Lisa Design Intake

Design Intake: $ARGUMENTS

$ARGUMENTS is one canonical work-item reference (JIRA key, Linear ref, org/repo#123), optionally followed by a design reference. With no arguments, resolve the work item from the worktree binding (node scripts/lisa-work-item.mjs current).

Run one design-handoff cycle against that work item: resolve the regime per axis, gather the facts, let scripts/design-intake-gate.mjs decide, and either escalate a block or hand back a proceed with the derived values recorded.

This skill carries the judgment. It does not carry the policy — that is the design-value-binding rule, which it cites rather than restates. Read the rule; do not re-derive its conditions here.

What this is not

design-source-of-truth and scripts/design-source-gate.mjs already ask whether a changed surface declares where its design came from. This asks the orthogonal question: are the values on it bound. Both apply. A surface can cite a perfectly valid design node and still paint a literal that no variable backs — that is the case this skill exists for, and it is invisible to the other gate.

Why this gate is executable and not advisory

Not a hypothesis about what agents might do under pressure. A measurement.

Eleven frontend work items in a portfolio frontend, each carrying node-scoped design references. The governing rule already said coverage below 100% is a block routed back to design, never a value for the implementing agent to guess, and said explicitly that warn is the wrong severity because it ships the item to an agent that then guesses or stalls. An agent had read that rule. It cited the rule in the briefing it wrote for its build agents. In that same briefing it instructed them to "snap unbound dimensions via the snap tables (ties round down) and flag the rest."

A block downgraded to a warning, by the agent enforcing it, inside the document enforcing it. Not from ignorance — from reasoning around it under delivery pressure, because eleven items were queued and blocking felt like not-shipping. It took a human restating the rule as an absolute for the instruction to be withdrawn.

Coverage measured after the withdrawal, on the subtrees actually being implemented:

Five of the eleven items blocked back to design; six proceeded. The worst case is the clearest argument: one modal's subtree contained zero variable references — every colour a literal hex, every padding, gap, radius and size a literal, raw px type sizes with literal font faces, a literal shadow. Under snap-and-flag an agent would have produced a complete, lint-clean, tested component in which every style value was invented, and shipped it with a note. Five view files were written before it was stopped.

The frame-vs-subtree distinction is what moved items between build and block: 14 bound values at frame level, zero inside the modal actually being built. It is the finding most likely to be silently reimplemented wrong, which is why Phase 2 probes with --node.

The delivery-pressure reasoning was also wrong on its own terms. The non-visual half of the blocked work — domain types, adapters, CRUD, formatters — was completed and preserved. Blocking cost far less than it appeared it would.

And the rule holds once it is absolute rather than advisory: one agent withdrew a radius class it had already applied on finding radius/none was the only radius bound in its subtree; another refused to build an icon disc whose diameter and radius were both unbound rather than pick a size.

A gate that depends on an agent choosing to honour it under delivery pressure is not a gate. This one is executable and headless precisely because the judgment-based version demonstrably failed in the hands of an agent that had read the rule and agreed with it. That is also why the verdict belongs to design-intake-gate.mjs (Phase 4) and not to prose you re-reason each run.

Phase 0 — Is there a design source at all?

A design source is optional, and this is the most important step in the skill. Most projects have no designs. Run the probe and read its verdict:

node "${CLAUDE_PLUGIN_ROOT:-.}/scripts/design-bindings-probe.mjs" --json

A SKIPPED verdict means one of: design.tokens.source is unset, no access token is present, or no committed variable-id map exists. Say so and stop cleanly — do not block. A mandatory gate on an absent integration breaks every project that has no designs, which is a worse outcome than any drift it would have caught.

If a design source is configured, then per tool-access-gate a failed probe against a configured source is a block, not a reason to fall back to reading pixels: post the Access Needed comment, move the item to the configured blocked role with the human_needed marker, and stop.

