Behind the Scenes: Migrating 77,000 Files & Building a Model Handover System

The Problem

Every project reaches a point where your storage setup becomes the bottleneck.

The original workspace lived on the D:\ drive — but as the project grew to 5,286 source files across 1,343 directories, with 19 AI agents, 18 skills, and 37 HTML reports, a few realities became unavoidable:

• The C:\ drive was dangerously low on space (29 GB free)
• D:\ was filling up and becoming a single point of friction
• The workspace needed to move to F:\ — a dedicated daily driver

The challenge? Move everything without breaking a single reference, path, or configuration. With 19 agents and dozens of scripts that reference absolute paths, one missed D:\ could break the entire system.

The Migration: 4 Phases

Phase 1: Robocopy (77,113 Files)

The first phase was pure data transfer — every file, every directory, every hidden config:

77,113 files  |  9,137 directories  |  2 GB copied

Robocopy handled the transfer with its native multi-threading, preserving timestamps and permissions. Exit code 1 — success with a few expected skips (locked files, temp caches).

Phase 2: Path Replace (17,258 Files)

The critical phase. Every D:\Projects\Project2Jarvis reference needed to become F:\sana-ai-dev-storage\Project2Jarvis — across 17,258 files.

Before: D:\Projects\Project2Jarvis\...
After:  F:\sana-ai-dev-storage\Project2Jarvis\...

A blanket find-and-replace operation patched every config file, script, documentation reference, and hardcoded path. The result was verified via Select-String audit: zero D:\ references remain on F:\.

Phase 3: Verification

Every subsystem was tested:

System	Check	Result
Git	git fsck	Clean — 5 commits intact
Obsidian	4 plugins, 10 themes	All present
Scripts	All 19 scripts	Identical on D: and F:
MCP Filesystem	Serves F:/ paths	Correct
Node modules	98 packages	Intact
Path references	Select-String audit	Zero D:\ found

Phase 4: Registry & Shortcuts

• Obsidian vault registry updated to point to F:\
• Desktop and Start Menu shortcuts cleaned — no stale D:\ references
• Residual install data cleaned from %LOCALAPPDATA%

Recovery: C:\ Drive Space

With the migration complete, I reclaimed 31 GB on the C:\ drive:

Action	Space Recovered
Disk Cleanup (temp files, caches)	+2.2 GB
npm cache cleared	+6.7 GB
Docker / WSL unregistered	+14.0 GB
Playwright browser cache	+2.6 GB
NVIDIA DXCache cleared	+5.6 GB
Total	+31.0 GB

From 29 GB → 60.4 GB free. The C:\ drive went from critical to comfortable.

The /onboarding System

The migration solved the storage problem, but there was a second, subtler challenge: model continuity.

Working with AI agents means occasional model switches — when the primary model exhausts its token budget, a fallback model takes over. Without proper context handover, the fallback model starts cold, wasting time and tokens re-establishing context.

The solution: a low-token onboarding system (~878 words YAML) that any fallback model can load in seconds.

How It Works

/onboarding command
    → Identity check — are you the primary or fallback?
    → Session verification — was state saved before handover?
    → Context load — AGENTS.md, MEMORY.md, active-context.yaml
    → Report back — "I'm [model] acting as [agent]. Ready."
    → Token warnings — guidance before the fallback runs dry

Key design decisions:

Decision	Rationale
YAML over Markdown	~1,500 tokens vs. 3,500+ — 57% less context overhead
Failover chain documented	Big Pickle → GLM-4.5-Air → any model
Dual deployment	Onboarding files live on both D:\ (reference) and F:\ (active)
Always save first	/onboarding warns: never switch models without saving session state

The Failover Chain

Step 1: opencode/big-pickle        → Primary — verify via /onboarding
Step 2: openrouter/glm-4.5-air      → Recommended fallback — session-save first
Step 3: any model via select-model  → Tertiary — manual override

The Agent Ecosystem

The project now runs 19 agents organized into three tiers:

Tier	Agents	Permission
Builders	JARVIS-Builder, ui-ux-pro-max, maintainer	Full access
Council	architect, security, performance, quality, docs, debugger	Review only
Agency	code-reviewer, mcp-builder, security-engineer, evidence-collector, workflow-architect	Specialized

Plus 8 slash commands for rapid workflows — from /onboarding for model handover to /awesome for the CSS POP design process.

Results

• 77,113 files migrated across drives with zero broken references
• 17,258 files path-patched automatically
• Zero D:\ references remain on the F:\ drive
• 31 GB C:\ drive space reclaimed (29 GB → 60.4 GB)
• /onboarding system deployed — 5-step model handover in ~878 YAML tokens
• 19 agents operational with 8 slash commands and 18 skills

Takeaways

1. Storage migrations are path-reference problems, not file-copy problems. The robocopy was the easy part — patching 17K references was the real work.
2. Verify with automation, not eyeballs. A Select-String audit across the entire tree catches what manual review misses.
3. C: drive hygiene compounds. Clearing npm cache alone freed 6.7 GB — unused caches are silent space killers.
4. Model handovers need a protocol, not hope. Without /onboarding, a fallback model wastes its first 3,500+ tokens reorienting.
5. Dual-drive deployment prevents lock-in. With onboarding files on both D: and F:, neither drive is a single point of failure.