Parent: #66345 — GPT 5.4 Enhancement v3: Hermes Parity Sprint

Priority: P1 — MEDIUM
Type: Architecture / Refactor

Problem

PR 1 (#66346) adds mandatory tool-use categories for GPT-5. Currently, the only way to inject this is via stablePrefix (a flat string) or by hardcoding it into the execution_bias section override. Neither is clean:

• stablePrefix doesn't participate in the section override system, so providers can't selectively replace tool enforcement without touching execution bias.
• Stuffing it into execution_bias conflates two concerns: how the agent executes (bias) vs when it must use tools (enforcement).

Proposed Implementation

Architecture

  ┌──────────────────────────────────────────────────────┐
  │        System Prompt Section Pipeline                 │
  │                                                       │
  │  ┌─────────────────┐  ┌─────────────────┐            │
  │  │ interaction_style│  │ tool_call_style  │            │
  │  │ (personality)    │  │ (narration rules)│            │
  │  └─────────────────┘  └─────────────────┘            │
  │                                                       │
  │  ┌─────────────────┐  ┌─────────────────┐            │
  │  │ execution_bias   │  │ tool_enforcement │ ← NEW     │
  │  │ (act first,      │  │ (mandatory tool  │            │
  │  │  don't plan)     │  │  categories)     │            │
  │  └─────────────────┘  └─────────────────┘            │
  │                                                       │
  │  Each section: provider can override via               │
  │  sectionOverrides, or fall back to default             │
  └──────────────────────────────────────────────────────┘

File: src/agents/system-prompt-contribution.ts

Add "tool_enforcement" to the section ID union:

export type ProviderSystemPromptSectionId =
  | "interaction_style"
  | "tool_call_style"
  | "execution_bias"
  | "tool_enforcement";  // ← NEW

File: src/agents/system-prompt.ts

Add a new buildOverridablePromptSection call after execution_bias:

...buildOverridablePromptSection({
  override: providerSectionOverrides.execution_bias,
  fallback: buildExecutionBiasSection({ isMinimal }),
}),
...buildOverridablePromptSection({
  override: providerSectionOverrides.tool_enforcement,  // ← NEW
  fallback: [],  // empty by default — only providers that need it inject content
}),

File: extensions/openai/prompt-overlay.ts

Move tool enforcement from stablePrefix to sectionOverrides:

sectionOverrides: {
  execution_bias: OPENAI_GPT5_EXECUTION_BIAS,
  tool_enforcement: OPENAI_GPT5_TOOL_ENFORCEMENT,  // ← moved here
  ...(params.mode === "friendly" ? { interaction_style: ... } : {}),
},

Benefits

1. Separation of concerns: Execution bias and tool enforcement are distinct prompt concerns
2. Provider extensibility: Google, Anthropic, or custom providers can inject their own tool enforcement via the same mechanism
3. Testability: Each section can be tested independently
4. Future-proofing: New providers (e.g., Google Gemini in PR 6) can use tool_enforcement without modifying core

Verification

• Unit test: GPT-5 models receive tool enforcement section in system prompt
• Unit test: Non-GPT-5 models receive empty tool enforcement section
• Unit test: Custom provider can override tool enforcement independently of execution bias