metis

Known Issues

Last updated: 2026-05-14

Carryover findings from prior implementation reviews that haven’t been fixed yet. These are bugs that look correct in isolation — invariants the specs claim that the code doesn’t quite honor, or capability declarations that are honest in intent but wrong in practice. AI agents working in the repo should know about them so they don’t:

1. Build new code that depends on the broken invariant.
2. Write tests that lock in the broken behavior.
3. Quote the spec back at the user when the impl quietly disagrees.

Severity legend:

• 🔴 high — silent correctness bug or substitutability invariant violation. Fix before the surface in question grows more consumers.
• 🟡 medium — spec promise unfulfilled; works correctly today but will bite when the dependent surface lands.
• 🟢 low — cosmetic, documented divergence, or design decision worth revisiting.

When you fix one, delete the entry. This file is not a changelog; it’s a watchlist.

---

Gaps that aren’t bugs (but worth tracking)

Things that aren’t promised by any spec but probably should be. AI agents proposing work in adjacent areas should know they’re missing.

• No context-assembler spec for history compression or behavior near the context window. docs/specs/context-assembler.md v1+§5.1 covers cache-breakpoint placement and the minimum-cacheable-prefix rule; §5.2 (v3) covers skill activation. History compression and behavior near the context window remain in §7 “Out of scope (later iterations).”

This site is open source. Improve this page.