Specification: Expose/Cull Hidden Invisible Prompting

1. Overview

The goal of this track is to review the codebase to identify, expose, or cull any hidden or invisible prompting injected by the system during discussion/sessions. This ensures the user has full control and visibility over the exact context sent to the AI API.

2. Functional Requirements

2.1 Identify Hardcoded Prompts

Audit src/ai_client.py to identify the hardcoded _SYSTEM_PROMPT and any tool execution instructions appended to requests.
Audit src/aggregate.py to identify headers and contextual markers injected during context aggregation (e.g., [SYSTEM: FILES UPDATED]).

2.2 Expose Prompts in GUI

For prompts that are necessary for the system to function (e.g., tool usage instructions, [FILES UPDATED] logic), expose them in the GUI (e.g., in "AI Settings" or "Project Settings").
Create editable text areas or configurable options so the user can modify or disable these prompts per-project or globally.
Ensure the modified prompts are correctly persisted and loaded by the AppController.

2.3 Cull Redundant Prompts

Remove any legacy or redundant prompting that no longer serves a purpose or duplicates user-defined system prompts.

3. Acceptance Criteria

All hardcoded system prompts in ai_client.py and aggregate.py are identified.
Necessary system prompts are exposed as editable fields within the GUI.
Users can modify or disable the default tool instructions or aggregation markers.
The ai_client utilizes the user-configured prompts instead of hardcoded strings.
Unnecessary or redundant hidden prompts are removed from the codebase.

4. Out of Scope

Modifying the Tiered MMA worker prompts in mma_prompts.py (this track focuses on the core discussion/session loop).
Adding a "Raw Prompt Preview" modal (this was an alternative option not selected).

1.9 KiB Raw Blame History