v0.12.1

Fixed

• Tolerate Claude plugin list --json truncation. On a machine with many plugins, claude plugin list --json can hit a ~64 KB stdout cap and return malformed JSON, which previously broke every plugin-aware path (plugin list, mirror, plugin rm, add plugin). The Claude plugin adapter now reads ~/.claude/plugins/installed_plugins.json directly when the CLI is missing, exits non-zero, or returns unparseable/unexpected JSON (with a hint when stdout is exactly 65 536 bytes — the telltale of a buffer cap). marketplaceSources() gets the same fallback against ~/.claude/plugins/known_marketplaces.json, so mirror's <name>@<marketplace> resolution and provisioning still work when claude plugin marketplace list --json is unreadable too. Same shape, same data — just read from Claude's own state files instead of round-tripping the CLI. New exported parseClaudeInstalledPlugins for tests.
• Skill already-synced guard is now per-agent, not global. addSkillsFromPlugins previously skipped a source repo whenever every skill name was globally present in npx skills list -g --json — even if those skills were registered only on some agents and the current sync was targeting agents that didn't have them. After a TUI "Add from repo" that picked a subset of agents, a later syncthis run (cohort) would falsely report "already synced" and leave the missing agents missing. Now a repo is skipped only when every skill it contributes is already registered on every requested agent. force: true (interactive Add-from-repo / Sync-from-plugins) still bypasses the guard. Default run/sync behavior unchanged for the common case where the cohort already has everything.