v3: rewrite wisdom as single-purpose skills in one plugin

[cmj0121]

Summary

Rewrites wisdom from scratch around three rules: one skill, one purpose, magic words only where they help, and fewer tokens. The 19 v2 plugins are replaced by one wisdom plugin with 21 skills, and smith stays the entry point.

• Skills: smith · plan design spec code test review docs · commit changelog release resolve · secure audit ops challenge · diagram compact brief lingua shortcut
• Docs: README.md rewritten and README.zh-TW.md added. CLAUDE.md is rewritten for v3 (221 → 69 lines).
• Checks: rewritten for the single-plugin layout. They enforce the token budget (description ≤ 160 chars, body ≤ 100 lines, smith ≤ 150), unique magic words, README tables that match exactly what ships, and wisdom:<skill> references that resolve. validate-fixtures is rewritten to match (16 fixtures).
• Hooks: lingua's SessionStart hook moves into plugins/wisdom/hooks/ unchanged, and the lingua.md format is the same, so existing preferences carry over.

Why it uses fewer tokens

Cost	v2	v3
name + description, loaded every session	4,063 chars (19)	3,244 chars (21)
All skill bodies, loaded per run	2,444 lines	996 lines
Largest body (the coordinator)	327 lines	108 lines
CLAUDE.md, loaded every session	221 lines	69 lines

• Descriptions: capped at one short sentence (160 characters) instead of 320.
• One purpose per body: each body describes a single job, so it carries no multi-phase role-play or persona framing.
• Delegation, not duplication: a second job is called as a skill instead of being restated inside the body (for example, the security check that v2's agent-ellis carried is now only secure).
• Plain result lines: skills that feed another skill end in one line (VERDICT:, RESULT:, …) instead of multi-field verdict blocks.
• PLAN.md as shared memory: subagents are pointed at it instead of being handed restated context, and a finished unit shrinks to one row.

Breaking changes

• Every v2 plugin is gone. Uninstall them and remove them from enabledPlugins, then install wisdom@wisdom. If both are enabled, lingua's hook runs twice.
• Skills are renamed to their purpose (agent-hale → code, agent-ellis → review, …), and most magic words change. smith is kept.
• Dropped from v2:
  • the per-plugin trigger-eval corpora
  • the model pins
  • per-skill versions
  • the "open a PR aligned with its issue" half of pr-flow (only resolve remains)

Test plan

• make test: all four checks pass
• bash scripts/validate-fixtures: 16/16 pass
• Install wisdom@wisdom in a fresh session and run a small task through smith
• Confirm magic words dispatch via shortcut

🤖 Generated with Claude Code