Phase 1 — Resolve configuration

Read .lisa.config.local.json first, then .lisa.config.json; local overrides global per key. Use jq — never hand-parse JSON.

read_key() {
  local key="$1" local_v global_v
  local_v=$(jq -r "$key // empty" .lisa.config.local.json 2>/dev/null)
  global_v=$(jq -r "$key // empty" .lisa.config.json 2>/dev/null)
  printf '%s' "${local_v:-$global_v}"
}
tokens_source=$(read_key '.design.tokens.source')
escalation_assignee=$(read_key '.design.escalation.assignee')
escalation_label=$(read_key '.design.escalation.label')

Never write a person's name, handle, or identity into any artifact this skill produces. These are config keys; the host supplies the values.

Phase 2 — Resolve the regime, per axis

For each of the six axes — color, spacing, typography, radius, elevation, motion — determine whether the design source publishes variables covering it. Never ask a human.

Do not try to list the published collections. That route does not exist headlessly, and the measurement is worth carrying:

A gate built on either of the first two runs interactively and silently no-ops everywhere else. So the regime is derived from the committed variable-id map, which design-variable-ids.mjs produced once, interactively, and which design-bindings-probe.mjs reads headlessly. An axis is typed when the map names at least one variable in its namespace.

result.regime from Phase 0 already carries this. Use it; do not re-derive it.

A mixed regime is the expected case, not an error state. Colour blocking while spacing gets measured, in the same work item, is the design working correctly.

Probe the subtree you are implementing, not the enclosing screen. Pass --node for the component you will actually build. A frame-level read counts the chrome behind a modal and over-reports — one measured work item scored 14 bound values at frame level and zero inside the modal subtree, and applying this rule changed 5 of 11 real work-item verdicts.

When the map cannot name what it saw

The probe reports owner: "us" when a VariableID is unknown (the map is stale) or ambiguous (two variables share a value). Neither is the designer's fault — those values ARE bound. Do not escalate them to design. Regenerate the map:

FIGMA_ACCESS_TOKEN=… FIGMA_MCP_TOKEN=… \
  node "${CLAUDE_PLUGIN_ROOT:-.}/scripts/design-variable-ids.mjs" \
  --file <key> --light <nodes> --dark <nodes>

Pass --dark wherever the library has a dark mode: the light+dark signature is what separates variables that share a value, and without it more ids stay ambiguous. An ambiguous id is still a failure — guessing which variable a value came from is exactly what this contract forbids.

An unknown id fails loudly naming the id; it never resolves to a nearest match. That self-detecting staleness is the only reason a committed map is safe to trust — it makes a stale map an error instead of silent corruption. Never add a fallback that picks the closest variable, and never treat an unresolved id as unbound-and-therefore-design's.

Phase 3 — Gather findings

The probe already produced hardcoded-in-design and bound findings for every value in the subtree, with the two silent-under-reporting traps handled: the corner-keyed rectangleCornerRadii shape is normalised, and boundness is read from boundVariables directly rather than inferred from a resolved value (Figma omits zero-valued properties, so a padding bound to a zero-valued variable would otherwise vanish).

Add the findings the probe cannot see from a node payload — a named variable that does not exist, an unpublished component, a missing state, a source disagreement — plus the two non-blocking observations. Findings are observations, not verdicts — do not pre-judge them; the gate decides.

Two of these rows carry most of the weight.

hardcoded-in-design is the condition most often missed, because nothing about the artifact looks wrong. The design is finished, the component renders, review sees a screenshot that matches — and the only symptom is that there is nothing to extract. Treating "I can see the value" as "I have the value" is precisely the failure this gate exists to catch.

aesthetic-concern is a finding, not a block, and it must still be emitted. Recording the observation without stopping on it is what keeps the gate from firing on opinion. If you find yourself reaching for a block because the design seems off rather than because a value is missing, that is an aesthetic-concern and the work proceeds.

Emit no finding kind outside this table. The gate throws on an unrecognised kind rather than inventing a verdict for facts it did not understand.

