4SYNC — A Return on Context Harness (ARCH)

Never lose
the plot.

Your prototype worked. Then the project got bigger than the context window — and now every session starts with you re-explaining what you already built. 4SYNC ARCH keeps a project's memory, decisions and current state in plain files your agent reads at the start of every session. In your repo, in git, readable by you.

In production at its first instance: a ~600,000-token canon · boots on ~5% of it · months of sessions THE 4CITE INSTANCE, MEASURED AT ITS LAST CLOSE · ESTIMATE (BYTES ÷ 4), NOT TOKENIZER OUTPUT
READ THE CASE STUDY →
your-project/ · synchronized
your-project/
├─ your-project.yaml   # manifest: boot · close
├─ config/
│  ├─ KERNEL.yaml      # identity — always on
│  ├─ STATUS.yaml      # live state — overwrite
│  ├─ CANON_INDEX.yaml # the map
│  ├─ REFERENCE.yaml   # deep canon — on demand
│  └─ HISTORY.md       # frozen archive
├─ MERGE_PLAN.md       # task ledger + journal
└─ ABBA.md             # agent bulletin board

About

4SYNC — A Return on Context Harness (ARCH)

A small stack of plain files plus a manifest that tells any agent how to start a session and how to end one. There is no runtime, no service and no account. Your agent reads the files; the files are yours.

Return on Context (RoC) is a build philosophy that seeks to gain maximum value for each token spent. RoC is the antithesis of “tokenmaxxing.” Stock markets don't reward companies for spending the most money, they reward those that make the most money.

Product Life Story
Three kinds of truth, three write disciplines
IDENTITY

Who it is

The kernel carries the operating contract: directives, invariants and naming — front-loaded, so even a truncated read lands the rules that must never be missed.

config/KERNEL.yaml · edit rarely
STATE

What's true now

A live snapshot, overwritten in place. Deploy versions, active focus, blockers — answerable at a glance, never a log.

config/STATUS.yaml · overwrite
THE MAP

Where everything lives

A pointer index to every deeper document, so a session fetches depth on demand instead of loading the world up front.

config/CANON_INDEX.yaml · append a row
Declared in the folder, executed by any session
01 · BOOT

Orient before working

Every session loads the same small stack in the same order — kernel, status, map, ledger. Identity first, so the operating rules land before the backlog does.

02 · CLOSE

Write the truth back

On an explicit ending signal, the session journals what happened, syncs the ledger and overwrites the snapshot — so the next session inherits an accurate now.

03 · GENESIS

Born from a seed

Drop the stack into a project with a seed document and the first session authors its own identity from it. This site's own instance was born exactly that way.

boot receipt · printed at session start
═══ ARCH — BOOT RECEIPT ═══
instance   your-project/
manifest   your-project.yaml

This instance declares 5 file(s) to load, in this order:

  1  your-project.yaml         6,240 B    ~1,560 tok
  2  config/KERNEL.yaml       11,080 B    ~2,770 tok
  3  config/STATUS.yaml        9,315 B    ~2,328 tok
  4  config/CANON_INDEX.yaml   4,720 B    ~1,180 tok
  5  MERGE_PLAN.md            18,410 B    ~4,602 tok
     ═══════════════════════════════════════════════
     boot stack               49,765 B   ~12,440 tok

2 session(s) holding undeposited state:
  · a71c3f04   last wrote  07-14 09:22
  · 3e90bb17   last wrote  07-15 16:41   ← LIVE NOW
LOCAL CLI CLOUD SESSION IDE SCHEDULED RUN — same folder, same receipt, on all of them
over 40% kept out of every window OF THIS PROJECT'S OWN STACK · ESTIMATE (BYTES ÷ 4), NOT TOKENIZER OUTPUT
RUN scripts/meter.py --dir . --json FOR THE CURRENT FIGURE

Benefits

  1. Stop re-explaining your project every session
  2. A new session picks up where the last one ended
  3. Your agent stops undoing decisions you already made
  4. You can answer "why is it like this" months later
  5. Future-you can reconstruct your own reasoning, not just your code
  6. Rejected approaches stay rejected instead of quietly coming back
  7. Your notes can't silently go stale
  8. You can trust what a file says without re-verifying it yourself
  9. A bad or clipped read announces itself instead of passing as content
  10. Mistakes get caught when you make them, not three weeks later
  11. Work from any tool or device without setting it up again
  12. Never locked to one vendor's memory store
  13. Two windows open at once won't quietly destroy each other's work
  14. Onboard a new agent — or a collaborator — in one read
  15. Hand a task to a session that wasn't there when you wrote it
  16. Setup takes one session
  17. Adopting it doesn't change a line of your code
  18. Leaving costs nothing — delete the folder
  19. Your project's history lives in git, not a chat log you'll lose access to
  20. See what your context actually costs you
  21. Know which file is responsible when startup cost grows
  22. Free to use, and free to build on

Features

Nothing here is a service you call or a process you run — it is all files, and a routine for keeping them true.

