thierryc/glyphs-mcp/repair

Accompanying agent skill design

Skill: glyphs-curve-repair. Its entry, helper and reference now live under skill/glyphs-curve-repair/ in the companion and ship in the standalone build. Global installation and managed-catalog registration have not occurred.

The skill package contains a concise SKILL.md, thin helpers importing the shared bundled engine, references for stars/SuperTool/master verification, and examples. It must not contain another implementation of curvature, harmonization or deviation bounds.

Planned workflow instructions

1. Diagnose the request: static G1/G2, visual balancing, or an interpolation-only

kink. Route qualified HOI problems to dekink-native; disclose its narrower domain.

1. Read the advertised environment and bind document, exact glyphs, layers,

foreground/background and selected joins once. Expand the entire explicit scope independently of the Reporter's display cap.

1. Resolve method and tolerance. Default to native stars and 0.001 em maximum

effective-outline deviation; balancing is off. Explain when extraction or general-master verification is unqualified.

1. Obtain immutable source state and coverage. Follow V2's saved-baseline

requirements for native script work.

1. Analyze and propose using the shared engine and detached native adapter.

Explain continuity, maximum deviation, join movement, excluded targets and sampled verification region.

1. Apply only within existing authorization. Preview-only requests stay

non-mutating. Recheck exact native guards immediately before application.

1. Verify actual results, including neighboring joins and declared native

locations. Keep recovery available until the result is resolved.

1. Report success, no-op, refusal, partial application or incomplete coverage

explicitly, with recoverable state and proposal/job identity when relevant.

The skill must not infer that enabling Curve Inspector authorizes edits, that an advertised companion capability is callable, or that an unchanged Boolean smooth flag proves a star proposal is fresh.

Authorization and recovery language

Respect an existing request to repair. Do not ask again merely because an unsupported typed edit needs an authorized narrow native script route. A request to preview authorizes no live edit. A changed method, larger tolerance or wider scope must be surfaced and resolved before dependent application.

Explain native script restoration precisely: it reloads the saved font and discards all later unsaved changes. It is not selective rollback. Do not save, overwrite, install or restart without the user's actual authorization. Existing job ownership and saved-baseline gates still apply.

Helper boundaries and packaging

Helpers translate V2 reads and native results into shared engine structures, construct bounded jobs, summarize proposals and verify engine/package consistency. They import a bundled module at a pinned revision. They do not load a developer checkout, download algorithm code at runtime or reproduce formulas in scripts.

A native script can import a qualified installed engine bundle only when its presence and version are established. If unavailable, the packaging milestone must provide a supported reproducible delivery route; silently depending on an absolute path in this repository is not acceptable.

Planned references:

references/native-stars.md          # exact connections, effective outline gate
references/supertool.md             # pinned adaptation, license and refusals
references/master-verification.md   # families, region, lattice and coverage
references/unsupported-cases.md     # degeneracy, topology, HOI qualification
references/recovery.md              # V2 scripts, guards and whole-font reload

The concise skill entry links these references instead of embedding a large mathematical tutorial or host-specific qualification log.

Example interactions

Skill acceptance

Milestone 3 must validate discoverable metadata, concise instructions, packaging and imports; then execute examples against disposable fonts with the native qualification record. Skill tests must exercise intent, preview authorization, exact scope and coverage/refusal reporting. Merely generating a SKILL.md or having an MCP connection does not qualify the repair workflow.