Skip to content
Runic
English
Esc
navigateopen⌘Jpreview
On this page

Scope

What v1 is, what got cut from the earlier execution-layer plan, and why.

Runic wasn’t always a cache. Earlier plans (archived under _deprecated/ in the repo, and in ALPHA.md/ROADMAP.md for historical context) described an execution runtime: intercepting tool calls, enforcing permissions, sandboxing filesystem/network access, and metering a hard budget. This page explains why that plan was cut, not extended.

In scope for v1

  • Caching a decision’s outcome, keyed on an exact-match normalized signature — { intent, params }, never raw task text, never a reasoning trace.
  • Logging tokens spent (agent-reported) vs. tokens saved on a cache hit.
  • Per-session scope. Cross-session/cross-agent sharing is a later phase, once per-session reuse is proven to matter in practice.

Explicitly out of scope

  • Execution of any kind. Runic never calls a capability, deploys anything, or runs code. The old packages/runtime is dead.
  • Permissions or sandboxing. There’s nothing to check if nothing executes. The old packages/permissions and packages/sandbox are dead.
  • Semantic or fuzzy task matching. This is deliberately narrow — exact match only. It’s a design constraint meant to stay a constraint, not a v1 shortcut to be widened later without re-opening the question of whether a cache hit is actually the same decision.

Why this is a pivot, not an addition

Most of the old execution-layer plan — permission enforcement, sandboxing, capability execution — turned out to duplicate what MCP already standardizes for tool calls. Building a parallel permission/execution system on top of that isn’t a defensible wedge on its own; it’s redundant infrastructure competing with a standard that already exists.

The cache-and-ledger idea survived the cut because it answers a question MCP doesn’t: has this exact decision already been resolved, and what did it cost? That’s a narrower, more honest claim than “Runic manages your agent’s execution” — and it’s the one this repo actually delivers on, with real benchmark numbers to back it up rather than an architecture diagram.

If you’re looking for the execution-layer plan

It’s preserved for history in ALPHA.md and ROADMAP.md at the repo root, and the code itself is archived under _deprecated/. Neither is maintained, and neither reflects what Runic ships today — read Introduction for that.

Was this page helpful?