v0.9.0

Added

• superskill-cli register [path] --slug <name> (and the MCP register tool): map a repo to a vault project so vault-backed commands auto-detect without -p. skill init now registers automatically and reports the mapping.
• superskill-cli doctor (and the MCP doctor tool): one-shot health check across runtime, install-vs-running-MCP freshness, vault, project mapping, project graph isolation, rules catalog, compile toolchains, telemetry, and MCP clients.
• Watchdog: session review + environment optimization. superskill-cli watchdog dig|fix|status and the watchdog MCP tool review harness session traces (OpenCode, Claude Code, Codex; harnesses without a reader fall back to SuperSkill session notes, graph, and telemetry), pin findings against the repo, and produce severity-ranked findings with evidence and proposals. dig scopes: one session, a window (--since/--count), the machine environment, or all; reports persist as projects/<slug>/watchdog/ notes with machine-readable finding status. fix is dry-run by default; file reclamation quarantines into ~/.superskill/quarantine/<timestamp>/ with a manifest for undo. status tracks open findings and recurrence across digs.
• Watchdog findings cover navigation thrash, verification gaps, tool error loops/duplicates, oversized outputs, steering bloat/duplication/staleness, dead skills, unused or erroring MCP servers, repeated prompt corrections, missing repo guardrails, and environment bloat (leaked .tmp files, runaway caches, harness store growth).
• New watchdog catalog pack (catalog/watchdog/dig.md, catalog/watchdog/fix.md) and a /watchdog slash command installed by setup; the router selects the pack for natural-language requests like "run a watchdog dig" or "clean up the agent environment".
• Session data quality: session complete records the registry's real started_at, closes the repo graph session with a mapped outcome and insights, and learn increments learnings_captured on the matching session note. Graph schema v3 adds lastUsed per skill (updated on every activation).
• Docs: new Watchdog README guide (loop, finding categories, CLI/MCP/slash/natural-language access, safety contract, env overrides).
• Worktree shared-cache policy: agents can share cheap caches (package stores, compiler caches, module caches) across git worktrees while keeping per-worktree build state isolated. worktree activate writes .git/superskill/policy.json and installs a guarded post-checkout hook that seeds new worktrees (reflink where available) and always exits 0.
• Host session adapters for Claude Code, OpenCode, Codex, Cursor, Gemini CLI, and GrokBuild (plus a generic AGENTS.md fallback) inject the shared-cache environment at session start; the OpenCode plugin merges worktree env --json into every shell.
• MCP tools worktree_status, worktree_audit, worktree_env, worktree_activate, worktree_apply, worktree_gc, worktree_uninstall (mirrored by superskill-cli worktree …). worktree status / worktree audit report per-worktree safety verdicts, cache duplication, seeded-manifest drift, and a --budget flag.
• worktree gc defaults to a dry-run report, quarantines selected dirs with --apply (reversible via --undo <journalId>), and permanently deletes only quarantine entries with --purge --yes; filters cover scope (--all, --worktree, --project, --tool), age (--older-than/--min-age, --newer-than), size, tier, globs, and --keep-latest.
• worktree apply applies audit items (policy, hook, seed, prune) behind consent; worktree uninstall removes hooks/adapters and keeps policy, caches, and quarantine. Seeding can be disabled with SUPERSKILL_WORKTREE_BOOTSTRAP=0; the cache root is configurable with SUPERSKILL_CACHE_ROOT.
• Docs: new Worktree caches README guide (agent-first usage, safety contract, full CLI reference, filter cookbook, host support matrix, troubleshooting, uninstall) and per-language ## Worktree & caches notes in catalog/code/.
• Slash-command installer: setup writes /review, /worktree, and /superskill command files for every host that supports them (markdown for most hosts, TOML for Gemini CLI); installs are idempotent and marker-guarded, and teardown removes them.
• doctor now includes worktree checks (shared policy present, guarded post-checkout hook installed) alongside runtime, install, MCP freshness, vault, graph isolation, catalog, toolchains, telemetry, and clients.
• Pre-commit hook runs gate check only when SUPERSKILL_GATE_TARGET is set; without a target it is a silent no-op.

Fixed

• CLI bin symlinks (superskill, superskill-cli) execute correctly when invoked through node_modules/.bin.
• Quarantine containment hardening: worktree-local reclaims stay inside the worktree, cache-root quarantines stay inside the cache root, and symlinked or escaping quarantine roots are refused.
• Worktree activation is consent-gated: policy and hooks are written only with explicit --yes / confirm: true.
• MCP worktree_gc tier validation rejects unknown tiers, and MCP write tools are rate-limited.
• Seeded-manifest verification covers directory entries, concurrent writes, and containment before a manifest is trusted.
• Hook backup/restore preserves bytes exactly.
• VaultFS rejects symlink escapes on every read/write path.
• Skill installer rejects path traversal and fails closed when audit status is unavailable.
• prune validates the mode before acting.
• Project-scope commands fail closed when no project slug can be resolved.
• Session registry, auto-number, and evidence writes are race-safe.

---

Install: npm install -g superskill or via Claude Code /plugin install superskill

Full Changelog: v0.8.1...v0.9.0