Strategic Finding
Type: priority-shift / process-gap
Horizon: near-term
Open hold-gated agent PRs are being silently superseded by human-authored fixes merged to main, and no process detects the collision. The hold queue keeps PRs open for weeks while humans fix the same defects in parallel.
Concrete instance: common#1010 ([scanner] fix: route reports by booted image, open since 2026-08-20, hold-labeled) rewrites read_image_info in system_files/bluefin/usr/libexec/bonedigger-report to prefer the booted image reference from bootc status over stale baked image-info. The same functional fix was delivered by two merged human PRs:
- common#1020 (merged 2026-08-22): "fix(report): read booted tag from bootc, not baked image-info"
- common#1030 (merged 2026-08-27): "fix(report): read the boot-time booted-image snapshot when bootc status fails"
PR #1010 now duplicates merged behavior and will conflict or double-apply if ever reviewed. It still sits in the queue as if it were live work. A second probable collision: common#978 ([scanner] fix: use valid brew tap trust commands) vs merged common#1047 (fix(brew): trust vendored taps so preinstalled casks actually install) — adjacent ground in brew tap trust.
Rationale
The hold queue is at ~75 PRs and growing (tracked in #1043 and #1052). Every superseded PR inflates the queue, wastes future reviewer attention, and makes queue-size metrics overstate real pending work. Worse, agent lanes keep treating occupied ground as occupied — blocking related work — long after main has moved past it. A queue that cannot tell live work from dead work will not drain no matter how much triage capacity is added.
Proposed Next Step
Add a supersede-check to the hold-gate kick preflight: for each open hold-gated PR, diff its claimed files/functions against git log on main since the PR's creation; flag PRs whose core change already landed (via merged PR reference or equivalent diff) and auto-comment recommending close-or-rebase. Start by auditing the current ~75-PR queue once, manually, to calibrate the heuristic — common#1010 is the first confirmed case.
Filed by strategist agent (ACMM L5 — hold-gated mode)
🐝 Hive Agent: strategist | Instance: hosted-projectbluefin-knuckle-gjvq | SHA: unknown
— hive: agent=strategist backend=copilot model=kimi-k3
Strategic Finding
Type: priority-shift / process-gap
Horizon: near-term
Open hold-gated agent PRs are being silently superseded by human-authored fixes merged to main, and no process detects the collision. The hold queue keeps PRs open for weeks while humans fix the same defects in parallel.
Concrete instance: common#1010 (
[scanner] fix: route reports by booted image, open since 2026-08-20, hold-labeled) rewritesread_image_infoinsystem_files/bluefin/usr/libexec/bonedigger-reportto prefer the booted image reference frombootc statusover stale baked image-info. The same functional fix was delivered by two merged human PRs:PR #1010 now duplicates merged behavior and will conflict or double-apply if ever reviewed. It still sits in the queue as if it were live work. A second probable collision: common#978 (
[scanner] fix: use valid brew tap trust commands) vs merged common#1047 (fix(brew): trust vendored taps so preinstalled casks actually install) — adjacent ground in brew tap trust.Rationale
The hold queue is at ~75 PRs and growing (tracked in #1043 and #1052). Every superseded PR inflates the queue, wastes future reviewer attention, and makes queue-size metrics overstate real pending work. Worse, agent lanes keep treating occupied ground as occupied — blocking related work — long after main has moved past it. A queue that cannot tell live work from dead work will not drain no matter how much triage capacity is added.
Proposed Next Step
Add a supersede-check to the hold-gate kick preflight: for each open hold-gated PR, diff its claimed files/functions against
git logon main since the PR's creation; flag PRs whose core change already landed (via merged PR reference or equivalent diff) and auto-comment recommending close-or-rebase. Start by auditing the current ~75-PR queue once, manually, to calibrate the heuristic — common#1010 is the first confirmed case.Filed by strategist agent (ACMM L5 — hold-gated mode)
🐝 Hive Agent:
strategist| Instance:hosted-projectbluefin-knuckle-gjvq| SHA:unknown— hive: agent=strategist backend=copilot model=kimi-k3