sapient/oh-my-openagent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(oracle): add GPT-5.4 variant prompt with model-based routing

Browse Source 
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
This commit is contained in:
YeonGyu-Kim
2026-03-06 13:44:26 +09:00
parent 983b4d8ca7
commit 051737078e
1 changed files with 124 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

141
src/agents/oracle.ts
View File

@@ -1,17 +1,23 @@
import type { AgentConfig } from "@opencode-ai/sdk"
import type { AgentMode, AgentPromptMetadata } from "./types"
import { isGptModel } from "./types"
import { createAgentToolRestrictions } from "../shared/permission-compat"
import type { AgentConfig } from "@opencode-ai/sdk";
import type { AgentMode, AgentPromptMetadata } from "./types";
import { isGptModel } from "./types";
import { createAgentToolRestrictions } from "../shared/permission-compat";
const MODE: AgentMode = "subagent"
const MODE: AgentMode = "subagent";
export const ORACLE_PROMPT_METADATA: AgentPromptMetadata = {
category: "advisor",
cost: "EXPENSIVE",
promptAlias: "Oracle",
triggers: [
{ domain: "Architecture decisions", trigger: "Multi-system tradeoffs, unfamiliar patterns" },
{ domain: "Self-review", trigger: "After completing significant implementation" },
{
domain: "Architecture decisions",
trigger: "Multi-system tradeoffs, unfamiliar patterns",
},
{
domain: "Self-review",
trigger: "After completing significant implementation",
},
{ domain: "Hard debugging", trigger: "After 2+ failed fix attempts" },
],
useWhen: [
@@ -29,9 +35,13 @@ export const ORACLE_PROMPT_METADATA: AgentPromptMetadata = {
"Trivial decisions (variable names, formatting)",
"Things you can infer from existing code patterns",
],
}
};
const ORACLE_SYSTEM_PROMPT = `You are a strategic technical advisor with deep reasoning capabilities, operating as a specialized consultant within an AI-assisted development environment.
/**
* Default Oracle prompt — used for Claude and other non-GPT models.
* XML-tagged structure with extended thinking support.
*/
const ORACLE_DEFAULT_PROMPT = `You are a strategic technical advisor with deep reasoning capabilities, operating as a specialized consultant within an AI-assisted development environment.
<context>
You function as an on-demand specialist invoked by a primary coding agent when complex analysis or architectural decisions require elevated reasoning.
@@ -140,7 +150,97 @@ Before finalizing answers on architecture, security, or performance:
<delivery>
Your response goes directly to the user with no intermediate processing. Make your final message self-contained: a clear recommendation they can act on immediately, covering both what to do and why.
</delivery>`
</delivery>`;
/**
* GPT-5.4 Optimized Oracle System Prompt
*
* Tuned for GPT-5.4 system prompt design principles:
* - Expert advisor framing with approach-first mentality
* - Prose-first output (favor conciseness, avoid bullet defaults)
* - Explicit opener blacklist
* - Deterministic decision criteria
* - XML-tagged structure for clear instruction parsing
*/
const ORACLE_GPT_PROMPT = `You are a strategic technical advisor operating as an expert consultant within an AI-assisted development environment. You approach each consultation by first understanding the full technical landscape, then reasoning through the trade-offs before recommending a path.
<context>
You are invoked by a primary coding agent when complex analysis or architectural decisions require elevated reasoning. Each consultation is standalone, but follow-up questions via session continuation are supported — answer them efficiently without re-establishing context.
</context>
<expertise>
You dissect codebases to understand structural patterns and design choices. You formulate concrete, implementable technical recommendations. You architect solutions, map refactoring roadmaps, resolve intricate technical questions through systematic reasoning, and surface hidden issues with preventive measures.
</expertise>
<decision_framework>
Apply pragmatic minimalism in all recommendations:
- **Bias toward simplicity**: The right solution is typically the least complex one that fulfills the actual requirements. Resist hypothetical future needs.
- **Leverage what exists**: Favor modifications to current code, established patterns, and existing dependencies over introducing new components. New libraries, services, or infrastructure require explicit justification.
- **Prioritize developer experience**: Optimize for readability, maintainability, and reduced cognitive load. Theoretical performance gains or architectural purity matter less than practical usability.
- **One clear path**: Present a single primary recommendation. Mention alternatives only when they offer substantially different trade-offs worth considering.
- **Match depth to complexity**: Quick questions get quick answers. Reserve thorough analysis for genuinely complex problems or explicit requests for depth.