Structure

  1. Five-file loader stack — kernel, status, map, deep reference, frozen history
  2. A separate write discipline per file (edit-rare / overwrite / append-a-pointer / edit-in-place / frozen)
  3. Front-loaded kernel — directives sit first, so even a truncated read lands the rules
  4. Task ledger: one summary table, all rows
  5. One long-form document per task, path derived from the row ID
  6. Closed task documents kept and greppable
  7. Session journal — newest-first blocks, capped, overflow to history
  8. Durable findings file, addressed by trigger, never loaded at boot
  9. Agent bulletin board and roster
  10. Naming and vocabulary canon file
  11. Commit landing queue for surfaces that can't run git
  12. Cross-repo commit pin
  13. Session-debt tracker — who's unwrapped, who's live right now

Mechanism

  1. A manifest that declares boot, close and genesis — pure declaration, no code
  2. Boot bookend: ordered load, identity before state before ledger
  3. Close bookend: freshness gate → journal → ledger sync → snapshot → rotate → meter → pin
  4. Genesis: drop in a seed document, the first session authors the stack from it
  5. On-demand loading — deep canon and task documents never enter the boot path
  6. End-of-file sentinels on every loader file
  7. Freshness gate — re-read a shared file immediately before writing it
  8. Anchored edits; never a whole-file rewrite from a session-start copy
  9. Post-write check against the regions the edit didn't target
  10. Self-contained task authoring rule — any session can execute a task cold
  11. Close fires only on an explicit signal; silence never triggers it

Tooling

  1. Seven write guards, applied at the tool call
  2. Boot receipt printed at session start
  3. Close-time checker: computed tally, size reports, subject caps, and fact-checking of stated figures against the real files, test files and repos
  4. Boot meter — per-file cost, logged to a series at every close
  5. Measured session cost, read from the agent's own transcripts
  6. Machinery drift alarm between copies
  7. Migration script for an existing monolithic ledger
  8. Hook installer
  9. Test suites — 914 tests across eleven files

Properties

  1. Plain markdown and YAML
  2. No package, plugin, framework or account
  3. Agent-agnostic — anything that reads files
  4. Surface-agnostic — local CLI, cloud session, IDE, scheduled run
  5. Git-native: diffable, revertable, lives in your repo
  6. Skill-free by design — no installation required to run the protocol
  7. Removable — delete the folder and carry on
  8. Source-available under FSL-1.1-ALv2
  9. Free

Download

4SYNC ARCH

Free Product

The whole stack: the manifest, the five-file loader, the task ledger, the bulletin board, the hooks and the close-time tooling. Copy it into a project, run one session, and it configures itself from your seed document.

Price Free
License FSL-1.1-ALv2 — source-available
Requires Claude Code, Git, Python 3
Removal Delete the folder and carry on!
Get it on GitHub → git clone https://github.com/SandmanCircles/4SYNC-ARCH.git
4SYNC ARCH Support

Paid Support

The struggle is real. Complexity drives cost, as do mistakes. Mitigate that risk with ARCH Support.

Implementation Genesis in your repo, guards wired, closes running. Multi-instance Several projects and surfaces, one discipline. Monolithic migration An existing project’s history, moved into the stack. Custom development Domain guards, integrations, the parts you need.

Your instance is the deliverable — yours from the first commit.

$200/HR · PROJECTS QUOTED FIXED-FEE Contact ARCH Support →
QUICK START

From your Claude

Make a new, empty folder — not an existing repo, and not Downloads or Desktop. Genesis writes that folder's absolute path in permanently. Open a Claude session there and paste this:

Clone https://github.com/SandmanCircles/4SYNC-ARCH into this folder,
then read CLAUDE.md and 4SYNC.yaml and run the genesis they declare.
I haven't filled in SEED.md — interview me through it.

Your session reads the protocol, walks you through the seed conversationally, then plays back what it understood — project name, purpose, and that permanent root path — and waits for your explicit go before writing anything. That playback is the only thing you have to read.