Moru Core · Rules for coding agents

Moru Core opens only the ~4.15 rule documents the target file needs right before the first edit, and only allows completion when four checks are recorded: requested scope match, passing tests & screen checks, ≤300-line role split, and Moru Guard safety.

Blocks “done” until the requested work is actually checked.

Moru Core opens only the ~4.15 rule documents the target file needs right before the first edit, and only allows completion when four checks are recorded: requested scope match, passing tests & screen checks, ≤300-line role split, and Moru Guard safety.

Speed is already everywhere. We built Moru Core to stop repeating “fix it again.”

Treating AI as a hasty assistant leads to skipped rules, giant single-file dumps, and unchecked “done” reports. Moru Core turns that loop into a colleague workflow—delivering only the rules the target file needs and holding completion until verification is on the record.

Ask in your own words. The checks adapt to what you asked.

There are no extra commands to memorize after setup. Ask a question, ask for a small fix, or hand over a feature as usual. Questions get a direct answer right away, and code changes go through the checks they need.

Reads your project rules first

Opens AGENTS.md or CLAUDE.md first so your project's build, test, and architecture rules always take precedence.

Picks the right path and freezes required checks

Read-only questions answer directly (≤4 hops) with edits blocked; code changes freeze the exact verification gates upfront.

Delivers ~4.15 needed rules before the first edit

Selects only the ~4.15 rules matched to the target file and opens any unread rule right before the first file edit.

Done Check 1 & 2 · Scope match + test & screen proof

Requires real test/build outputs—and rendered screen checks for UI tasks—recorded in the run ledger.

Done Check 3 & 4 · ≤300-line split + Moru Guard

Verifies changed files stay within the 300-line role budget and runs Moru Guard (0 LLM tokens) for exposed keys or unsafe commands.

Opens “Done” only when all 4 proofs match file hashes

Compares recorded proofs against current file SHA hashes; if anything is missing or stale after a later edit, completion is blocked.

Your part stays simple.

Write your project's rules in one file, ask for work in plain words, and step in only when a choice belongs to you.

Put your project's rules in AGENTS.md

Write build commands, test commands, and coding rules in AGENTS.md or CLAUDE.md in your repository. Moru Core always follows your project's file first.

Follow up in a few words

Short follow-ups can build on the active task. The agent checks the requested scope and reuses earlier verification results when the relevant code and conditions have not changed.

Step in only when a decision is needed

If the answer is already in the code, the agent reads it instead of asking you. It stops and asks only when requirements conflict or a step—like pushing, deleting data, or adding a paid service—needs your approval.

Save what you learned for next time

The agent can save a brief verified cause, pitfall, or decision rationale with its source and a review date. Future tasks recall a bounded set of relevant notes as local references, while current project rules and source code always take precedence.

Its own worktree for every task in Git

When several agents work in parallel, one agent's half-finished edit should never leak into main. Each task runs in its own worktree and branch, named by your rule, and merged branches and worktrees are cleaned up afterwards.

Required documents arrive before the first edit

Asking an agent to read the rules was not enough, because some runtimes skipped them. Any required document the agent has not read is delivered right before its first file-modifying tool call; extra references open only for unresolved questions.

Holds “done” when a test fails

Most of the times we had to repeat “fix it again” came from a done report that nobody had checked. If a required check fails or is missing, completion is held until the agent fixes the issue and the test passes.

Continue from a concise record

A brief record captures goals, decisions, modified scope, and remaining work, so a follow-up does not start from zero. The agent checks it against current files when resuming, while live conversation upkeep stays with each native tool.

Project memory with sources and review dates

Only verified root causes and decision reasons are saved, each with its provenance and a review date. After two days of automatic recall at startup, we turned it off on Oct 24 so memory stays on demand: the agent opens a note only when the task calls for it.

One standard across four runtimes

Claude Code, Codex, Antigravity, and Gemini CLI read the same shared guidance and the same definition of done, instead of four siloed memories, while each tool keeps its own permissions, built-in tools, and context.

Moru Guard

Moru Guard is the safety check Moru Core runs before an agent edits your project and again before it finishes. Low-risk setup like .gitignore rules for .env files and commit checks is handled automatically, while hard-to-undo changes pause and wait for your decision.

Secrets

Finds exposed API keys and tokens without printing their values, and moves them out of code, commits, and logs.

Env files

Keeps .env and .env.local out of Git automatically, and makes sure shared .env.example files contain no real keys.

Server-only keys

Stops database URLs, admin keys, and webhook secrets that belong on the server from leaking into browser or app code.

Hard-to-undo work

Database resets, file deletion scripts, migrations, production deploys, and credential changes wait for your explicit approval.

Cost

Before adding a paid service, database, or model call, the agent checks whether caching or a simpler server helper is enough.

Git pushes

Before pushing, it checks the remote URL, whether the repository is public, and which files are being sent.