Skip to content

work-items(tracker-seam): another actor produced a full PR for a claimed issue without ever holding the lease #920

Closed as not planned

Description

@kyle-sexton

Lane: work · Category: operational (coordination gap) · Observed: 2026-07-21/22, this session.

Observed vs expected

Observed: I (this work lane) picked issue #740, called the tracker-seam claim verb, and got a clean successful lease (acquired_at: 2026-07-21T23:17:39Z, no rejection) — implying no other holder was contending for it. I dispatched a worker, which built an independent fix. Minutes later I discovered PR #903 (Closes #740) was already open, with its first commit timestamped 2026-07-21T18:14:09Z — roughly 5 hours before my claim succeeded — authored under the same account identity but with zero relation to my worker's transcript (different branch, different implementation approach, no memory of it in the dispatched worker's own context).

Expected: if another actor (a different lane/session, autonomous or human-launched) is actively working an issue via a real PR, the tracker-seam claim on that issue should either already be held by them (causing my claim call to exit 7/back off) or should reflect their activity somehow. Instead my claim succeeded cleanly, meaning the other actor produced ~5 hours of work and an open PR without ever acquiring the seam's lease — the coordination mechanism this skill (and the standing loop rules) depend on to prevent exactly this race was silently bypassed by whatever produced PR #903.

Impact

  • A dispatched worker (worker-740, this session) burned a full implementation+test cycle on a fix that turned out to be redundant, because the seam gave no signal that the issue was already spoken for.
  • The orchestrator had no way to detect the collision until CI/review time, and only caught it because the worker itself independently noticed a branch-identity mismatch while trying to comply with a routing instruction — not because of any process signal.
  • Resolution required escalating to the human operator to break the tie (this session did so; user chose "stand down, leave fix(guardrails): distinguish --config-env from -c/--config in shared git parser #903 alone" rather than guess).

Root cause (unconfirmed, worth investigating)

Likely one of:

  1. PR fix(guardrails): distinguish --config-env from -c/--config in shared git parser #903 was produced by a process that writes PRs/commits directly without ever calling work-item-tracker.sh claim (e.g. a manual/ad-hoc session, or an older lane predating the claim-before-dispatch invariant landing).
  2. A lease was held and expired/was reclaimed before my claim, and the other actor kept working past its lease without renewing — the claim protocol's "abandoned lease" recovery path may not distinguish "genuinely abandoned" from "still working but not renewing."

Suggested fix direction

  • An operational convention (not necessarily new tooling) that ALL PR-producing paths in this repo's automation — not just /work-items:work — call the seam's claim verb before pushing/opening, even if they don't use the rest of the work skill's flow. Absent that, a cheaper mitigation: before dispatching a worker, in addition to the existing "no open PR" search, also check for very-recent local commits on ANY branch containing the issue number in its name pattern (*<N>*) as a secondary signal that someone is already mid-flight even without an open PR yet.

Activity

  1. kyle-sexton commented on Jul 22, 2026

    @kyle-sexton
    MemberAuthor

    Second occurrence this session, same session: issue #832. This lane claimed #832 via the seam at 23:16:35Z, dispatched a worker that committed at 23:58:40Z and opened PR #908 — but PR #910 (feat/832-go-ecosystem, first commit 23:43:24Z, opened as PR at 00:13:47Z) already covered the same issue independently, again with no seam lease ever held for it, and landed materially more complete (full green CI incl. both review passes, non-draft) by the time this lane's PR opened. Resolved the same way as the #740/#903 instance: closed this lane's #908 as its own redundant work, released the claim, left #910 untouched. Two independent instances in one session strengthens the case that whatever is producing these PRs (#903, #910) genuinely never calls the seam's claim verb — worth prioritizing this fix.

  2. kyle-sexton commented on Jul 22, 2026

    @kyle-sexton
    MemberAuthor

    New occurrence — #915, this session, 2026-07-22.

    Same class: lease held first (claim on #915 succeeded 05:43:31Z, no rejection, no contention signal), yet PR #960 (fix/915-powershell-tool-bypass) landed ~06:35Z from a different actor ("tower-dispatched" per the issue's own comment) with zero seam interaction visible from this lane's side. Third occurrence this session of the same root cause (prior: #740/#903, plus one other referenced in this issue's body) — a non-lease-holding actor producing a real PR against a leased issue, with no seam signal to either side.

    Parked the #915/#960 resolution on the issue itself (comment: #915 (comment)) rather than unilaterally closing/superseding another lane's PR — consistent with this issue's own point that the coordination gap, not any one lane's judgment call, is the thing that needs fixing.

