board-taxonomy

name	board-taxonomy
description	How 33GOD models ticket state on a Plane board — which facts belong in a state, which in a label, and which axis a new label joins. Use when adding or renaming a Plane label, deciding between a label and a state, wiring automation that writes to a board, scaffolding a new project board, or reconciling boards that have drifted apart.

33GOD Board Taxonomy

A Plane board carries several different kinds of fact, and Plane offers only two carriers for them: states and labels. Left alone, everything becomes a label. Measured across the 25 fleet-bound boards on 2026-09-04: 75 distinct label names, 72 of them on exactly one board, and bug in three different colours. See references/field-survey.md.

That happened because nothing was wrong with any single decision. Labels are authored at runtime by whichever agent needs one first — the grooming prompt says "add lifecycle:triaged", the agent finds it missing, and the label MCP tool creates it. Same prompt, same model, different name every time. The fix is not discipline; it is a declared, closed set.

The one test

Do not classify a label by what it means. Classify it by who writes it and what clears it. That single question resolves every case below.

kind	cleared by	carrier	exclusive
Position — where it is in the pipeline	a human or agent moving it	Plane state	yes, structurally
Latch — automation recorded a fact	the next pipeline stage	lifecycle: label	one per axis
Lease — a process is holding it right now	the holder, or a sweeper	agent: label	one per axis
Facet — a durable property of the ticket	someone deciding it is false	xp:, risk:, needs:, bare words	no

Latch vs lease — the distinction that matters

Ask: if the process holding this dies, does the label become a lie?

• agent:working — yes. It is a lease, not a state. It must self-clear, and it must be safe to force-clear, because the thing it describes is a running turn that can vanish.
• lifecycle:triaged — no. Grooming really did finish; that stays true forever. It is a latch.

This is operational, not cosmetic. Leases need a TTL and a sweeper. Latches need neither. Getting it wrong leaves permanent lies on the board.

The axes

Closed set. Anything with a colon must appear here.

axis	kind	writer	values
lifecycle:	latch, exclusive	automation only	triaged, needs-input, blocked
agent:	lease, exclusive	automation only	working, failed
xp:	facet, exclusive	either	internal, external
risk: / needs:	facet, multi	either	open within the axis

Everything else is a bare word, descriptive, per-project, and nobody's business but that board's: bug, docs, spike, security.

Rules

1. Position never goes in a label. You have a 9-state machine; use it. A label that answers "where is this in the pipeline" is a bug. This is what rules out phase:0/1/2, scope:proposed, blocked-by:M0.
2. One writer per exclusive axis. Plane cannot enforce exclusivity — GitLab has scoped labels (key::value replaces within key), Plane has nothing — so the writer enforces it. Corollary: if a human has to hand-set it, it belongs in the state machine instead.
3. A colon means machine-readable and closed. Bare words are descriptive and open. This is checkable: any colon label not in the table above is either a mistake or an axis someone forgot to declare.
4. Colour encodes the axis, not the value — one family per axis, one shade per value within it. Not hypothetical tidiness: by the time the set was closed, lifecycle:triaged had already reached two colours (teal on JIMB, amber on 33GOD), exactly the way bug reached three.
5. Cap the functional set at ~8. The cost was never filtering — 75 labels filter fine. It is that every agent re-invents a name it cannot look up. Eight is memorizable. Descriptive labels can grow freely per project.

Where this is declared

A standards document rots exactly the way those 75 labels did. The taxonomy is real only where it is (a) in a file the agent already reads and (b) provisioned by a tool. Both seams exist:

• Declaration — .project.json, activity_report.board.exposure_labels. Already on groomingPrompt's read path. Today only james-brennan declares it, and only for the xp: axis.
• Provision — bloodbank/bin/bb-board-scaffold, FUNCTIONAL_LABELS.

Keep those two in agreement. When they disagree, the declaration wins and the scaffolder is wrong.

Scaffolding a new board

Plane exposes no template API on this instance — /templates/, /project-templates/ and /workitem-templates/ all 404 — and no issue types (404), which is why classification has to be labels at all. States are fully writable over v1.

bb-board-scaffold                          # dry run over every fleet board
bb-board-scaffold --to DELO,DNET --apply
bb-board-scaffold --from JIMB --all --apply

It clones the reference board's state machine and provisions the functional labels.

States it only adds and aligns:

• never deletes a state, and never reorders one it did not create;
• never touches default or is_triage — a board must have exactly one default state, and moving it silently re-homes every new ticket;
• refuses a board reporting zero states, which is the signature of a soft-deleted project, not an empty one.

Labels it adds, and recolours where they have drifted. Colour is cosmetic — unlike a state's group it can be corrected without moving anything — so this is what actually keeps one name to one colour across 25 boards. Any colon label outside the closed set is reported as ? undeclared colon labels and never touched: deleting a label strips it from every ticket carrying it, and that does not come back.

33GOD and JIMB hold byte-identical 9-state machines and either is a sound reference. Note that --all excludes the reference board itself, so its own labels go unreconciled; follow with bb-board-scaffold --to <ref> --apply.

The ack chip, as a worked example

agent:working exists because a ticket sat visibly untouched for the entire time its agent was booting — 377s, 633s and 2596s measured on three JIMB tickets. No instruction in an agent's prompt can close that gap, because the gap is the agent starting up. The acknowledgement has to come from the lane, at dispatch.

Both ticket lanes stamp the chip the moment the Fleet node reports invoked: true, and clear it when that agent's turn ends. Two constraints shaped it, and both generalise to anything that writes labels:

• Plane has no per-issue label sub-resource. Every label write is a full replacement of the array, so a writer must re-read the issue first. Trusting the webhook payload instead is how you silently delete the agent's own lifecycle:triaged.
• A bloodbank.agent.invocation.* event carries no board and no ticket — only a correlation id, and that id is inherited, not the recomputable ticket uuid. So the dispatch side records correlationid -> ticket and the turn-end side spends it. Store and spend; never try to invert.

Measured: chip on in 339ms of node time, off 271s later, leaving the three labels the agent had added itself untouched.

Known gaps

• The lease has no board-side sweeper. The chip's expiry lives in n8n workflow static data, so if n8n restarts mid-turn a chip can strand with nothing to clear it. A --sweep on bb-ack-labels is the fix.
• .project.json declares only the xp: axis, in one repo. The other three axes are baked into the scaffolder instead of declared.
• groomingPrompt still says "using the label names the board already uses" — which, on a board with no labels, tells the agent to invent. It should read the declared axes.
• CANDY and CANDYS accept nothing — zero states, and a label create 409s as a duplicate while GET reports count 0. They are soft-deleted, not empty.
• Thirteen undeclared colon labels are still live, reported by the scaffolder's lint and left alone because deleting one strips it from its tickets. JIMB carries phase:0/1/2, scope:proposed|gated|exploratory| completed, sprint:shadow-foundation, violation:quiet|amber-attention; DECK carries blocked-by:M0, needs:hardware, risk:unverified; INFR carries project:DeLoContainers, stack:media. phase:, scope: and blocked-by: are Rule 1 violations — position wearing a label — and want migrating into the state machine. needs: and risk: are declared open axes and are fine. The rest are undeclared axes someone should either declare or demote to bare words.