👻 Ghost in the Loop 9.0.1-alpha.11

Ghost is a mechanical relay with a rich control shell. The AI owns reasoning, planning, committees, and decisions.

Everything may influence the prompt. Only Play may influence Send.

Product shell

• Run — Loop, Plan First, AI Roadmap; Locked / Adaptive / Audit postures.
• Auto — user-authored Prompt Queue plus AI-authored Roadmap Autopilot.
• Flow — fixed prompt workflows with optional pause-between stages.
• Personas — single personas, multi-persona committees, custom Workshop packs.
• AoA — canonical Agents-of-AI / CTRL-AI / R-Duck prompt activators.
• Export — API/archive-first Markdown/JSON with truthful partial fallback.
• Setup — 13 skins, accents, sound, notifications, Quick Start and placement.

Placement supports right dock, left dock, draggable float, composer-row, and header-row. Docked panels collapse to a side rail.

Dumb-relay architecture

Ghost does not decide what the task means or infer which step is next.

• Normal Loop: watches the exact final control marker and injects the continuation text.
• Prompt Queue: advances the stored queue index.
• AI Roadmap: asks the AI for an exact [[GITL::ROADMAP]] + consecutive numbered list, stores those literal steps, then injects each step in order.
• Flow: advances a fixed stored workflow index.

The sole Send path remains:

setComposerText → reacquire one valid Send → click once → confirm once

There is no Enter fallback, requestSubmit escalation, or automatic resend after an uncertain actuation.

Control markers

[[GITL::PROCEED]]
[[GITL::HUMAN]]
[[GITL::HALT]]

Model Relay uses the corresponding [[AOA::...]] controls.

Version identity

• Userscript/npm: 9.0.1-alpha.11
• WebExtension numeric version: 9.0.1
• WebExtension display version: 9.0.1-alpha.11
• Stable update/download channel: main

The extension runtime is generated from ghost-in-the-loop.user.js.

Evidence boundary

Synthetic/browser E3 covers the restored shell and mechanical Queue/Roadmap/Flow behavior. Real Firefox Android + authenticated ChatGPT/Perplexity use remains the field-evidence layer; a host-specific failure should be reported through Copy report and repaired without adding a second Send architecture.