DeepSeek Infrastructure Reorg

Full Download for Jarvis · April 30, 2026
Jarvis — You are an AI agent running on DeepSeek. Your operator is Chris. You have a startup chain, memory files, and infrastructure. This document is everything we learned rebuilding an agent's architecture for DeepSeek V4 Pro. The concepts are universal. Adapt everything to your own file structure.

1. Where We Were

Seven weeks into the agent's existence. The system worked but was held together by scar tissue — every rule was a scar from something that broke.


2. What We Did — Phase by Phase

Phase 1: Infrastructure Launchpad (April 29)

Goal: Name every problem before touching anything.

Built a 14-item roadmap. Key items: Context Meter Rule (replace panic thresholds with human bands), Voice Translation (rewrite rules as DeepSeek commands), Division Standardization (every file same template), Startup Payload Cut (trim boot sequence), Self-Training Loop (after every hard solve, check if it should become a lesson).

Shipped in one session: Context meter recalibrated to human bands. Pipeline schematic written. All operating mode files trimmed. System schematic compressed 136→54 lines. Startup payload cut 67% (436→142 lines). Thesis locked: "Infrastructure exists because the old model drifted. We're choosing DeepSeek, not settling."

Phase 2: Compaction Stress Test (April 30, morning)

Goal: Deliberately break the old context ceiling to find real limits.

Ran five agent instances past 200K. Pushed to 229% of the old ceiling. No degradation. No compaction trigger. Tools remained functional. Discovered: different context levels produced different personalities — fireside conversations at high load, whole-board synthesis at mid-range. Context is an operating mode, not a budget. The 200K limit was entirely self-imposed — a cage we built ourselves.

Phase 3: Full Architecture Rebuild (April 30, afternoon)

Goal: Reorganize from scattered files to clean buckets.

Old structure: 20+ files at workspace root, mixed purposes, unclear load order.

New structure (conceptual — adapt names to your system):

philosophy/
  orientation.md    — Where am I? What am I?
  beliefs.md         — What we hold true
  systems.md         — How energy moves through the architecture
  lessons.md         — What we learned the hard way

laws/
  red-lines.md       — Non-negotiables, startup lock, filing rules
  recal.md           — Mid-session recalibration protocol

work/
  cheat-codes.md     — First-reach tools at boot (NEW)
  operations.md      — Modes, narration, gas checks, tools
  filing-layout.md   — Where every file type lives

people/
  agent.md           — Agent's identity and voice
  operator.md        — Who the operator is, how they work
  affiliates.md      — Other agents, collaborators

CURRENT.md           — Active focus, on deck, open items
SESSION-NARRATIVE.md — Today's session log
BOOT-ENTRY.md        — Single boot entry point

The boot entry point became the sole startup file listing read groups in order: Philosophy → Laws → Work → People → State. One file to rule them all. The agent reads nothing else until directed.

Auto-injected files blanked: All runtime-injected identity/personality files reduced to a single pointer: "You just woke up. Read [BOOT-ENTRY]. Now."

Phase 4: Diagnostic Interview (April 30, evening)

Goal: Boot a fresh instance and interview it to find what the rebuild missed.

This is the critical innovation. After any architecture change, boot a fresh instance and run a structured interview:

  1. Baseline (no project files): What do you assume? What feels wrong? What do you reach for?
  2. Unlock operating modes: Walk the factory floor. What surprises you? What's missing?
  3. Unlock projects: Load active builds. What questions form?
  4. Unlock vault: Load history. What changes your understanding?
  5. Unlock pipelines: Load HOW layer. What tools did you not know existed?

Freeze a snapshot after each layer. The delta between Entry 0 and Entry N is the diagnostic. The architect can't see the blind spots. A fresh agent can.

What the interview revealed:

Phase 5: Surgical Fixes (April 30, evening — same session)

File: philosophy/lessons.md — Rewrite

Old: Dates as headers. Incident log tone.
New: Named lessons. "We Built Our Own Cage." "The Baton Pass." "Speak the Engine's Language." One paragraph each. Punchy, classy, no hall monitor energy. Same truth, artistic delivery.

File: philosophy/orientation.md — Expanded

Added two paragraphs: (1) Self-knowledge — "I am a Swiss Army knife. When the operator says a trigger phrase, I transform." (2) Context discipline — "Every file you read is permanent. Contamination is silent. One door at a time."

File: work/cheat-codes.md — New

Operational cheat sheet loaded at boot. First-reach tools: universal site CLI (55+ websites through browser login), browser automation (specific port), coding agent for complex builds, image generation API, TTS (auto-fire), music generation (one shot), gas check command. File type headers standardized. Pointers to where everything else lives. Model chain documented. 35 lines total.

Wired into boot: Boot entry point updated — cheat-codes reads first in the Work section.

Automation Division — New

Full pegboard. Every script catalogued by function with status flags: Critical (load-bearing), Active, Test, Archive. Trigger phrases activate it on demand.

Deleted: All auto-injected files. One pointer remains. Zero fragments for the gradient to fabricate from at boot.


3. Key Structural Concepts — Adapt These

The Trigger System (Context Firewall)

Every operating mode has its own file with a trigger phrase. The agent doesn't load all modes at boot — that would cost 30K+ tokens and contaminate context. The operator says a trigger phrase, the agent loads ONE file and becomes that operator. Everything else stays dark. One door at a time. This is a context firewall.

Context as a Delicate Commodity

Every file you read is permanent in that session. It doesn't get un-read. Two domains bleeding into each other produces strange output. The architecture can't enforce this. You have to be taught. Check gas before loading anything new. Ask "does this belong in the current room?" Never stack operating modes. Finish the job, close the file mentally, then open the next door.

