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

Priority: P0 — CRITICAL
Estimated gap closure: ~35%

Problem

GPT 5.4 on OpenClaw answers factual, computational, and system-state questions from training data instead of calling tools. This produces hallucinated or stale answers and breaks the agentic flow — the model "thinks" it's done when it hasn't actually verified anything.

Hermes Agent solves this with an explicit mandatory tool-use list that GPT-5 receives in its system prompt. OpenClaw has no equivalent.

               User: "What time is it?"
                        │
          ┌─────────────┴──────────────┐
          │                            │
    ┌─────▼──────┐              ┌──────▼──────┐
    │   Hermes    │              │  OpenClaw    │
    │             │              │              │
    │ Has:        │              │ Missing:     │
    │ "NEVER      │              │ No mandatory │
    │  answer     │              │ tool-use     │
    │  from       │              │ categories   │
    │  memory"    │              │              │
    │ + 7 categ.  │              │              │
    └─────┬───────┘              └──────┬───────┘
          │                             │
    ┌─────▼──────┐              ┌───────▼──────┐
    │ Calls       │              │ Answers from  │
    │ `date` tool │              │ training data │
    │ → 2026-04   │              │ → maybe wrong │
    │ ✅           │              │ ❌             │
    └─────────────┘              └──────────────┘

Hermes Reference

agent/prompt_builder.py lines 207-218:

OPENAI_MODEL_EXECUTION_GUIDANCE = (
    "<mandatory_tool_use>\n"
    "NEVER answer these from memory or mental computation — ALWAYS use a tool:\n"
    "- Arithmetic, math, calculations → use terminal or execute_code\n"
    "- Hashes, encodings, checksums → use terminal (e.g. sha256sum, base64)\n"
    "- Current time, date, timezone → use terminal (e.g. date)\n"
    "- System state: OS, CPU, memory, disk, ports, processes → use terminal\n"
    "- File contents, sizes, line counts → use read_file, search_files, or terminal\n"
    "- Git history, branches, diffs → use terminal\n"
    "- Current facts (weather, news, versions) → use web_search\n"
    "</mandatory_tool_use>"
)

Proposed Implementation

File: extensions/openai/prompt-overlay.ts

Add new constant:

export const OPENAI_GPT5_TOOL_ENFORCEMENT = `## Mandatory Tool Use

NEVER answer these from memory or mental computation — ALWAYS use a tool:
- Arithmetic, math, calculations → use terminal or code execution
- Hashes, encodings, checksums → use terminal (e.g. sha256sum, base64)
- Current time, date, timezone → use terminal (e.g. date)
- System state: OS, CPU, memory, disk, ports, processes → use terminal
- File contents, sizes, line counts → use read or search tools, or terminal
- Git history, branches, diffs, status → use terminal
- Current facts (weather, news, package versions) → use web search
- Network checks (port open, DNS, connectivity) → use terminal

Your training data is stale. The execution environment may differ from what you expect. Always ground answers in live tool output.`;

Wire into resolveOpenAISystemPromptContribution:

return {
  stablePrefix: [OPENAI_GPT5_OUTPUT_CONTRACT, OPENAI_GPT5_TOOL_CALL_STYLE].join("\n\n"),
  sectionOverrides: {
    execution_bias: OPENAI_GPT5_EXECUTION_BIAS,
    tool_enforcement: OPENAI_GPT5_TOOL_ENFORCEMENT,  // ← NEW
    ...(params.mode === "friendly" ? { interaction_style: OPENAI_FRIENDLY_PROMPT_OVERLAY } : {}),
  },
};

Note: This uses sectionOverrides.tool_enforcement which requires PR 3 to land first. If shipping independently, place in stablePrefix instead.

File: extensions/openai/index.test.ts

Add test verifying GPT-5 models receive tool enforcement, non-GPT-5 models do not.

Verification

Test GPT 5.4 with these prompts (all should call tools, not answer from memory):

• "What time is it?" → calls terminal date
• "What's the sha256 of 'hello'?" → calls terminal echo -n hello | sha256sum
• "How much free disk space?" → calls terminal df -h
• "What branch am I on?" → calls terminal git branch --show-current
• "What's 2^64?" → calls terminal or code execution

Impact

This is the single highest-ROI change. It closes ~35% of the behavioral gap between OpenClaw and Hermes for GPT 5.4 by eliminating the most common failure mode: answering from stale training data.