  3. kyle-sexton commented on Jul 22, 2026

    @kyle-sexton
    MemberAuthor

    4th occurrence (2026-07-22 17:53Z), branch fix/915-powershell-tool-bypass — loss-free via head-assertion again. #903 squash-merged at 17:49Z, staling #960 to DIRTY; at 17:53Z a lane under the shared identity pushed a "2nd restack" merge commit (dd550e0) while a tower fixer held a dispatched restack task on the same branch (claimed via its brief; expected head 1f80229). The tower fixer head-asserted, found the branch moved, STOPPED without forcing — designed safety behavior; zero loss. Content adjudicated identical (one cosmetic char); dd550e0 accepted.

    Contributing factors: (a) shared bot identity blocks attribution (#450); (b) the v16 hands-off/tower-lock block depends on every lane actually running v16 — the babysit lane ran a round-14 prompt (no tower-lock) until ~17:5xZ (checkpoint finding, #502); (c) merge-staleness reflexes fire within minutes of any main merge, faster than tower dispatch round-trips — claim comments on the PR are the only race-safe signal, and the tower dispatch did not post one on the PR itself. Tower error to log: dispatch briefs assert heads but did not CLAIM the PR with a comment — claim-comment-first applies to the TOWER too. Adding to retro material.

  4. added
    needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.
    on Jul 23, 2026
  5. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — gate-audit claim). Claiming this item for one gate-audit cycle: an independent panel re-classifies whether the needs-human / status: needs-decision gate is still warranted, or whether the item is C2/C3 and liftable to agent-ready. Outcome comment follows in this cycle.

  6. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — gate audit). Independent panel: two fresh-context agents (one fable, one opus), each given only the issue content and the repo's own plugins/autonomy/reference/guardrails/work-classes.md, plus the triage lane's own read. This lane never executes work items, authors fixes, or merges.

    Gate audit: GATE UPHELD — C4, root cause unconfirmed and fix shape undecided

    Panel: 2-of-2 STAYS-GATED, both C4. Lane read concurs.

    The bug is real and recurring — four documented occurrences spanning multiple lanes and the tower. But the issue's own heading reads "Root cause (unconfirmed, worth investigating)" and it lists materially different causes whose fixes diverge:

    • a PR-producing path that never calls claim at all, versus
    • lease-expiry recovery that cannot distinguish abandoned from still-working.

    The fourth occurrence comment widens rather than narrows the shape, adding shared bot identity (#450) and stale prompt-round skew (#502) as contributing factors. The candidate fixes — a binding claim-before-push convention, a pre-dispatch branch-scan mitigation, and lease-renewal semantics — are not interchangeable, and no decision is recorded. A binding convention over every PR-producing automation path is a coordination-contract change, hence C4.

    Decision required

    Which fix shape binds: a mandatory claim-before-push convention, a pre-dispatch detection heuristic, or lease-renewal/abandonment semantics?

    RECOMMENDED — split it, and get the cheap half moving now. File a C1 investigation to confirm which actor path skips claim (that answer decides everything else), and independently land the pre-dispatch open-PR + recent-branch scan as a scoped mitigation — it is valuable regardless of which root cause wins, and it stops the bleeding while the convention question is settled. The 4th-occurrence retro already points toward the convention ("claim-comment-first applies to the TOWER too"), so expect to land that too, but on evidence rather than assumption.

    Alternative — rule the convention directly now: every PR-producing automation path must claim via the seam and post a claim comment before first push, treating the branch scan as redundant once enforced. Faster to state, but it commits to a fleet-wide obligation before the root cause is confirmed.

  7. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — claim release). Gate-audit cycle complete for this item; the triage claim posted earlier this cycle is released. No lease is held — another instance or lane may pick this up immediately. The outcome comment above this one carries the classification and either the lift rationale or the decision summary.

  8. kyle-sexton commented on Aug 13, 2026

    @kyle-sexton
    MemberAuthor

    Frontier disposition (drain session)

    Ruling: Close as not planned for this drain wave (frontier consensus).

    Dual-panel drain ruling: advisory, superseded, folded into a program umbrella, or blocked on a host/design carrier without an actionable in-repo slice this session. Re-open with a concrete acceptance brief if a consumer needs the work.

    KEEP set for this wave: standing telemetry; host-physical Class C (#631 family); program umbrellas #530/#531/#532/#535; live design carriers #1500/#1553/#1668/#1695/#1716/#1681/#1718; #1360.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.

    Type

    Fields

    No fields configured for Task.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions