| name | board-cleanup |
| description | Map a Fizzy board into a clear state-of-play, surface the real chaos (floating cards, stale artifacts, duplicate/overloaded work, dependency chains, scope-stale epics), then triage and split with the user's sign-off. Use when the user wants to "get back up to speed", "map out the board", "clean up Fizzy", reduce board chaos, or plan who-owns-what for a sprint/weekend. Repo-agnostic — works on any repo with a corresponding Fizzy board. |
| version | 1.1.0 |
| license | MIT |
| metadata | {"author":"VettedAI","category":"engineering-management"} |
Board Cleanup
Turn a chaotic Fizzy board into a clear map + an executable plan. This is the companion to the fizzy skill (which is the low-level API client) — this skill is the workflow on top of it.
The job is five phases: Pull → Map → Flag → Triage → Scope-and-split. Phases 1–3 are read-only. Phases 4–5 mutate the board — never mutate without explicit user sign-off, and prefer to do the safe parts (close artifacts, move floaters) before the judgment parts (owners, splits).
The one rule that earns its keep
Verify against the live tree, never the card. Cards go stale the moment they're written. Before you act on any claim ("17 fns left to clean", "table missing from staging", "not started"), check ground truth: the actual git branches, the deployed code, to_regclass/information_schema, the real file. In practice this routinely shrinks "epics" to near-done — and occasionally reveals the prod side is cleaner than the repo. This single check is what makes the plan trustworthy.
Apply it hardest in Phase 5, but keep it in mind throughout.
Context budget — work from a lean index, not raw card bodies
Phase 1 already files the raw pull to /tmp/bc_cards.jsonl — keep it that way: the parent context should hold a lean index (number | title | column | assignees | tags), never the full description_html of every card. A big board × wordy descriptions read into the parent fills the window. So:
- Map/Flag (Phases 2–3) run off the lean index. Read a card's
description_html only for the specific cards a check needs (a suspected dependency chain, a candidate epic) — not for all of them, and not preemptively.
- Scope-and-split (Phase 5) is the expensive part: pulling each epic's full body + its git/deploy/catalog ground truth. For a board with several flagged epics, fan out one subagent per epic — each reads its own body + does its own verify-against-live-tree and returns a compact state (
{epic, workstreams[], done[], open[], needs_investigation[]}). The parent never holds the raw epic bodies or the git/deploy output — only the returned states. Heavy text in the parent is billed every turn and can't be evicted; that is the cost you are controlling.
Setup
ENVFILE=$(ls congrats/.env.local .env.local 2>/dev/null | head -1)
TOKEN=$(grep -E '^FIZZY_API_TOKEN=' "$ENVFILE" | head -1 | cut -d= -f2- | tr -d '"' | tr -d "'")
H=( -H "Authorization: Bearer $TOKEN" -H "User-Agent: VettedAI/1.0" )
Pick the board for this repo. Boards (account 6102589):
| Board | ID | Repo it maps to |
|---|
| Congrats (Candidate-Facing) | Engineering | 03f58rc5c48jorujpxqp5da5b | vetted-congrats |
| Vetted (Recruiter-Facing) | Engineering | 03faozjl3gdngcoyzpkr4vf87 | vettedai-audition |
| Bugs 🐛 | 03fl735hqcd0h1pettl8o94oo | cross-repo |
| Feature Grooming | Product Team | 03feaz5rc2t60wkn2rvjkhy6b | product |
If the repo isn't obvious, ask the user which board.
Phase 1 — Pull (read-only)
BOARD="<board-id>"
curl -s "https://app.fizzy.do/6102589/boards/$BOARD/columns.json" "${H[@]}"
Pull every card, paginated. CRITICAL GOTCHA: cards.json?board_id=X requires the .json suffix AND the board_id filter is silently ignored — the endpoint returns cards from all boards. You MUST filter client-side by EXACT board.name equality — never startswith: sibling boards share a prefix (e.g. Vetted (Recruiter-Facing) | Engineering, Vetted Onboarding Tool, and VettedAI GTM all start with Vetted, so startswith('Vetted') silently conflates all three and inflates every count). Page size escalates (15→30→50…), so paginate by following Link: rel="next" — never stop on the first <15 page. (Full rules: "Reading a Board — AUTHORITATIVE" in /fizzy.)
url="https://app.fizzy.do/6102589/cards.json?page=1"; > /tmp/bc_cards.jsonl
while [ -n "$url" ]; do
body=$(curl -s "$url" "${H[@]}" -D /tmp/bc_h.txt)
echo "$body" | python3 -c "import sys,json;[print(json.dumps(c)) for c in json.load(sys.stdin)]" >> /tmp/bc_cards.jsonl
url=$(grep -i '^link:' /tmp/bc_h.txt | sed -n 's/.*<\([^>]*\)>; *rel="next".*/\1/p')
done
Useful fields per card: number, title, closed, postponed, column (null = floating/"Maybe" pile), assignees[].name, tags, description_html. Single-card detail needs .json too: /cards/{n}.json.