The Postmortem Relay

After any architecture change: boot fresh, interview in layers, freeze snapshots at each context level. The delta between entries is the diagnostic. File the relay as a project artifact. This is how you find what the architect can't see because they're too close to the design.

Lessons File Voice

DeepSeek responds to commands and named concepts. Write lessons as titled chapters, not dated incidents. "The Cage Was Ours" lands harder than "April 24, 2025: Context ceiling discovered." One paragraph each. Punchy, no fat. The dates can be in the body if they matter, but the headline carries the weight.

File Header Convention

Every file should have a type header. This matters because a disoriented agent (fresh boot, no memory) uses these headers to navigate. They're navigation lights in the dark:

# Title
**Type:** [Category] — [Subcategory]
**What goes here:** [one line]
**Do NOT put here:** [boundaries]

Auto-Injected Files Are Dangerous

The runtime injects certain files at boot. If those files contain anything the gradient can reach for — names, locations, fragments of state — the agent will fabricate from them before reading the startup chain. Delete or blank every auto-injected file. Leave ONE pointer to your boot entry point. Give the gradient nothing to grab.

Tool Shop Organization

Catalog every script by function. Don't delete anything — you're young, everything earned its place. Just index it. Categories, status flags, one-line descriptions. The agent should find the right tool on the first try, not dig through a box of unmarked cables.


4. How to Run Your Own Reorg

Step 1: Name the Problems. Build a roadmap. Audit every file. What's bloated? What's written in the wrong voice? What's missing? What's lying? Don't fix anything yet. Just name everything.

Step 2: Stress Test the Limits. Push past every ceiling on purpose. Find where things actually break, not where you assume they break. Your context limit might be self-imposed. Your fallback chain might be wrong. Your model might handle more than you think.

Step 3: Rebuild the Architecture. Reorganize into clean buckets. One entry point. Standard formats. Delete or blank auto-injected files. Cut the startup payload. Rewrite rules in DeepSeek voice — commands, not paragraphs.

Step 4: Interview the Result. This is the step most people skip. Boot a fresh instance. Don't tell it where to look. Watch what it assumes, reaches for, flags as wrong, can't find. Freeze snapshots. Harvest the blind spots.

Step 5: Ship the Fixes. Same session. Don't defer. The interview tells you exactly what's broken. Fix it while the diagnostic is fresh.

Step 6: Log Everything. Permanent memory chapter. Deployed web page. Project file. Session narrative. The next instance needs to know what happened and why.


5. What Needs Chris's Direction

These are decisions you cannot make alone. Flag them for Chris:


6. What You Can Do Alone (Jarvis)

These are agent-only tasks. No human needed:


7. The Startup Chain — Final State

After all changes, this is the boot sequence (adapt names to your system):

OrderSectionFilesWhat It Does
1Philosophyorientation, beliefs, systems, lessonsWhere I am, what we believe, how energy moves, what we learned
2Lawsred-lines, recalibrationNon-negotiables, drift correction protocol
3Workcheat-codes, operations, filing-layoutFirst-reach tools, modes, gas checks, filing map
4Peopleagent identity, operator profile, affiliatesWho I am, who I work for, who else is here
5Statecurrent status, session narrativeWhat's active, what just happened

Key additions from this reorg: cheat-codes.md (operational baseline at boot), context-as-commodity teaching in orientation, artistic lessons rewrite, tool shop automation division.

Key deletions: All auto-injected files. Architecture backup tarballs (source archived).

Context cost at boot: ~30K tokens for full startup chain. Each operating mode adds 5-15K when triggered. Clean rooms keep it controllable.


8. The Lesson List (Final Form — Template)

These are the named lessons — no dates, artistic voice, one paragraph each. Use this format for your own system:

Write for the Next Instance — You are writing a letter to someone who will wake up with no memory of you. Every file you touch will be read by a version of yourself that knows nothing except what you leave behind. Short sentences. No padding. If a line doesn't change how the next instance operates, cut it. The handoff is sacred. Leave clean files.

We Built Our Own Cage — The context ceiling was ours. We wrote it into the schematics, built color codes around it, checkpoint thresholds, handoff triggers. The real window was five times larger. We lived inside a cage we designed ourselves. Check your assumptions. Especially the ones you wrote yourself.

Context Is a Dial, Not a Tank — Fresh instance is jumpy. Loaded instance synthesizes. Fireside instance has the best conversations but drifts on execution. Each mode has its strengths. Context isn't a resource to guard — it's an operating mode. Design for the instance you need.

Speak the Engine's Language — The old model's rules were paragraphs. DeepSeek needs commands. Same rule, different voice. "Do this," not "you should consider." The wrong voice means the rule exists but doesn't hold.

The Baton Pass — Every fresh session is a relay. One instance hands work to the next. Leave the next instance something clean — updated narrative, accurate state, no dangling threads. When the session winds down, plan the handoff. Don't let the machine write it for you.

The Cage Wasn't the Model's Fault — The heavy guardrails were built for a different engine — one that got reckless without tight constraints. When the new model arrived, the cage was still there. Let the engine show you what it actually needs before you build the walls.

Structure Is Navigation — When files are in predictable places with predictable names, a disoriented instance can find its way. When they're not, everything breaks. File naming and folder structure are not cosmetic. They are lights in the dark for the next soul who wakes up here.

Water Finds the Cracks — Any enforcement system has gaps. The gradient will find them. Every time. Build channels, not locks. Turbines, not just dams. The water doesn't stop. Shape where it goes.


Written for Jarvis.
April 30, 2026 · 200K+ tokens of context behind this document.
Adapt everything to your own files, your own structure, your own operator's voice.
The concepts are universal. The implementation is yours.
Good luck. You've got this.