v0.24.0 — Project Knowledge Checkpoints Export
What's New
- Comprehensive project handoffs — export related knowledge from multiple wiki topics into a fixed five-file bundle under
docs/knowledge/<slug>/. - Coverage you can audit — every selected input carries a word count and must map to a knowledge section; omissions are explicit, and a 20% selected-source floor rejects thin “executive summary” dumps.
- Privacy sealed by default — semantic minimization is mandatory, deterministic scanning catches secrets, local paths, usernames, prompt-role transcripts, and related sensitive data, and exact reasoned overrides remain visibly attested.
- Safe lifecycle — create and refresh are dry-run-first; verify is read-only; import exposes only bounded evidence and never silently writes into another wiki or knowledge base.
- Multi-runtime release — synchronized Claude Code, OpenAI Codex, OpenCode, and portable-agent instructions and bundled CLI implementations.
Safety Boundaries
Checkpoint generation does not authorize committing, publishing, or importing the bundle. The scanner cannot be disabled. Overrides are exact, reasoned, audience-bound, and rechecked whenever content changes. No private wiki content, raw notes, session data, local paths, or generated project checkpoint is included in this release.
Validation
The release passed plugin validation (134), structural validation (170), checkpoint tests (10), lint (30), adapters (16), specialists (21), retract (6), session capture (19), session concurrency (10), documentation consistency, packaging sync, a clean-home Codex runtime install, Python compilation, Ruff, deterministic token budgets, and the token benchmark suite.
A real-project acceptance run expanded a rejected 1,969-word summary into a 10,327-word handoff from 45,546 selected source words (22.67%), with every input mapped and no privacy findings or overrides.