
Blume helps you build an agentic-native codebase with consistent and correct context. Fewer mistakes, fewer headaches.


"split importer discovery from sync rules; keep the public api the same"
"make the plan change page clearer; keep checkout behavior unchanged"
"mark archived tickets as resolved everywhere; add regression tests"
Both refactors this run were told to preserve existing behavior, again. Capture it as a shared rule and note why.
These refactors shuffled code across modules with no typecheck. Enforce one on every change.
Works with
Three steps to an agent setup your agents actually understand.
Immediately keep track of your running agents across harnesses and the hidden files that shape them.
Blume watches for mismatches between what your agents assume and what you actually want, then asks before changing anything.
No more repeating yourself. Your agents understand your domain. You spend your time building, not fighting the agent.
One narrow sidecar that watches your harnesses, tracks the hidden files behind them, and keeps your rulebook honest.
"move plan changes to the new checkout adapter; keep the existing public api"
Converting plan updates to the checkout adapter and preserving the current validation path.
All 5 retry scenarios pass. Root cause: worker env vars weren't propagated to the queue test harness.
Creates an on-demand skill that teaches future agents how you diagnose flaky test retries.
Recent chats ask for 2-space indent, but the rule still enforces four.
Skill: debug flaky retries from your chats
Skill · 1 change · Project
Review the exact change below. Nothing is applied until you confirm.
Expected effect
Loads only when an agent hits a failing retry loop. (project)
Align the indent rule with your chats
Rule, doc · 2 changes · Project
Review the exact changes below. Nothing is applied until you confirm.
Expected effect
Updates .cursor/rules to a 2-space indent. (project)
Expected effect
Notes why the change was made for the team. (project)
Today, and the corners of the agent setup we're building toward.
Always know when your agent finished, is working, or needs your approval.
Track the hidden files, skills, hooks, and rules that quietly shape what your agents do.
Cursor, Codex, Claude Code, omp, and Pi. More on the way.
Detect mismatches between your agent setups and your chat instructions. Blume proposes the fix; you approve.
See whether your agent setup is improving or degrading over time.
A single source of truth for intent across your project.
See what your agents are doing, catch drift early, and keep your codebase's rulebook current.