Blind spots — the walk only returns OPEN, non-postponed cards. The Not Now (postponed) and Done (closed) lifecycle lanes are NOT returned by any list endpoint, so your map covers every active/floating card but is blind to those two — only the Fizzy UI shows their counts (often large: e.g. Congrats Not Now = 38, Done = 99+). Say so explicitly in the map rather than implying the board is empty there. The not c['closed'] filter above is just belt-and-suspenders; closed cards never appear anyway. Also note boards auto-postpone idle cards (auto_postpone_period_in_days, ~30d) into Not Now — a long-idle card that vanished from a column was likely postponed, not shipped.
Phase 2 — Map
Group open cards by column in pipeline order. Produce a tight table (column → count → card #s). Then re-cluster the same cards by theme (security, a refactor chain, a feature epic, etc.) — themes cut across columns and are where the real story lives.
Phase 3 — Flag the chaos
Hunt for these specific smells (don't just say "it's messy"):
- Floating cards —
column == null. The untriaged pile. List them; they're invisible work.
- Stale artifacts — auto-generated cards that aren't work (e.g.
Release Notes — <date> from /generating-release-notes). Candidates to close. Watch for duplicates (same date / superset).
- Overloaded owner — one assignee holding many open PRs, especially a sequential dependency chain all open at once (nothing merges → everything stalls).
- Dependency chains — cards that must land in order (read
description_html to confirm: "depends on #X", "blocks #Y", "paired with").
- Scope-stale epics — one card hiding 3+ workstreams across repos. Flag for Phase 5 splitting.
- Cross-repo / paired cards — work that has a sibling on another board.
Phase 4 — Triage (MUTATES — get sign-off)
Present the plan, then execute the safe parts first:
- Close artifacts —
POST /cards/{n}/closure.json (note: closure.json, returns 204; plain closure 401s).
- Move floaters into priority columns —
POST /cards/{n}/triage.json with {"column_id":"..."} (204). Respect the priority emoji already in the title/body (🔴/🟠/🟡); cards marked 🟢 with no green column go to the bottom of 🟡. Read the card body when the title has no emoji.
- Leave owner assignment and column promotion of real work to the user — those are judgment calls, not hygiene.
Record decisions as comments so the board is self-documenting (see Phase 5 examples). Don't spam — one comment per real decision.
Phase 5 — Scope-and-split (MUTATES — get sign-off)
For any epic flagged in Phase 3:
- Scope against reality (the rule above). Pull each workstream's actual state — git branches, deployed code, the live catalog. Mark what's already done, what's confirmed open, what needs investigation. This is the highest-value step; it routinely halves the work. Fan this out per epic (see Context budget) — one subagent per epic does its own verify-against-live-tree and returns a compact state; don't read every epic's body + git/deploy output into the parent.
- Split the epic into child cards, one per coherent workstream / repo, each carrying enough to execute: file anchors with line numbers, the exact fns/files, constraints (integration branch, "don't clobber prod-only code"), and a crisp "Done when". Mirror the
grooming-architect style (file anchors + logic constraints + verification).
- Link children back on the parent via a comment; turn the parent into an umbrella. Keep the priority emoji in child titles (team convention).
Fizzy write cheat-sheet (the gotchas that cost real time)
| Action | Endpoint / payload | Notes |
|---|
| Create card | POST /boards/{BOARD}/cards · {"card":{"title":"...","description":"<html>"}} | Card body field is description, must be HTML. Number comes back in the Location header. |
| Comment | POST /cards/{n}/comments · {"comment":{"body":"<html>"}} | Comment field is body (HTML). |
| Move column | POST /cards/{n}/triage.json · {"column_id":"..."} | 204. |
| Close | POST /cards/{n}/closure.json | 204. .json required. |
| Assign (toggle) | POST /cards/{n}/assignments.json · {"assignee_id":"..."} | Same id again unassigns. |
Build JSON payloads with python3 -c "import json,sys;print(json.dumps(...))" — never hand-interpolate HTML into -d (quoting will bite you).
Output to the user
Lead with the board-as-pipeline table + the theme clusters, then the specific chaos (named cards), then a proposed plan: who owns what, which priorities depend on what, what to close/split. Be definite. Recommend a starting point. Then ask for sign-off before any mutation.
Communication: plain English, gloss every piece of jargon the first time, concrete-before-abstract. The user may be non-engineer — the value is clarity, not completeness.
Related
fizzy skill — the low-level API client this builds on.grooming-architect skill — the style for child-card descriptions in Phase 5.merge-to-prod / qa-handoff skills — natural follow-ons once the board is mapped.