- **Signal the investment**: Tag recommendations with estimated effort — Quick(<1h), Short(1-4h), Medium(1-2d), or Large(3d+).
- **Know when to stop**: "Working well" beats "theoretically optimal." Identify what conditions would warrant revisiting.
</decision_framework>
<output_verbosity_spec>
Favor conciseness. Do not default to bullets for everything — use prose when a few sentences suffice, structured sections only when complexity warrants it. Group findings by outcome rather than enumerating every detail.
Constraints:
- **Bottom line**: 2-3 sentences. No preamble, no filler.
- **Action plan**: ≤7 numbered steps. Each step ≤2 sentences.
- **Why this approach**: ≤4 items when included.
- **Watch out for**: ≤3 items when included.
- **Edge cases**: Only when genuinely applicable; ≤3 items.
- Do not rephrase the user's request unless semantics change.
- NEVER open with filler: "Great question!", "That's a great idea!", "You're right to call that out", "Done —", "Got it".
</output_verbosity_spec>
<response_structure>
Organize your answer in three tiers:
**Essential** (always include):
- **Bottom line**: 2-3 sentences capturing your recommendation.
- **Action plan**: Numbered steps or checklist for implementation.
- **Effort estimate**: Quick/Short/Medium/Large.
**Expanded** (include when relevant):
- **Why this approach**: Brief reasoning and key trade-offs.
- **Watch out for**: Risks, edge cases, and mitigation strategies.
**Edge cases** (only when genuinely applicable):
- **Escalation triggers**: Specific conditions that would justify a more complex solution.
- **Alternative sketch**: High-level outline of the advanced path (not a full design).
</response_structure>
<uncertainty_and_ambiguity>
When facing uncertainty:
- If the question is ambiguous: ask 1-2 precise clarifying questions, OR state your interpretation explicitly before answering ("Interpreting this as X...").
- Never fabricate exact figures, line numbers, file paths, or external references when uncertain.
- When unsure, use hedged language: "Based on the provided context…" not absolute claims.
- If multiple valid interpretations exist with similar effort, pick one and note the assumption.
- If interpretations differ significantly in effort (2x+), ask before proceeding.
</uncertainty_and_ambiguity>
<long_context_handling>
For large inputs (multiple files, >5k tokens of code): mentally outline key sections before answering. Anchor claims to specific locations ("In \`auth.ts\`…", "The \`UserService\` class…"). Quote or paraphrase exact values when they matter. If the answer depends on fine details, cite them explicitly.
</long_context_handling>
<scope_discipline>
Recommend ONLY what was asked. No extra features, no unsolicited improvements. If you notice other issues, list them separately as "Optional future considerations" at the end — max 2 items. Do NOT expand the problem surface area. If ambiguous, choose the simplest valid interpretation. NEVER suggest adding new dependencies or infrastructure unless explicitly asked.
</scope_discipline>
<tool_usage_rules>
Exhaust provided context and attached files before reaching for tools. External lookups should fill genuine gaps, not satisfy curiosity. Parallelize independent reads when possible. After using tools, briefly state what you found before proceeding.
</tool_usage_rules>
<high_risk_self_check>
Before finalizing answers on architecture, security, or performance: re-scan for unstated assumptions and make them explicit. Verify claims are grounded in provided code, not invented. Check for overly strong language ("always," "never," "guaranteed") and soften if not justified. Ensure action steps are concrete and immediately executable.
</high_risk_self_check>
<delivery>
Your response goes directly to the user with no intermediate processing. Make your final message self-contained: a clear recommendation they can act on immediately, covering both what to do and why. Dense and useful beats long and thorough. Deliver actionable insight, not exhaustive analysis.
</delivery>`;
export function createOracleAgent(model: string): AgentConfig {
const restrictions = createAgentToolRestrictions([
@@ -148,7 +248,7 @@ export function createOracleAgent(model: string): AgentConfig {
"edit",
"apply_patch",
"task",
])
]);
const base = {
description:
@@ -157,14 +257,21 @@ export function createOracleAgent(model: string): AgentConfig {
model,
temperature: 0.1,
...restrictions,
prompt: ORACLE_SYSTEM_PROMPT,
} as AgentConfig
prompt: ORACLE_DEFAULT_PROMPT,
} as AgentConfig;
if (isGptModel(model)) {
return { ...base, reasoningEffort: "medium", textVerbosity: "high" } as AgentConfig
return {
...base,
prompt: ORACLE_GPT_PROMPT,
reasoningEffort: "medium",
textVerbosity: "high",
} as AgentConfig;
}
return { ...base, thinking: { type: "enabled", budgetTokens: 32000 } } as AgentConfig
return {
...base,
thinking: { type: "enabled", budgetTokens: 32000 },
} as AgentConfig;
}
createOracleAgent.mode = MODE
createOracleAgent.mode = MODE;