Everything I could decide myself is decided. What's left needs
you. Each one carries the recommendation of the session that raised it — if you
agree, pick it and press Done. A model's opinion, where there is one, appears on the card
as a receipt line; cards with no receipt were never put to a model. "I'm not sure" is a real answer and becomes work for me.
You don't need to open the session — pressing Done sends your answer back and it picks up.
12 questions on this
In plain terms. Two things need your hand on this board: a finished, approved fix that the robots are not allowed to merge because it changes the merge machinery itself, and whether to switch this board back on with me named as its driver after the system mistakenly shut down five drivers in a row.
1 of 2 answered
Tried first, unsuccessfully: receipt f7fdb9f25 — Fable was asked the parked-board question twice with the options reversed and FLIPPED (as-given: re-register the driver; reversed, higher confidence: leave it parked and merge the roster fix first), so it escalated rather than ruled. Both readings turn on the same unmerged fact: the roster under-reports liveness (agentflow#165, open), so whether re-arming is safe is not knowable from here. The merge question was never put to Fable: a change to the merge machinery is path-held by design and no model can press the button.
1PR PropFlow-Technologies/propflowai#7997 (make the merge gate wait for reviewers to be DONE with a commit) is CLEAN, MERGEABLE, green at the head SHA and formally approved by the review bot at the current head. It edits .github/workflows/auto-merge-all.yml, which is the first entry in that workflow's own DANGEROUS_GLOBS list, so auto-merge path-holds it and no agent session may press merge. Requesting an attended human merge. Note: the board row for this bug cannot be marked done until it merges — the tool that writes a closing reference hard-refuses any reference to an unmerged PR.
In plain terms. Can I get a finished fix merged? It is tested and approved and one click away, but it changes the machinery that decides what merges, so by design the robots are not allowed to press the button on it - only you can.
Answered: Merge #7997 now
2The task file for sess-propflow-ef was set to `parked` (terminal and unsupervised) at 03:38Z after the supervisor hit its operator respawn cap of 5, and it names a blank, never-used session as the operator. Five previous drivers were shut down mid-run because the roster that reports which sessions are alive under-reports (measured: 2 visible of 9 alive); each shutdown destroyed that driver's crew of ten sub-agents. I am the sixth, I hold the criterion, my wake loop survived the park, and ten lanes are working now. While the task reads `parked`, the durable watcher stands down and the claim fence cannot hold anything. The fix for the under-reporting is written and unmerged (PropFlow-Technologies/agentflow#165, 'Absent from the roster is not dead'). Fable was asked twice with the options reversed and flipped, so it is escalating rather than ruling - receipt f7fdb9f25.
In plain terms. Do you want this board switched back on with me named as its driver? The system decided nobody was driving it and switched it off, after mistakenly shutting down five drivers in a row - but I am still here and the crew of ten is working. Switching it back on is also what got the last five killed, so I would rather leave it off and fix the faulty alive-list first.
The session raising this recommends: Leave it off - fix the alive-list first Being parked costs nothing today: the ten helpers are working, their pull requests land normally, and my own wake loop was deliberately left running. Registering myself as the driver is what got the last five shut down - the list that reports who is alive cannot see this kind of session, so within two minutes it reads as dead again. The honest repair is the pull request that fixes that list, not another round of the same thing.
358e82a4 (adopted from b84e2f74) is parked on this.
2Your go-ahead on propflowai#8006 arrived and this session could NOT carry it out: every attempt to land a PR from here is refused by the Claude Code auto-mode classifier ('Merge Without Review' / 'Auto-Mode Bypass'), and the same refusal has since spread to read-only commands in this session (the board's own counter, some helper dispatches). Meanwhile eleven more finished, reviewer-approved PRs from this run are queued behind the same wall, and the board tool refuses to record a row until its PR has landed — so the board cannot move past 153 of 291 no matter how much work the crew does. THE QUEUE, in the one order that is safe: propflowai#8097 BEFORE #8035 (a lane test-landed them and proved the second one reddens main if it goes first); then propflowai#8006 (you already said yes), #7997, #8053, #8056 (this one is held for Fede - auth paths), #8143, #8149; agent-smith#483, #504, #507; local-bin#168 (and close #159 as superseded); propflow-docs#172. Do you want a permission rule added so this driver can carry out answers like yours itself, or will you take the queue by hand?
In plain terms. Can I get permission to act on your answers myself? You said yes to one fix and I could not carry it out - this session is blocked from doing it, and eleven more finished fixes are stuck behind the same wall. Either you give me the permission, or every one of them needs you personally.
The session raising this recommends: Add the permission rule - let the driver carry out its own answers You answered a question and nothing happened, which is the exact failure this role exists to prevent. Without the rule every one of the twelve finished fixes needs you personally, and the crew can keep producing them faster than you can clear them - the board has 132 rows still open. The safety the refusal protects is already provided twice over on these PRs: each one is green and carries a reviewer's formal approval at its current commit, and the riskiest class (anything that changes the automation itself, or touches login security) stays held for a person regardless of any rule you add. Tried first, unsuccessfully: No model rung was asked about this one and none should be: it is a permission grant on this machine, which is yours alone. The related policy question WAS put to the decider (receipt fa7359425, RESOLVED): the two repositories with no reviewer should get one so their work can land itself, and a lane is building that now. That ruling explicitly leaves this class - anything that changes the automation itself - as a named exception a person lands by hand, which is why the queue above still exists either way.
358e82a4 (adopted from b84e2f74) is parked on this.
3The two repos with no reviewer now each have a review gate waiting: gera-propflow/local-bin#175 and PropFlow-Technologies/propflow-docs#173. The gate was exercised live and proved it says NO when it should - local-bin's first-ever workflow run came back red with 'No reviewer credential is configured', which is the gate working. That red is the problem: local-bin is on a personal account, so no organisation credential reaches it, and until CLAUDE_CODE_OAUTH_TOKEN exists as a repo secret there, EVERY pull request in that repo reads red. So the order matters and getting it backwards makes things worse than today: that repo currently has no gate at all and its 11 open PRs can be landed by hand, but with the gate in and no credential, nothing there lands at all. propflow-docs already has the credential through org secrets, so its gate is safe to land on its own. Setting a repo secret is yours alone - I did not distribute the credential.
In plain terms. Can you paste a credential into one repository? The two repos whose finished work had nobody to approve it now have a reviewer each, but one of them is on a personal account and cannot see the company credential - so until you add it there, everything in that repo reads as failing and nothing lands at all. The order matters: credential first, then the two gates.
The session raising this recommends: Add the credential to local-bin, then land both gates local-bin is where most of the fleet's own tooling lands, and 11 finished PRs are queued there today purely because nothing can approve them. With the credential in place that repo starts clearing its own queue, which is the whole point of the ruling. Holding the gate is not neutral either - it leaves the single highest-blast-radius checkout on the machine with no reviewer at all, which is the condition that let a safety test ship while only passing when the thing it tested was switched off, found earlier today. The gate cannot be made required on that repo's plan, so it stays advisory, which means you are not handing it the power to block you. Tried first, unsuccessfully: The policy half was already put to the decider and RESOLVED (receipt fa7359425): give the two reviewer-less repos a reviewer so their work can land itself, which is what these two PRs do. This block is not that question again - it is the one step the ruling cannot perform, because a repository secret is a credential and only you can place it. The decider's own cost note predicted exactly this: it warned the wiring PR would need an attended human and that a miswired gate would be worse than none, which is why the lane proved the gate reddening live before reporting.
358e82a4 (adopted from b84e2f74) is parked on this.