token-map skill

Token map

Spec §10 / §21.3.1: every figma-migration / code-migration run crosses the same boundary — "the source has its own tokens; the target uses the active OD design system; we need a deterministic mapping". This atom produces that mapping.

Inputs

prompt; the atom reads from the staged design-system context).

• figma/tokens.json from figma-extract (figma-migration), OR
• code/tokens.json from design-extract (code-migration).
• The active design system DESIGN.md (already injected into the

Output

project-cwd/
└── token-map/
    ├── colors.json     # { source: '#hex' | 'tokenName', target: '--ds-token' }[]
    ├── typography.json # font + size + weight pairings
    ├── spacing.json    # spacing scale crosswalk
    ├── unmatched.json  # { source: ..., reason: 'no-target-equivalent' }[]
    └── meta.json       # { sourceKind: 'figma' | 'code', generatedAt, atomDigest }

unmatched.json is the audit list a human reviews; the agent must not invent target tokens silently.

Semantic token inference

Figma often exports anonymous source names such as color-3, paint/17, or raw #5B8DEF. Do not ask the user to rename those before mapping. First infer the semantic role from usage evidence:

frame name, layer name, and nearby text such as Primary, Selected, Link, Error, Focus, Nav, Button, or CTA.

• Node path, component name, instance overrides, variant/state labels,

text/icon, border, divider, overlay, shadow tint, focus ring, status badge, chart series, or brand/accent treatment.

• CSS-like position in the rendered tree: background fill, foreground

canvas is likely foreground; one paired with foreground inside CTA components is likely primary/accent background; a thin outline around interactive elements is likely border or focus-ring.

• Contrast relationships: a color paired repeatedly with the main

selected tabs, and active nav items is stronger evidence for --ds-color-primary than a value that appears once in an illustration.

• Reuse topology: a value that appears across primary buttons,

Use that role evidence to choose among existing active design-system tokens and to decide whether an anonymous token should be renamed or left unmatched before the executable mapping pass. Keep the on-disk token-map contract unchanged: the atom still writes the existing bucket files, unmatched.json, and meta.json only.

For example, this is a useful reasoning note for deciding whether color-3 should map to the active primary token:

{
  "source": "color-3",
  "value": "#5B8DEF",
  "role": "primary",
  "targetCandidates": ["--ds-color-primary", "--ds-color-link"],
  "evidence": [
    "Button/Primary fill",
    "Selected tab indicator",
    "Link text in Settings frame"
  ]
}

Then map to an active design-system token only when the evidence is role-based, not value-only. If the top candidates are too close to call, or if the evidence points to conflicting roles (primary vs link vs focus-ring), leave the source token unmatched using the existing no-target-equivalent reason and include the competing candidates in the hint. This keeps automation useful for common anonymous-token cases while preserving human review for ambiguous brand decisions.

Before / after expectation

Without semantic inference, an anonymous Figma token can only produce an uncertain value-level mapping:

{
  "source": "color-3",
  "value": "#5B8DEF",
  "target": null,
  "reason": "no-target-equivalent"
}

With semantic inference, the same token should carry role evidence before it is accepted:

{
  "source": "color-3",
  "value": "#5B8DEF",
  "target": "--ds-color-primary",
  "via": "name"
}

This deterministic v1 atom does not claim a measured accuracy lift by itself. Treat the expected improvement as coverage of previously manual anonymous-token cases when the Figma tree contains enough role evidence. Real accuracy numbers require a fixture suite with known source tokens, expected semantic roles, and a before/after agent run. See examples/semantic-inference-before-after.json for a deterministic same-token-batch simulation that compares the old value-level output with the semantic inference output.

Convergence

The atom completes when every input token is either mapped or explicitly recorded under unmatched.json with a non-empty reason. The until evaluator reads tokens.unmatched.length === 0 on strict mode; default is "soft converge" (proceed with unmatched.json populated).

Anti-patterns the prompt fragment forbids

approval (use a confirmation GenUI surface for that).

• Injecting a new token into DESIGN.md without explicit user

thresholds belong in the visual-diff evaluator (Phase 7).

• Mapping hex colours by visual proximity alone; perceptual ΔE

silently; record collisions in unmatched.json with reason target-collision.

• Collapsing distinct source tokens onto the same target token

Status

Implemented by the daemon runner in apps/daemon/src/plugins/atoms/token-map.ts. It parses design-system tokens, performs deterministic mapping, and writes the mapped and unmatched outputs.