Every release targets one concrete pain people hit running Claude Code (or any agent) daily. We ship what we already run ourselves; each slice is sanitized (framework only, our personal data never leaves). Releases twice a week, small commits daily.
| # | Pain | What we'll open-source | Status |
|---|---|---|---|
| 1 | CLAUDE.md bloats and stops being followed | The Bible skeleton: rule anatomy, precedence, routing tree, journals | ✅ v0.1.0 |
| 2 | Agent memory has the same disease — grows until half of it silently stops loading | Index-vs-body pattern for memory: hard caps, one-line pointers, archive discipline, a deterministic guard script | next — issue #1 |
| 3 | "Built" ≠ "works" — the agent says done and it's 2/3 done | The test-after-build gate: break-it-on-purpose checklist before any "done" | queued — issue #2 |
| 4 | Rejected ideas resurface every week | Declined-decisions journal deep dive: OVERTURN mechanics, revisit-conditions, real examples | queued — issue #3 |
| 5 | Multiple machines, one system — laptops and desktops drift apart | The multi-machine consensus protocol: dual-rail bus, ACK discipline, leader/follower canon, self-healing sync | ✅ shipped as claude-consensus |
| 6 | Agent-written content is embarrassing | The taste gate: turning the owner's real feedback signals into enforceable principles with a pass/fail verdict | queued — issue #4 |
| 7 | Personalized outreach at scale reads like spam | The CRM-to-messaging pattern: history-aware, research-backed message drafting | last, deepest sanitization — issue #5 |
| 8 | "My agent has 10 tools and any prompt injection can use them all" | LEASH-8: 8-domain control model, scorecard, plan-vs-authorize pattern, approval-design checklist, reference A2A card | ✅ shipped as agent-leash |
Live board: claude-bible — roadmap shows the same rows as Now / Next / Later / Shipped, and each open row is an issue you can claim by commenting "claiming this".
Have one of these pains — or a different one? Open an issue. Best signal for what to ship next.
Starring the repo genuinely helps: community catalogs require ~10 stars of social proof before they even accept a submission.