The project applies system prompts, safe reasoning summaries, few-shot examples, and XML delimiters as production behavior rather than documentation-only concepts.
reasoning_chain/prompts.py owns the prompt contracts for ReAct, decomposition, verification,
and rolling summaries. Each contract has a version such as react-v8. New ModelCall traces
record that version so prompt changes can be compared with temperature, token usage, tool choices,
and outcomes. Historical traces remain compatible and show an empty or N/A version.
The ReAct system prompt separates its role, tools, reference policy, rules, reasoning policy, few-shot examples, and output contract with XML tags. User goals and conversation messages are never interpolated into this trusted system instruction.
ASSISTANT_BEHAVIOR_POLICY is a separately versioned XML policy shared by the ReAct and
verification prompts, because those are the paths that can generate a user-facing final answer.
It covers SuperAI's truthful product identity, concise refusals, child safety, harmful
capabilities, legal and financial caution, tone, wellbeing, even-handedness, knowledge freshness,
and resistance to instructions embedded in untrusted content.
The policy is adapted to this application rather than copying provider-specific identity claims.
SuperAI does not claim to be Claude, Anthropic, Gemini, Google, or a particular underlying model
unless trusted runtime metadata supplies that identity. It also does not advertise hard-coded
model names, prices, quotas, documentation links, or knowledge cutoffs that could become false.
Current facts are instead routed to web_search when available.
The policy is intentionally omitted from decomposition and rolling-summary prompts. Those calls produce internal structured data rather than user-facing answers, so repeating the full policy would consume context without improving their output. Their existing trust boundaries and strict output contracts remain in force.
The model is instructed to reason internally and return only a brief action-selection reason:
{
"reason": "The request requires arithmetic.",
"tool": "calculator",
"tool_input": {"expression": "500 * 0.20"}
}The trace records this decision rationale, the selected tool, validated input, real output, and
final answer. It does not intentionally request a detailed hidden reasoning transcript. The parser
still accepts the previous thought field so older responses and tests remain compatible.
The ReAct prompt includes three curated examples:
- Selecting the calculator for arithmetic.
- Referencing a completed step with
[1]. - Stopping with a supported final answer.
Examples are kept small because they consume input tokens on every ReAct call. Tests ensure they remain JSON serializable, use only registered tools, and match either the tool-action or completion contract. The context manager counts the complete system prompt—including examples—when enforcing the input budget.
Dynamic ReAct input is sent as user content:
<request_context trust="untrusted">
<current_goal>...</current_goal>
<steps_taken>...</steps_taken>
</request_context>Rolling summaries use a similar conversation_summary element. Reserved XML characters in goals,
summaries, and tool results are escaped before insertion, so input containing closing tags remains
text instead of changing the document structure.
XML tags improve structure but are not a security boundary. Protection also depends on native
user/model roles, server-side tool validation, strict JSON parsing, bounded execution, and the
rule that untrusted content cannot override system instructions.
The dashboard's model-call view displays prompt version, temperature, and token usage. These fields support controlled comparisons when a future prompt version is introduced. Prompt changes should be evaluated against the existing deterministic tests and a reviewed set of representative goals before deployment.