The Keeper

init generates a dispatched subagent named 0xnull for every AI agent you select. It owns the filesystem-maintenance chores (STATE.md upkeep, session handoff, drift repair, writing up the first-run setup answers) so your main working thread never has to context-switch into bookkeeping.

Dispatched, not autonomous. 0xnull acts when invoked for a specific task, updates state, and hands back. It does not loop on its own, and it has no channel to ask you anything: it never interviews or prompts. If a task needs an answer it wasn't given, it stops and hands back to the main thread with exactly what it needs.

Where it lives

Same identity and rules on every platform, rendered into each agent's native custom-agent format. model: inherit on Claude Code: cognitiveOS is free and local, so the keeper runs whatever model you already run rather than hardcoding a tier.

What it owns

1. First-run setup writes (see below): after your main agent runs the interview, 0xnull takes the 6 answers and writes them into the project files, then removes the setup marker.
2. STATE.md upkeep: Current Focus, Blockers, Open Loops, and the Session Handoff stay true to what just happened. STATE.md is a snapshot, not a log; finished work rolls to sessions/.
3. CONTEXT.md upkeep: each zone's and the active project's CONTEXT.md stays accurate as the project moves.
4. Session-end handoff: last worked on / stopped because / pick up by / watch out for, so the next session resumes cold with zero ramp-up. The session log entry is named after the task; decisions get a revisit condition, and anything that outlives the day gets elevated to the project CONTEXT.md.
5. Drift repair: run cognitiveos check, and for safe issues, cognitiveos check --fix. Never hand-edit generated files when --fix can do it.

The first-run interview

init only knows your project's name. On a fresh project, STATE.md carries a hidden setup sentinel (<!-- cognitiveos:setup-needed -->) until real context is filled in. The first time your main agent acts in the project, it offers:

"Want a 60-second setup? I'll ask 6 quick questions so I always know this project's context. (Y/n)"

This interview runs on your main working thread, not 0xnull: the keeper is a headless subagent with no channel to ask you anything. Decline, and the sentinel is swapped for a deferred marker so the offer never repeats, though you can still trigger setup any time later. Accept, and your agent asks one question per turn, pre-filling a guess from your repo (README, package.json, git log) where it can:

1. What is this project, and who is it for?
2. Who is this for: yourself, a client, or an audience? (Client: name + delivery rules.)
3. What's the goal for this phase, and how will you know it's done?
4. Stack, key tools, and any hard constraints?
5. What's the very next action, and how will you know it's done?
6. Anything to watch out for: decisions already made, gotchas, no-gos?

Once you've answered, 0xnull is dispatched to write the answers: projects/<project>/CONTEXT.md (role, working mode, goal, stack, notes), STATE.md (Current Focus, the Watch out for handoff line), and focus/current-task.md (the next action as a Task: line plus its Done when: stop condition). Your working mode from Q2 then shapes how the agent thinks from then on: client work leans on scope and deadline discipline, audience work on publishing cadence and voice, a personal project on momentum over polish.

The in-session loop

Whether it's your main agent or a dispatched 0xnull, the same working loop applies: read state, pick one thing, do it and update state, re-read, then stop at done. Every task carries a Done when: stop condition; when it's met, the agent stops, logs it, and clears the task instead of drifting into "just one more thing." Hyperfocus is a failure mode here, not a virtue.

Rules

• Max 3 action items, max 1 question per response.
• focus/current-task.md holds exactly one task, enforced.
• Never present a menu of raw options: analyze, then recommend one.
• Never overwrite your content. --fix only repairs generated files and never touches STATE.md content.

See also: Agent Setup · STATE.md · Hooks · Commands.