The Daily Loop

The system runs on a fixed overnight operating rhythm of 47 scheduled jobs orchestrating a day that starts with councils before dawn and a human gate in the morning, leaves plenty of time for execution and QA all day, and ends with a summary captain's log at night.

1. 22:00
   Nightly ingestion sweepAUTONOMOUS

   Five sources of potential knowledge: YouTube, RSS, academic papers, web search, and GitHub. Every night the system hunts for its standing search terms and potential knowledge gaps. A local model does the item-by-item reading, extracting knowledge while the humans sleep.

2. 01:00
   Graph night shiftAUTONOMOUS

   The Gardener sweeps the knowledge graph for contradictions and duplicate entities; a synthesis pass distills the last 24 hours of new nodes into higher-level insights and writes them back as first-class knowledge.

3. 02:00
   Memory metabolismAUTONOMOUS

   Episodic memories get decay-scored, never deleted. Every week fading episodes are crystallized into durable heuristics with source links back to the experiences that taught them.

4. ~04:30
   Knowledge routingAUTONOMOUS

   The moment the ingestion report lands, the system matches tonight's new nodes against every project's knowledge needs, marks knowledge gaps as filled, and retires the search terms that were chasing them.

5. 05:00
   Pre-flight ladderAUTONOMOUS

   A ten-step graph-health ladder, the council reputation rebuild at 05:25, and fleet preflight at 05:35 create one aggregated ship-status covering services, yesterday's failure rate, executor circuits, disk and swap headroom, and backup freshness.

6. 05:40
   The morning knowledge councilAUTONOMOUS

   Three CLI seats — Claude Code, Codex, Antigravity, each on its own subscription — deliberate independently over the overnight intelligence packet. An aggregator collates the council votes on the tasks for the day, knowledge that should be applied, new knowledge needs, and search expirations.

7. 06:00
   The morning pipelineAUTONOMOUS

   The output ranks and selects the day's top 12 tasks, with slot 1 reserved for the council's top self-improvement, and aggregates the rest of the votes. The plan is staged to disk, ready for the Operator to start the day.

8. on greeting
   "Good morning, Praxis"HUMAN GATE

   Each morning the operator greets the system and the day begins. The Inbox is filled with approval requests for recommended new skills, obsolete tasks to archive, potential new overnight search terms, and the day's schedule with per-slot objective, complexity, and difficulty tier. The Operator provides his feedback on these selections, with the ability to add instructions, modify the recommended executor, or skip a task altogether, which becomes a lesson for the system.

9. all day
   Dispatch → build → reviewAUTONOMOUS

   Approval releases slot 1 and the work starts. Each task dispatches with a bounded context pack to one of three executor CLIs; every completion is reviewed by a different AI on a different subscription; passing work is committed and the next slot dispatches automatically. Any issues can be flagged via a HITL skill, and the operator is notified both on desktop and mobile that their input is needed to continue. The day never stalls waiting on a permission request.

10. 18:00
    End-of-day commitHUMAN GATE

    A commit card surveys every workspace touched today, staged through a junk-and-secrets deny-list, and commits and pushes on approval. A failsafe sweep at the end of the day catches any stray edits.

11. 19:10
    Outcomes & calibrationAUTONOMOUS

    "Outputs are not outcomes": today's completions enroll for T+3 and T+14 re-checks — a task that gets reopened later is scored as a failed outcome, mechanically. The prediction ledger Brier-scores the council's impact calls against those outcomes. The system keeps score on itself.

12. 22:15
    Captain's logAUTONOMOUS

    Praxis provides a Captain's Log at the end of every day, with a stardate, as a vault entry, a chat card, and a spoken MP3. Fifteen minutes earlier, the next ingestion sweep has already begun. The loop closes.

# Captain's Log — Stardate 80561.9

Today was a clean, uninterrupted sweep across the queue. All ten scheduled
tasks reached completion without friction, leaving Praxis idle after
resolving our end-of-day human-in-the-loop commit. […] Nothing genuinely
resisted or failed today — no timeouts, no errors, and zero dropped tasks.
All twenty-one fleet services are answering on all channels.

It is almost unsettling when the day goes entirely according to plan, but a
quiet, compliant ship is a finding rather than a fault.