Manta

Inspiration

I'm a designer who ships through coding agents. On any afternoon I have 3–5 Codex and Claude Code sessions running. One night, I suddenly realized the bottleneck was no longer the models but me. I rebuild context after every switch, miss the one decision that blocked everything, compulsively check sessions that didn't need me. Agent interfaces stream tokens while none of them manage the human's attention. Manta is named after the calmest way to move through deep water.

What it does

Manta is a local, read-only attention layer over your real agent sessions (~/.codex + ~/.claude). A derivation engine turns raw session logs into explainable state. Without LLM guessing, every claim citing its evidence (seq + excerpt): what each thread is doing, whether it's waiting on a judgment / a permission / a review, what changed since you left, which two threads are touching the same files, and whether anything is truly on fire (the only red in the product). A metronome serves interventions at your rhythm instead of the agents. One write action exists: send your decision straight back into the parked session via codex exec resume. Guarded by a freshness check, so a stale reply can never land on a question the agent already moved past. At the end of the day, a close-out screen accounts for every thread; it only ever says "safe to close" when that is literally true.

How Codex built it

Every line of product code was written by Codex (GPT-5.6) across chaptered sessions like data foundation, derivation layer, UI, bridge, etc, each of which is logged in docs/LOG.md. The meta part: Manta watched itself being built. Its own build sessions appear live in its inbox, their timelines showing the arc (exploring → implementing → verifying), and the demo's "send a reply back into the session" moment was first exercised on the very session that built it.

Challenges

Session logs have no format contract (we survived 9 schema versions); 1.4GB of logs demanded byte-offset incremental tailing; and the hardest part was restraint that decides the product would guarantee arrival, never predict timing, and never manufacture urgency.

What we learned

Trust kills the slot machine. Once "if it truly needs you, it will surface" is credible, compulsive checking loses its purpose.

What's next

Claude-side direct replies, thread lineage across forks, a public replay site, and the close-of-day ritual as a first-class habit toward a sustainable pace for human-agent work.

Built With

• chokidar
• codex
• css-modules
• express.js
• gpt-5.6
• node.js
• react
• server-sent-events
• typescript
• vite