knowcards: v0.2.0

Highlights

• Knowledge injection is now title-first, providing each matching card’s slug, title, and optional “use when” guidance. Agents can fetch full card text with knowcards query or the MCP query tool.
• Prompt injection is capped at 8 cards and 8,000 characters, reducing unnecessary context usage.
• Retrieval now uses a streamlined MiniSearch ranking path, with prefix matching limited to short queries for more focused results.

Improvements

• Claude Code and Codex avoid reinjecting the same card slugs within a session.
• Cursor avoids rewriting its context rules when retrieved content is short or unchanged.
• Empty retrieval results no longer add host context.
• Reflection hooks skip runs when the transcript contains no file edits, reducing unnecessary follow-up turns.
• Documentation now clearly describes knowcards’ memory layers and the separation between filesystem memory and host adapters.
• The optional init command remains available but is hidden from CLI help; propose continues to create card directories as needed.
• Harbor A/B evaluation now uses four official SWE-bench Verified tasks covering pytest, requests, pylint, and Sphinx.
• Evaluation comparisons now separate uncached prompt tokens from cache tokens and report agent execution time instead of total Harbor wall-clock time.

Fixes

• Retrieval skips corrupt card files instead of failing the entire query.
• Claude Code, Cursor, and Codex hooks now fail open so hook errors do not interrupt host workflows.
• Stop-hook safeguards prevent repeated reflection when a hook is already active, including Cursor looped follow-ups.