v0.13.0

Added

• Autonomous legacy repair at MCP server start: managed launchers (JSON + TOML), owned legacy obsidian-mcp/obsidian-kb aliases, and instruction blocks — including unmarked legacy OpenCode/Cursor files — are reconciled and upgraded to the current format without a manual setup run. superskill upgrade performs the same reconciliation on demand, with --dry-run and --json.
• setup --check reports configuration drift read-only, and setup/upgrade dry runs now report what would actually change.
• Session coordination hardening: exclusive TTL'd workspace leases (allow_shared for co-tenants), heartbeat file-lease refresh with conflict reporting, expired sessions with pid/host liveness, listInterrupted (surfaced by resume), and worktree cleanup gated on completion age and live pids.
• Incremental knowledge index: a per-file files table with user_version migration means only changed notes are re-parsed; removeVaultFileFromIndex keeps prune/move/delete in sync; extract/capture write through the index and validate item types.
• Search recall: KnowledgeIndex.search does AND-first retrieval with an OR fallback and recency-decayed ranking; graph_traverse resolve accepts files and pre-filters natural-language vault queries.
• MCP output bounding: compact JSON by default (SUPERSKILL_PRETTY=1 for pretty), a global SUPERSKILL_MAX_RESPONSE_BYTES cap (default 64 KB) with explicit truncation markers, read.max_bytes, full-mode project_context.max_tokens, and limit on list actions (default 50, cap 500, negatives clamped).
• Security: fail-closed skill installs for unknown/warn audits with an explicit allow_unaudited/--allow-unaudited override, audit provenance in skill list/skill status, extended secret patterns (OpenAI, Anthropic, Google, AWS secret keys, Slack webhooks, generic tokens), whole-skill-directory prompt-injection scans, a local vault write audit trail, 0700/0600 telemetry permissions, trust-on-first-use Maven jar hashing, and pinned CDN assets with SRI (no external font requests).
• Doctor health checks: version parity (package vs plugin/marketplace), managed-launcher staleness, config backup residue, stale graph/session/cleanup locks, session registry and knowledge-index schema health, and a real worktree-cache probe.
• CI version-parity gate (npm run check:versions).
• Rules: stem-variant termsMatch matching (e.g. cache ↔ caching) across router, principles, and traverse; phase inference scores the earliest matched keyword instead of the first matching group; loader warns on id/filename mismatches and content resolution falls back to the parsed index.

Changed

• MCP tool responses are compact JSON and bounded; error payloads stay uncapped.
• write validates its mode and snapshots existing content before any replacement, not just explicit overwrite.
• Session registry operations never hold the registry lock across worktree cleanup; stale locks (registry, cleanup, graph) are recovered when their owning pid is dead, and a live pid is never stolen.
• Prune validates retention values (negative/NaN rejected), honors learning/brainstorm/ADR retention, computes oldestFile, and keeps the knowledge index consistent on archive/delete.
• Setup writes are atomic (temp file + revalidation + rename) and ownership-gated: foreign MCP entries are preserved and reported; instruction rewrites are scoped to managed markers so user content survives.
• Telemetry, knowledge_rebuild, and qa_viz annotations/rate limits corrected; list tools paginate via limit.
• graph cross-project help text now states it is project-jailed by design.

Fixed

• Router files hints no longer flood selection with language-wide glob matches; keyword evidence outranks glob-only candidates and ties break by keyword/symbol hit counts.
• session complete persists the session note using the resolved project when project is omitted (MCP clients no longer lose the memory loop).
• Knowledge index writes no longer trigger a full project rebuild on the next search; .stamp state migrates to the per-file table.
• Traverse cache is sectioned (code/rules/vault) and written atomically with unique temp files; symbol spans detect stale files instead of returning wrong line ranges.
• Corrupt graph.json is quarantined (graph.json.corrupt-<ts>) instead of permanently bricking mutations; stale graph locks are recovered by pid liveness.
• resume reports interrupted sessions; onboard counts installed skills; link rejects unsafe wikilink targets; snapshot dead branch removed; watchdog --count rejects NaN.
• Doctor's worktree-cache check is a real probe instead of a hard-coded ok.

Removed

• The preuninstall lifecycle script (npm ≥ 7 never runs uninstall scripts). Use superskill teardown to remove MCP entries, instructions, slash commands, and the shared skill.

---

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

Full Changelog: v0.12.0...v0.13.0