fix(agentsskills): emit spec-conformant SKILL.md and diagnose name/description violations

[dyoshikawa]

Background

The agentsskills target is rulesync's implementation of the Agent Skills standard, yet the SKILL.md it generated was not conformant, and the normative name / description constraints were never checked.

Re-verified against the current specification (fetched 2026-07-27):

Field	Spec
name	1–64 chars, lowercase a-z0-9 and hyphens, no leading/trailing hyphen, no --, must match the parent directory name
description	1–1024 chars, non-empty
compatibility	optional string, 1–500 chars
metadata	optional map of string keys to string values
allowed-tools	optional space-separated string (experimental)

Changes

1. Export conformance

fromRulesyncSkill copied allowed-tools, compatibility and metadata verbatim out of the rulesync frontmatter, so a legacy list or object input was written as a YAML sequence / mapping that the spec's own skills-ref validate rejects. They are now normalized on the way out:

• allowed-tools: a list is joined with spaces — the same treatment DeepagentsSkill already applied to the identical field.
• compatibility: an object is flattened to key: value pairs.
• metadata: non-string values are stringified (scalars use their natural text form, containers are JSON-encoded) so the block stays a string→string map.

Input schemas are untouched, so existing .rulesync/skills/** frontmatter keeps working — only the emitted shape changes.

Before / after, for agentsskills: { allowed-tools: ["Read", "Bash(git:*)"], compatibility: { runtime: node }, metadata: { version: 1 } }:

# before                        # after
compatibility:                  compatibility: 'runtime: node'
  runtime: node                 metadata:
metadata:                         version: '1'
  version: 1                    allowed-tools: Read Bash(git:*)
allowed-tools:
  - Read
  - Bash(git:*)

2. Normative name / description diagnostics

Generation now reports, per skill:

• name empty, longer than 64 characters, containing anything but lowercase letters, digits and single hyphens, or not equal to its parent directory name;
• description empty or longer than 1024 characters;
• compatibility longer than 500 characters.

These are warnings, not errors. A conformant client only skips a non-conformant skill, import must stay lenient per the client-implementation guide, and failing the run outright would break existing skill directories. What must not happen — and no longer does — is emitting a skill that clients drop without rulesync saying anything.

Reporting needs a logger, so ToolSkillFromRulesyncSkillParams gains an optional logger that SkillsProcessor passes through. Other skill classes ignore it.

Verification

A skill directory My_Bad--Name with name: Totally-Different-NAME--x and an empty description now produces, on generate --targets agentsskills --features skills:

.agents/skills/My_Bad--Name/SKILL.md: `name` "Totally-Different-NAME--x" must contain only lowercase letters, digits and single hyphens, with no leading, trailing or consecutive hyphens
.agents/skills/My_Bad--Name/SKILL.md: `name` "Totally-Different-NAME--x" must match its parent directory name "My_Bad--Name"; conformant clients require them to be equal
.agents/skills/My_Bad--Name/SKILL.md: `description` is required and must not be empty; conformant clients skip a skill without one

Tests cover the three serializations, the already-conformant pass-through, every violation message, the length limits, and the no-warning case. docs/reference/file-formats.md documents the new contract.

pnpm cicheck passes.

Closes #2429

[dyoshikawa]

@dyoshikawa Thank you!