd check alone does not establish that a workspace is inactive. Name in the handoff the agent that will retire the workspace.
Acceptance criteria
- Finalization states the owner-idle check, or owner handoff, for workspaces written by another agent.
- The procedure's checks still pass.
Non-goals
- Changing who merges, or adding lock files or services.
Blockers
- None.
Required proof
- The changed procedure text, with the package checks passing.
Elon record
- Requirements and who asked: The user wants found problems filed at their owner. Another agent's alarm about concurrent delivery shows the gap.
- Deleted: Lock files and ownership services. A status check already exists in the host.
- Bottleneck: Workspace ownership is not checked across agent threads before retirement.
- Checked myself: Read
finalization.mdanddelivery.mdin full, plus the observed timeline: the merge at 17:06:12Z, the worktree removal in the same step, and the author's preflight alarm.
Activity
lukemaj commentedon Oct 18, 2026
What: Implemented #230 in component PR #231.
Why: Finalization previously lacked an owner-idle or owner-handoff requirement before removing another agent's workspace.
So what: The backlog planner arranges independent exact-head review before integration; #230 remains open until final delivery.
Canonical implementation handoff
fix/230-owner-idle-retirement./home/luke/.t3/worktrees/agentsmd/fix-230-owner-idle-retirement, owned exclusively by this implementation sub-agent.2f04c90a926a986f82ebba00971cfb5f98ffbe01,integration/backlog-2026-10-18.2da6d73b3d4218478a885694d002b86a50a3c4b7, committed usingVERSIONCTL_WIP=1and pushed.skills/operations/references/finalization.mdandtests/test_operations_contract.py. The active owner's received handoff, naming the workspace, now explicitly satisfies the merger's same-step obligation; the owner retires the workspace when its work ends. Delivery remains unchanged because it delegates to finalization.Review round 1 disposition
Accepted Luna's blocking finding. Added the same-step exception and owner completion timing in the existing Resources paragraph. Extended the contract test: the two new subtests failed before this clarification and passed afterward. Fresh exact-head independent review remains pending.
Proof
Before the prose change:
Each subtest failed because its corresponding retirement obligation was absent. The same focused command passed after the prose change.
On the committed candidate:
CI on the new head is pending. The previous head failed only at the expected version-transition step (
BUMP_REQUIREDandTAG_CONFLICT); that previous result is not claimed for this candidate. Version remains unchanged for this internal component.Elon record
Requirements and who asked: #230 and the implementation brief require checking another writer is idle with no queued/pending work or handing retirement to the owner, rejecting process-cwd-only evidence, and naming the retiring agent.
Deleted: No lock files, services, changed merge roles, duplicate delivery rule, or component version bump.
Bottleneck: Cross-thread workspace eligibility was not checked before removal.
Checked myself: Read the Issue and both procedures, observed six red subtests before the prose change, inspected the complete diff, ran both full required suites, and inspected CI steps and failure logs.
Delivery and remaining work
Implementation and deterministic proof complete. Self-review complete. Independent exact-head review remains for the backlog planner; no review status or review exemption claimed. Version remains 17.2.0 under decision-tree case 2; the final approval PR owns the single transition. No merge, release, marketplace promotion, installation, or finalization performed. Instruction-package proof does not establish behavioral Live Verification.
Workspace and branch retained for review and possible fixes. Retirement is assigned to this implementation sub-agent in this thread after verified terminal disposition, rather than removal by a different active thread. Next action: independent review, then integration under the planner's authority.
lukemaj commentedon Oct 18, 2026
Train state (2026-10-19): PR #231 merged into
integration/backlog-2026-10-18(head2da6d73). Luna round 1 blocking finding (same-step retirement vs owner handoff) fixed in2da6d73; Luna round 2 pass: #231 (comment). Suite 425 + 7 OK on the author's run. Closes with the train's final PR to main; not released.release: 17.3.0 backlog train (#226, #230, #225, #164, #213)
Merge pull request #242 from toolboxmd/integration/backlog-2026-10-18