board
Use when the user runs /board, /board add <KEY> [--active], /board remove <KEY>, /board refresh [--stability-scrape], /board attention <mrs|prs|jira> [--clear], /board stability-write <label> --from-j
SKILL.mdskills/board/SKILL.md
GitHubBoard
The situation board — one standing window onto the work: the tickets in progress
and the app's growth, gathered where a glance can hold them. Each source is a
channel — a JSON file under ~/.skadi/board/. The page polls the folder every
few seconds; a channel added, changed, or removed shows on its own, no rebuild.
All work routes through one entry, ~/.claude/hooks/board.sh, which orchestrates
focused helpers (the ticket writer, the growth writer, the shared manifest, the
page). Run the verb the user named.
Verbs
/board — serve
~/.claude/hooks/board.sh serve
Lays the page and theme into ~/.skadi/board/, boots a background server (or
reuses the one the lockfile names, if it still answers), and prints the URL.
Surface the URL inline; tell the user the tiles follow the channel files.
/board add <KEY> [--active]
~/.claude/hooks/board.sh add PSG-4478 --active
Fetches the Jira ticket and its subtasks, derives the AC rate from subtask
completion (a subtask counts met when its status is a done category or a
name listed in ~/.skadi/board/ac-done-statuses.json — the team's done-enough
set), and writes ticket-<KEY>.json. --active makes it the hero and clears the
flag on every other ticket, so exactly one hero stands. Re-running refreshes it.
Tracker routing. Before dispatching, resolve the tracker from the key's
project prefix via the tracker_routing memory (e.g. MET → youtrack,
PSG → jira), the same map /council and /glorfindel use. Default to
jira when the prefix is unmapped. Pass it through:
~/.claude/hooks/board.sh add MET-1 --active --tracker youtrack
board.sh add forwards all arguments to the ticket writer, so --tracker
reaches it unchanged. On refresh, the tracker is re-derived from each
channel's recorded source — no memory lookup needed.
/board remove <KEY>
~/.claude/hooks/board.sh remove PSG-4478
Drops the ticket channel and regenerates the manifest. The tile vanishes on the next poll.
/board refresh [--stability-scrape]
~/.claude/hooks/board.sh refresh
~/.claude/hooks/board.sh refresh --stability-scrape
Re-fetches every ticket already on the board (preserving which one is active),
the metis growth numbers, and the three Attention surfaces (mrs, prs,
jira — see below; mail is not a hook path, it stays out of refresh).
Growth is best-effort — a BigQuery hiccup skips it with a note rather than
failing the sweep. Attention never skips: a fetch failure there is itself a
successful write (see below), so refresh only prints "(skipped)" for an
Attention surface when the write failed, never when glab/gh/Jira did.
Plain refresh never touches Stability — the live dropdown already covers any
app BigQuery can answer, in real time, with nothing to gain from pre-caching it.
--stability-scrape is the explicit ask to go further: it sweeps every bound
app's crash-free number and prints which ones BigQuery couldn't answer (no GA4
pairing, or the query itself failed). That printed list is yours to act on next
— see The Stability tile below for what to do with it. The hook stops at
naming the gap; it cannot drive Chrome itself.
/board attention <mrs|prs|jira> [--clear]
~/.claude/hooks/board.sh attention mrs
~/.claude/hooks/board.sh attention prs --clear
Refreshes one Attention surface directly — the same three surfaces refresh
already sweeps, callable on their own when only one needs a poke. --clear
unlinks that surface's channel and regenerates the manifest instead of
fetching — the row simply drops off the band, rather than showing a stale
count wearing no age label. See The Attention band below for what the band
shows and why a failed fetch is still a successful write.
/board list
~/.claude/hooks/board.sh list
Prints the channels — tickets with status and AC (the active one starred), then the growth line.
The growth channel
/board refresh pulls growth alongside the tickets. To refresh growth alone,
~/.claude/hooks/board.sh growth runs the /growth hook, parses its headline
(WAU/MAU), copies the rendered dashboard in as the tile's Enter door, and writes
growth.json. Only metis has a GA4 → BigQuery export today.
The Attention band
Sits above the ticket bento — a band answering "what awaits me" outside the
curated ticket set. Each row reads <label> · <count> · <detail> · Enter →,
one row per surface that has ever written a channel; a surface never written
(or --cleared) simply has no row — no ghost tile, unlike the strip's
"channel not wired" convention. Row order is fixed (mrs, prs, jira, mail),
not the manifest's mtime order, so rows don't shuffle on every refresh.
mrs— open GitLab MRs you're reviewing or authored (mrs-check.sh).prs— open GitHub PRs you're reviewing or authored (prs-check.sh).jira— tickets assigned to you that moved in the last day (assignee = currentUser() AND updated >= -1d, all statuses — this reports movement, not outstanding work, so a ticket that reached Done today is exactly what belongs here).mail— unread Outlook mail in the last 24h. Not written byboard.sh attentionat all — see below.
A failed fetch is a successful write. A dead glab/gh token, an
unreachable Jira, no network — none of these skip the write. The channel still
gets written, with count: null (rendered as —, never 0) and
verdict: "unknown", and the writer still exits 0. This is deliberate: a
count: 0 must always mean "genuinely nothing awaits", never "couldn't look".
board.sh refresh's || echo "… (skipped)" guard is only ever tripped by a
write that could not happen at all (bad args, an unwritable board dir) — never
by the fetch itself failing.
