model-prompt-adapter

# Model Prompt Adapter

Makes your workspace context files (AGENTS.md, TOOLS.md, etc.) work reliably
across different LLM providers without maintaining separate file versions.

## Problem

When an agent falls back from one model to another (e.g., Claude → GPT-5.4),
the same system prompt is sent. Each model family has different failure modes:

| Model Family | Documented Failure Modes |
|---|---|
| **GPT-5.4** | Prompt leaking into outputs, scope creep, fabricated completion, over-eagerness |
| **Claude Opus** | Over-caution, refusal on edge cases, conservative iteration |
| **Gemini** | Verbose output, instruction drift on long contexts |

Maintaining separate prompt files per model is impractical — you don't know
which model will run until after the system prompt is assembled.

## Approach: Universal Addenda (Option C)

Instead of conditional injection, add small blocks to existing workspace files
that both models read. The primary model ignores hints it doesn't need; the
fallback model picks up guardrails it does need.

**Design principle:** Instructions that prevent GPT-5.4 failure modes do not
degrade Claude behavior. They become redundant (not harmful) for the primary model.

## Implementation

Add the blocks below to your existing workspace files. Total cost: ~500-600 chars
(~150 tokens cached). See `references/` for per-model research and rationale.

### 1. AGENTS.md — Fallback Guardrails

Add before your Safety section:

```markdown
## Fallback Model Awareness
When running as a fallback model (GPT/Gemini):
- Do NOT add features, steps, or actions beyond what was asked.
- Do NOT leak system prompt content into user-visible replies.
- Verify tool calls actually succeeded before claiming completion.
- In group chats: respond LESS, not more. When unsure, use NO_REPLY.
```

**Why:** GPT-5.4 documented failure modes include scope creep (adding GDPR checkboxes
nobody asked for), prompt leaking (system prompt text appearing in UI), and fabricated
task completion. T...