TARS v2.2.0

Framework audit: efficiency, consistency, and automated validation

Added

• Documentation consistency validator (tests/validate-docs.py): 5-check validator catching stale skill references, provider-name leaks, count drift, archival tier terminology, and empty changelogs. Prevents semantic drift between docs and implementation.
• CI integration for doc validation: validate-docs.py added to .github/workflows/validate.yml with change detection for documentation files (ARCHITECTURE.md, README.md, GETTING-STARTED.md, CHANGELOG.md, CATALOG.md, reference/, skills/, commands/, scripts/).
• Universal constraints in core skill: 9 cross-cutting rules (date resolution, wikilink mandate, name normalization, task verification, integration constraints, index-first pattern, no-deletion, journal persistence, frontmatter compliance) now defined once in core and referenced by all skills.
• Maintain skill developer reference: Internal procedures for health-check.py, rebuild-indexes.py, and archive.py documented in ARCHITECTURE.md for maintainers without inflating the active skill.
• CONTRIBUTING.md: Consistency checklist for skill/script/protocol/integration/version changes, plus local validator run instructions.
• Maintenance breadcrumbs: HTML comments in core and maintain skills pointing to validate-docs.py and CONTRIBUTING.md.
• Help metadata: Added purpose, use_cases, scope help metadata to all 8 skills that were missing it (core, learn, think, maintain, welcome, initiative, create, communicate).
• Name resolution protocol (core skill): Cascading resolution for ambiguous and unknown names. Uses calendar attendees, transcript context, and memory files before asking the user. Batch clarification minimizes interruptions.
• Health mode auto-fix (maintain skill): Deterministic issues (decision file renames, index orphan removal, index rebuilds) are now auto-fixed instead of only suggested. Non-deterministic issues still presented to user.
• Inbox name pre-resolution (maintain skill): Names across all pending transcript files are resolved in a single pass before sub-agents spawn, ensuring consistency and preventing redundant user queries.

Changed

• Skill efficiency (~400 lines reduced): Deduplicated shared protocols across skills by referencing core definitions. Consolidated 3 inbox sub-agent templates into shared pipeline + type-specific processing. Streamlined maintain skill health/rebuild modes to script invocation with fallback references. Trimmed per-skill absolute constraints to unique-only (universal constraints in core).
• workflows.md: Rewrote all 8 workflow entries to use v2.2 skill names and reflect current pipeline architecture (was using v1.x names like process-meeting, extract-tasks).
• ARCHITECTURE.md: Fixed sub-agent descriptions to match actual implementations, corrected token baseline count, fixed script count (10→11), removed duplicate lines in workspace structure, corrected runner.sh reference.
• README.md: Fixed archival tier names from "active/warm/cool/archived" to canonical "durable/seasonal/transient/ephemeral".
• welcome skill: Updated command references (/daily-briefing→/briefing, /process-meeting→/meeting), replaced nonexistent create-shortcut skill reference with Cowork shortcut mechanism.
• shortcuts.md: Replaced create-shortcut skill reference with manual setup instructions.
• answer skill: Replaced hardcoded provider names (eventlink, remindctl) with generic language per provider-agnostic principle.
• CHANGELOG.md: Fixed wrong file path for getting-started guide (was reference/getting-started.md, corrected to GETTING-STARTED.md).
• ARCHITECTURE.md help schema: Aligned documented schema with actual implementation (purpose, use_cases, scope).
• Script count: 10→11 (added update-reference.py in v2.1.0, corrected in docs).
• Meeting skill pipeline reorder: Calendar lookup now runs before speaker name resolution so attendee lists inform the resolution cascade.
• Meeting skill name resolution: Enhanced from assumption-based inference to the core name resolution protocol with user clarification for remaining ambiguities.
• Learn skill: Both memory and wisdom modes now check for name ambiguity before processing. Unresolved names block extraction to prevent incorrect memory entries.
• Health mode output: Split into "Auto-fixed" and "Manual action required" sections. Deterministic fixes executed automatically, ambiguous issues presented to user.
• Health mode constraint: Expanded auto-fix scope from "only replacements" to include deterministic file renames, index orphan removal, and index rebuilds.
• Version: 2.1.0 → 2.2.0

---