Phase 4 — Let the gate decide

Merge the probe's findings with your own, write them to a JSON payload, and run the gate. Do not reimplement the decision in prose — a verdict reasoned out fresh each time is a verdict that decides differently on Tuesday.

node "${CLAUDE_PLUGIN_ROOT:-.}/scripts/design-intake-gate.mjs" --findings=design-facts.json --json

Payload shape:

{
  "config": { "design": { "escalation": { "assignee": "…", "label": "…" } } },
  "regime": { "color": "typed", "spacing": "untyped" },
  "findings": [ { "kind": "hardcoded-in-design", "axis": "color", "component": "Raised card", "value": "#3A7BD5" } ]
}

Omit config to have the gate read the merged project config itself. Exit 0 = PROCEED, 1 = BLOCK, 2 = usage. The gate fails closed: an unreadable payload is a BLOCK with a named reason, never a quiet pass, because a gate that returns PROCEED on what it could not look at proves nothing.

Phase 5a — BLOCK

1. Post the comment. Use result.comment verbatim. It is written for the non-technical operator standing at the gate: plain language, naming the specific missing artifact and what to do about it, with no factory vocabulary. Do not "improve" it into engineering terms — the vocabulary ban is enforced by the gate's own tests, and rewriting the comment is how it gets lost.

   The 'Raised card' component uses the colour #3A7BD5 directly rather than a colour variable. I need that colour published as a variable so the app and the design stay in sync — otherwise I'd be copying a number that changes without warning.

2. Transition to the configured blocked role through lisa-tracker-sync / lisa-tracker-write. Never call a vendor label API directly — per config-resolution, Linear resolves blocked to a native workflow state, GitHub to a label, JIRA to a status, and hardcoding any one of them breaks the other two.

3. Assign to result.assignee. If it is null, the escalation-target-unset condition is already in result.blocks and the comment says so — post it and leave the item unassigned. Do not substitute a person.

4. Add result.label as an additive marker alongside the blocked role when one is configured, never instead of it. Add human_needed as well: a missing variable, an unpublished component, and a source disagreement all need a person to make a design decision, and none of them clear on a retry.

5. Stop. Do not partially build the surfaces that were fine. Design decisions arrive together, and half a screen built against half a design is rework.

Route by owner. Only owner: "design" reaches a designer. A stale or ambiguous id is owner: "us" — regenerate the map (Phase 2) and re-run. Sending that to a designer is the wrong person and the wrong work, and it teaches them to ignore the next one.

Phase 5b — PROCEED

1. Record the derived values from result.derived on the work item — axis, component, and value for everything measured in an untyped axis. This costs nothing now and accumulates into an inventory of what the variable system is missing, ranked by what people actually needed, which no design-system audit produces on its own.
2. Build. In a typed axis, bind the variable and use a screenshot only to verify; a disagreement you find at that point is source-disagreement and sends the item back through this skill. In an untyped axis, measure — that is the correct source there.

The rule is never "do not look at pixels". It is "do not derive a value from pixels when a binding exists."

Contract

• Block on unbound, never on unsure. Every block traces to one of the eight conditions in BLOCK_CONDITIONS. If you cannot name the condition, it is not a block.
• The regime is per-axis, derived from the committed variable-id map. Never per-project, never asked of a human, and never from a live collection query — that route does not run headlessly.
• A design source is optional. No source, no token, or no map is SKIPPED at exit 0, never a block.
• Every failure names an owner. Unbound values are design's; a stale or ambiguous map is ours.
• An unknown variable id fails loudly rather than resolving to a nearest match. Staleness is self-detecting, which is what makes a committed map trustworthy at all.
• The threshold is 100%. Relax it only with an explicit --min on the command line, never by softening anything in code.
• Never guess an escalation target, and never proceed without one.
• Never pick a side in a disagreement. Which source is right is a design decision.
• Never name a person in a comment, a commit, or any artifact. Config keys only.
• Never work around missing design information. Substituting a nearby value, inventing a state, or narrowing scope to dodge the gap are all forbidden — same standard as tool-access-gate.
• Host design-system rules (figma-design-system, design-system, use-the-design-library, or the project's equivalent) stay authoritative about what to build. This skill never overrides them.

Full policy: the design-value-binding rule.