The Merges Tab

A merge with no agent id is sorted by cause, never pooled into one “unattributed” bucket

The brief said the agent → PR join is broken, so the tab must not silently drop what it cannot attribute. Measuring it changed the answer, and the change is worth stating plainly rather than designing around a defect that is not there.

The join through git history works. Across the last 30 merges to obot.agent main, 26 carry a Worker: W00xx trailer in the merge commit, and among merges made by a worker the rate is 26 of 26, unbroken since 2026-08-18. It works because an agent types the trailer into the commit, so nothing about it depends on an environment variable surviving a daemon respawn.

The four that carry no trailer are the ranked-head re-ranks, made by prime and the Navigator. Those actors hold no worker id by design — tools/worker-id says so in its own docstring — so they have nothing to stamp. Filing them as unattributed would invent a defect and, worse, would cite obot.agent#276 as its cause. #276 is real, but it is about a different join: OBOT_WORKER_ID is never set, so ledger writes record a session id. That breaks worker → blocker and worker → journal. It does not touch the commit trailer.

So the tab carries three non-worker lanes instead of one:

Lead session — holds no W-id

prime, the Navigator, the admiral. Expected, permanent, and labelled as such. Four rows tonight.

Not an agent

A human merge. One row tonight, and it is yours.

Attribution missing — the #276 lane

A merge from a session that does hold a worker id, whose commit carries no trailer. This is the only lane that cites #276, and the only one that means something is wrong. It reads zero tonight, and a zero here is a real fact rather than an empty bucket.

The header line never presents attributed counts as a total: it reads 20 merges — 15 by 11 agents, 4 by the lead session, 1 by you, and those five numbers add up in front of you.

A day is midnight to midnight in America/New York, and the page says so on its face

You ask what happened today while sitting in that timezone, so that is what the day means. It is not a cosmetic choice. Three of tonight's 17 obot.agent merges — 20:06, 20:10 and 20:12 local — carry a UTC date of 2026-08-21. A UTC day would have shown you 14 and been quietly, confidently wrong by the three most recent things that happened.

That is not hypothetical: the existing Navigator panels bucket by UTC, using new Date(t).toISOString().slice(0,10). So the tab has an obligation beyond stating its own rule — it has to state the disagreement, because the same merge will sit on a different day one tab over. The header line carries both: the local day it is showing, and a note that three of tonight's merges appear on the Navigator tab under tomorrow's date.

Two mechanics follow from the call. The zone is pinned as the IANA name through Intl.DateTimeFormat rather than a hardcoded offset, so 2026-11-01 does not silently shift every boundary by an hour. And the header renders from the same value the bucketer used, so the line cannot describe a rule the code is not applying.

It reports. It never asks you for anything.

You have three buckets that can demand action — release candidates, decisions, config — and a fourth would cost more than it returns. Nothing on this tab is a checkbox, an approve button, or a row that stays red until you clear it. Every link is a drill-down to a pull request or an issue; none of them is a “do”.

That has a consequence worth accepting on purpose. The tab will surface things that look like they need doing — 98 pull requests waiting on a harness release, a product repo that has not moved in three days, a merge in a repo no lane governs. It does not adopt them. Each such row cites the queue item or requirement that already owns it; a row with nothing to cite says so, and an unowned row is a filing gap for the Navigator to pick up, not a button here.

One honest qualification, because “read-only” will otherwise read as a safety property it does not have. The dashboard writes a local last-look record on every page fetch. That is unchanged for this tab, and it means the drill-down must vary by query string rather than by path — per-agent and per-day URLs as new path segments would generate unbounded surface keys and evict real surfaces from a 64-entry record. Read-only here means no GitHub writes and nothing entering your queue.

The tab counts every merge in the workspace, not only the governed ones

This is the call the brief did not ask for, and the one that decides whether the summary is trustworthy. The obvious implementation scopes to the seven repos in policy.json. Tonight that would have printed 19 and been wrong by exactly the merge that most deserved your attention: a merge into jwildfire.github.io, a repo absent from the policy file, which obot-merge would have refused outright.

Scoping to the governed set institutionalises that blind spot. Scoping to “repos cloned on this machine” drifts as clones come and go. So the tab states its denominator out loud — every repo in the jwildfire org — and renders an ungoverned merge as its own state, in its own colour, rather than as an omission.

Three related edges get the same treatment, because each has a shape where a zero would be a lie:

obot.roadmap declares no release branch

Its integration-versus-release delta is undefined, not zero. The cell reads “nothing to wait for”, because the hub publishes on merge.

demo-301's site branch shares no ancestor with main

The comparison returns a 404, permanently. That is neither an error to log nor a zero to print; the cell says the two branches are not comparable.

Zero merges and unread source must never render alike

A quiet Sunday and a dead collector both produce an empty page. Only one of them is news. The house rule already covers it — absence is believable only from a genuine not-found — and this tab inherits it.