Why Moru Core · How It Works & 5-Month Results

When we first paired frontier models with a 20-file rulebook in July, token usage jumped 3–5× per task. Moru Core takes a simpler path: stay in the coding CLI you already use, read only the rules the current task needs, reuse passed test results when the code hasn't changed, and wait to report “Done” until the checks actually pass.

Spend tokens on the actual fix—not on re-reading rules and re-running the same tests.

Four practical habits that keep your agent focused.

You don't need to switch to a separate agent app or paste a 20-page rulebook into every chat. Moru Core fixes the four places where coding agents usually waste time and tokens.

Answer questions right away without extra steps

When you just ask “Why is this API slow?”, the agent doesn't create a branch or start a workflow. It reads the relevant code (within 4 hops) and answers right away while keeping your files untouched.

Read only the rules the current task needs

When a task edits files, Moru Core picks the right workflow (a small fix, a feature, a bug fix, or docs) and opens only the few rule files that apply—an average of 4.15 files in October, down 79.7% from 20.45.

Reuse passed tests when the code hasn't changed

When you follow up on a task or ask to commit, Moru Core checks whether the relevant files changed. If the code is untouched, passed local tests carry over automatically instead of running the test suite all over again (-54.4% fewer steps).

Get faster every week—and check before calling it Done

When the agent finds a tricky root cause, it saves a short note with an expiration date (at most 3 notes are recalled per task). And when a recurring mistake is fixed once, all 26 connected repos inherit the fix—cutting brand-new tasks from 112 steps in Week 1 to 23 steps in Week 3 (-79.5%). Finally, the task only finishes when tests pass, new files stay under 300 lines, and Moru Guard passes.

Stock CLIs, prompt packs, separate agent apps, and Moru Core.

Default coding CLIs, prompt skill packs (like Superpowers or Agent Skills), and standalone agent apps (like Nous Research's Hermes Agent or OpenHands) each focus on different parts of the workflow. Moru Core works inside your existing CLI to guide what the agent reads, what it remembers, and how it verifies the result.

What matters in daily work
Stock Coding CLI (Plain Claude Code / Codex)
Long Rulebooks & Skill Packs (Giant CLAUDE.md / Superpowers)
Separate Agent App (Hermes Agent / OpenHands)
Moru Core (Rules Inside Your CLI)

What actually happens when you ask for a change.

Here is how each approach handles four common requests—comparing the documented behavior of stock CLIs, prompt packs, and Hermes Agent with our measured October 2026 numbers across 26 repositories.

Approach
What the agent reads
What happens during the task
Typical steps & token profile
How the result is checked

What happens when you hold the model constant.

Because our 5-month history spans several model updates, we also compared runs using the exact same model ID across our 26 repositories (419,857 Spill events; 265 linked runs in October 2026).

On claude-opus-5-5, median steps per task dropped from 591.0 in September to 35.5 in October (-94.0%), and median token usage dropped from 33.47M to 0.41M (-98.8%).
Cutting rule files per task from 20.45 (Jul) → 4.29 (Sep) → 4.15 (Oct, -79.7%) also shrank average context per step on claude-opus-5-5 from 488.3K to 301.0K (-38.4%).
On gpt-6-sol, median steps per task dropped from 855.0 in September to 174.5 in October (-79.6%), with initial request-setup steps falling from 1,215.2 to 6.0 (-99.5%).

Continuing an existing claude-opus-5-5 task took 54.4% fewer steps (20.5 median steps across 80 runs) than starting from scratch (45.0 median steps across 91 runs).
When committing verified work, continuing from the previous check took 13.0 steps (0.29M tokens) compared to 52.5 steps (2.02M tokens, -85.6%) when starting fresh.
Compounding over time: even on brand-new (fresh) claude-opus-5-5 tasks, median steps dropped from 112.0 in Oct Week 1 → 48.5 in Week 2 (-56.7%) → 23.0 in Week 3 (-79.5%) as saved project notes and shared rule fixes accumulated across all 26 repos.

Measured across 26 repositories and 4,984 Git commits (May–Oct 2026).

Because 98.6% of raw tokens are prompt cache reads, we compare months using Spill's standard weighted token index (Fresh Input × 1.0 + Cache Write × 1.25 + Cache Read × 0.1 + Output × 1.0; a consistent comparison index, not billed USD) divided by commits and changed files. Active repositories grew from 11 in July (2.68 commits/day/repo) to 17 in October (3.86 commits/day/repo, +44.0% per project), while token usage per changed file dropped 78.1% (432.5K → 94.6K) and token usage per commit dropped 46.9% (3.07M → 1.63M).

Month
Active repos
Git commits
Commits / day (per repo)
Changed files
Steps / commit
Tokens / commit
Tokens / file
Rule files / task

How does Moru Core pick the exact 4 rules out of hundreds in milliseconds? Meet Wikimap.

Instead of stuffing a 3,000-token skill list into every prompt or burning extra LLM turns searching raw chat logs, Moru Core uses Wikimap—an open-source, zero-LLM incremental document engine built for Markdown knowledge bases.

0 LLM tokens · Sub-second incremental index

Single Python stdlib file with no external DB or model calls. Re-parses only SHA-changed files in milliseconds.

CJK-safe section search & Rank Fusion (RRF)

Finds the exact heading, tag, alias, or section across Korean, Japanese, Chinese, and English without reading whole files.

Link graph & SHA-pinned freshness

Tracks outlinks, backlinks, and REQ-IDs across docs, and automatically expires notes when source files change.