Mail — the one surface board.sh cannot fetch
Mail has no hook path. The Outlook connector /triage uses
(mcp__claude_ai_Microsoft_365__outlook_email_search, or search outlook email via ToolSearch if the name differs) needs a model in the loop to call
it — a bash hook cannot. So there is no board.sh attention mail fetch verb,
and refresh never touches it.
To populate the mail row: search unread inbox mail from the last 24 hours the
same way /triage's step 1 does, write the count and a short detail to a
scratch JSON file ({"count": 3, "detail": "2 from humans · 1 flagged", "url": "https://outlook.office.com/mail/"}), then persist it the same way
stability-write persists a console scrape:
~/.claude/hooks/board-attention.sh mail --from-json /tmp/mail.json
If the connector is not connected, or its search errors, do not write a
channel at all — run ~/.claude/hooks/board-attention.sh mail --clear
instead (idempotent; a no-op if nothing was ever written). The mail row must
never show a ghost or a stale count from a connector that's gone dark; the
row is either accurate or absent, never guessed at.
The Stability tile
Sits in the top strip, after Metis growth. Unlike every other tile it carries no
channel file — its dropdown and number are fetched live from the browser via two
routes board-server.py adds:
GET /stability/apps — the app roster, for the dropdown
GET /stability/fetch?label=… — one app's crash-free users %, on selection
Both shell out to ~/.claude/hooks/board-stability.py, which builds the roster by
gluing two sources: every crash_routing.md any repo has ever bound via /beleg
(globbed across ~/.claude, ~/.claude-personal, ~/.claude-work), layered with
hand-added or hand-corrected entries from ~/.skadi/board/stability-apps.json —
the same "small local JSON you maintain by hand" pattern as
ac-done-statuses.json. No /board verb manages it — bind an app the normal
way, by running /beleg in its repo.
The percentage itself is a join: crashed users from the Crashlytics BigQuery
export, active users from GA4's. A crash_routing.md row now carries two optional
trailing fields — GA4 project · GA4 dataset — after the six /beleg already
writes (repo · flavor · Firebase project · bundle · platform · account); an app
bound without them shows in the dropdown but reads "GA4 not configured for this
app" instead of a percentage, rather than guessing at a denominator it doesn't
have.
Fetched on pick, not on the poll loop. The roster loads once per page load; selecting an app fires one live query. Neither auto-refreshes — a live BigQuery join is not something to run every 8 seconds, and a newly-bound app needs a page refresh to appear.
The console-scrape fallback
When BigQuery can't answer for an app — no GA4 pairing, or the query itself
failed — refresh --stability-scrape names it rather than leaving the tile to
show "GA4 not configured" forever. For each app it lists:
Read the printed
project/bundle/platform/accountand follow/beleg's own console-scrape flow (skills/beleg/SKILL.md§3) — the same navigation, the same account-index care — but read the dashboard's "Crash-free users %" headline directly, rather than the issue list/belegranks. No BigQuery export is needed for this number; the console computes it internally.Write what was found — a percentage, or
nullwith anoteif the console itself showed nothing usable — to a scratch JSON file:{"crash_free_pct": 93.1, "note": null}.Persist it:
~/.claude/hooks/board.sh stability-write "vitallink-ca · jp · ANDROID" --from-json /tmp/scrape.jsonThis writes
stability-<slug>.json(the label, lowercased and dashed) as an ordinary pulled channel — same auto-pickup as every other tile, no live endpoint involved.
The live dropdown reads this channel as its fallback whenever a BigQuery fetch
comes back with no percentage: it shows the pulled number in its place, no
label marking it as scraped or stating its age — a decided simplification
(docs/plans/board-stability-console-fallback.md), not an oversight. If both
BigQuery and the fallback channel have nothing, the tile still shows the
BigQuery note plainly.
This step needs a live, Chrome-connected model turn. --stability-scrape
only ever names the gap — a hook cannot drive Chrome, so an unattended
refresh --stability-scrape (cron, /loop, a background sweep) will print the
needs-scrape list into a log no one reads and go no further. Run it
interactively, or don't pass the flag.
Notes
- Auto-pickup, not auto-fetch. The page auto-shows any channel file change
within a poll (~8 s). It does not poll Jira or BigQuery — fresh numbers are a
pull: run
add/refresh. A scheduledrefresh(via/loopor cron) closes that gap when wanted. The Stability tile is a further exception — see above. - One server, one fixed port.
servebinds port 10000 (override withBOARD_PORT) and reuses it whenever it still answers, rather than multiplying servers — the URL never drifts across restarts. The handbook rides the same server:board-server.pyroutes/handbook/and/previews/toBOARD_SKADI_ROOT— the skadi repo, whichboard.shresolves from its own parent, or, running as the copy under a config root, from the pathinstall.shrecords at~/.skadi/install/skadi-root. So./handbook.shand the board's own "Handbook ↗" link both resolve through this one port, whichever copy booted the server. - Read-only. The board reads Jira, BigQuery, and the Crashlytics/GA4 exports; it never writes to a tracker.
- Tests ride beside the hooks —
board-ticket.test.sh,board-growth.test.sh,board-manifest.test.sh,board-server.test.sh,board-stability-write.test.sh,board-attention.test.sh, andboard-index.test.sh's Attention band cases run offline via injected fixtures;test_board_stability.pycovers board-stability.py's own logic the same waytest_beleg_crashes.pycovers beleg's. The console scrape and the mail connector search both stay unverifiable offline, same caveatdocs/plans/beleg-crash-analysis.mdalready carries for beleg's own fallback.