The Ω Protocol Family

ΩHERE

Heredoc-style command format with custom delimiters. Passes multi-line content — scripts, configs, file edits — with zero escaping. The workhorse of daily operations.

ΩHERE run_command
@command=bash
@stdin<<SCRIPT
echo "hello $USER"
SCRIPT
ΩEND

ΩBATCH

Structured batch execution with conditional chaining. Multiple independent operations in a single round-trip, with variable passing and error control between steps.

ΩBATCH{"steps":[
  {"tool":"run_command","params":{...},
   "saveAs":"v1"},
  {"tool":"eval_js","params":{...},
   "when":{"var":"v1","success":true}}
]}ΩEND

ΩPLAN / ΩFLOW

Higher-level task planning and workflow templates. ΩPLAN declares goals for intelligent decomposition; ΩFLOW invokes reusable workflow templates with parameterized steps.

ΩPLAN{"goal":"Deploy new feature"}
ΩFLOW{"template":"build-test-deploy"}

ΩCODE

Zero-corruption code transport channel. AI outputs raw code in the SSE stream; the extension intercepts it before DOM rendering and stores it in a dedicated storage backend. No JSON serialization, no shell escaping, no base64 bloat.

ΩCODE
var regex = /[^a-z]+/g;
var str = "hello$world";
return str.split(regex);
ΩCODEEND

JSON serialization

Backslashes doubled, quotes escaped, template literals interpreted

\n becomes \\n, ${var} interpolated

Shell expansion

$ and backticks trigger variable/command substitution

$USER replaced, $(cmd) executed

Regex in eval_js

Character classes mangled by nested escaping

[^a-z] breaks across JSON→JS→RegExp layers

Multi-layer nesting

Each transport layer adds its own escape rules

3+ layers = practically impossible to predict

Context Storage

Compression summaries, cross-session memory. Written by autoCompress, read on conversation restore.

writeContextStorage() / readContextStorage()

Code Storage

Dedicated to ΩCODE payloads. Intercepted from SSE stream and stored automatically, executed via one-line eval_js.

writeCodeStorage() / readCodeStorage()

Stream-Native

Commands and code ride the existing SSE stream. No side channels, no extra HTTP requests for transport. The stream is the bus.

Intercept Before Render

Ω markers are captured from raw SSE text before the chat UI renders them. The browser never tries to display or interpret the payload as markdown.

Same-Origin Advantage

Running inside genspark.ai means free access to authenticated APIs. No tokens to manage, no CORS issues. The browser is the auth layer.

Separation of Concerns

Context summaries and code payloads in separate storage slots. One cannot corrupt the other. Each has its own lifecycle and access patterns.

Compressed conversation summaries for cross-session memory. Auto-written by autoCompress, restored on new conversation boot.

Self-referential VFS registry. Maps slot names to metadata, sizes, and descriptions. The index of all other slots.

Dynamic prompt generator. Assembles injection payload from other slots on each conversation start, with live function name extraction from fn slot.

Operational reference: environment config, infrastructure IPs, tool selection rules, error handling patterns, and hard-won pitfall notes.

Compact system prompt template. Distilled rules and behavioral guidelines, injected alongside dynamic content.

22 skill quick-references with command templates and code snippets. Loaded on demand to save injection budget (~6.7K chars).

Executable function library. 20 JavaScript functions for API calls, DOM manipulation, data processing. Loaded via vfs.exec() into window.__tk.

CDP Snapshot

Chrome DevTools Protocol reads VFS slots directly from the browser via headless connection. No login needed — connects to existing Chrome instance.

PM2 Scheduled

PM2 ecosystem runs vfs-backup.py every 6 hours on the Oracle ARM server. Snapshots stored as timestamped JSON files.

Manual Trigger

vfs.backup() from any conversation exports all slots. vfs.snapshot(name) captures a single slot with timestamp.