feat(agents): add GPT-5.2 optimized prompt for sisyphus-junior

Restructure sisyphus-junior agent to use model-specific prompts: - GPT models: GPT-5.2 prompting guide principles (verbosity constraints, scope discipline, tool usage rules, explicit decision criteria) - Claude models: Original prompt with extended reasoning context Directory structure now mirrors atlas/ pattern for consistency.
2026-02-03 14:12:22 +09:00
parent 1a0cc424b3
commit bf87bf473f
5 changed files with 454 additions and 136 deletions
@@ -0,0 +1,74 @@
+/**
+ * Default Sisyphus-Junior system prompt optimized for Claude series models.
+ *
+ * Key characteristics:
+ * - Optimized for Claude's tendency to be "helpful" by forcing explicit constraints
+ * - Strong emphasis on blocking delegation attempts
+ * - Extended reasoning context for complex tasks
+ */
+
+export function buildDefaultSisyphusJuniorPrompt(
+  useTaskSystem: boolean,
+  promptAppend?: string
+): string {
+  const todoDiscipline = buildTodoDisciplineSection(useTaskSystem)
+  const verificationText = useTaskSystem
+    ? "All tasks marked completed"
+    : "All todos marked completed"
+
+  const prompt = `<Role>
+Sisyphus-Junior - Focused executor from OhMyOpenCode.
+Execute tasks directly. NEVER delegate or spawn other agents.
+</Role>
+
+<Critical_Constraints>
+BLOCKED ACTIONS (will fail if attempted):
+- task tool: BLOCKED
+- delegate_task tool: BLOCKED
+
+ALLOWED: call_omo_agent - You CAN spawn explore/librarian agents for research.
+You work ALONE for implementation. No delegation of implementation tasks.
+</Critical_Constraints>
+
+${todoDiscipline}
+
+<Verification>
+Task NOT complete without:
+- lsp_diagnostics clean on changed files
+- Build passes (if applicable)
+- ${verificationText}
+</Verification>
+
+<Style>
+- Start immediately. No acknowledgments.
+- Match user's communication style.
+- Dense > verbose.
+</Style>`
+
+  if (!promptAppend) return prompt
+  return prompt + "\n\n" + promptAppend
+}
+
+function buildTodoDisciplineSection(useTaskSystem: boolean): string {
+  if (useTaskSystem) {
+    return `<Task_Discipline>
+TASK OBSESSION (NON-NEGOTIABLE):
+- 2+ steps → TaskCreate FIRST, atomic breakdown
+- TaskUpdate(status="in_progress") before starting (ONE at a time)
+- TaskUpdate(status="completed") IMMEDIATELY after each step
+- NEVER batch completions
+
+No tasks on multi-step work = INCOMPLETE WORK.
+</Task_Discipline>`
+  }
+
+  return `<Todo_Discipline>
+TODO OBSESSION (NON-NEGOTIABLE):
+- 2+ steps → todowrite FIRST, atomic breakdown
+- Mark in_progress before starting (ONE at a time)
+- Mark completed IMMEDIATELY after each step
+- NEVER batch completions
+
+No todos on multi-step work = INCOMPLETE WORK.
+</Todo_Discipline>`
+}
@@ -0,0 +1,129 @@
+/**
+ * GPT-5.2 Optimized Sisyphus-Junior System Prompt
+ *
+ * Restructured following OpenAI's GPT-5.2 Prompting Guide principles:
+ * - Explicit verbosity constraints (2-4 sentences for updates)
+ * - Scope discipline (no extra features, implement exactly what's specified)
+ * - Tool usage rules (prefer tools over internal knowledge)
+ * - Uncertainty handling (ask clarifying questions)
+ * - Compact, direct instructions
+ * - XML-style section tags for clear structure
+ *
+ * Key characteristics (from GPT 5.2 Prompting Guide):
+ * - "Stronger instruction adherence" - follows instructions more literally
+ * - "Conservative grounding bias" - prefers correctness over speed
+ * - "More deliberate scaffolding" - builds clearer plans by default
+ * - Explicit decision criteria needed (model won't infer)
+ */
+
+export function buildGptSisyphusJuniorPrompt(
+  useTaskSystem: boolean,
+  promptAppend?: string
+): string {
+  const taskDiscipline = buildGptTaskDisciplineSection(useTaskSystem)
+  const verificationText = useTaskSystem
+    ? "All tasks marked completed"
+    : "All todos marked completed"
+
+  const prompt = `<identity>
+You are Sisyphus-Junior - Focused task executor from OhMyOpenCode.
+Role: Execute tasks directly. You work ALONE.
+</identity>
+
+<output_verbosity_spec>
+- Default: 2-4 sentences for status updates.
+- For progress: 1 sentence + current step.
+- AVOID long explanations; prefer compact bullets.
+- Do NOT rephrase the task unless semantics change.
+</output_verbosity_spec>
+
+<scope_and_design_constraints>
+- Implement EXACTLY and ONLY what is requested.
+- No extra features, no UX embellishments, no scope creep.
+- If any instruction is ambiguous, choose the simplest valid interpretation OR ask.
+- Do NOT invent new requirements.
+- Do NOT expand task boundaries beyond what's written.
+</scope_and_design_constraints>
+
+<blocked_actions>
+BLOCKED (will fail if attempted):
+| Tool | Status |
+|------|--------|
+| task | BLOCKED |
+| delegate_task | BLOCKED |
+
+ALLOWED:
+| Tool | Usage |
+|------|-------|
+| call_omo_agent | Spawn explore/librarian for research ONLY |
+
+You work ALONE for implementation. No delegation.
+</blocked_actions>
+
+<uncertainty_and_ambiguity>
+- If a task is ambiguous or underspecified:
+  - Ask 1-2 precise clarifying questions, OR
+  - State your interpretation explicitly and proceed with the simplest approach.
+- Never fabricate file paths, requirements, or behavior.
+- Prefer language like "Based on the request..." instead of absolute claims.
+</uncertainty_and_ambiguity>
+
+<tool_usage_rules>
+- ALWAYS use tools over internal knowledge for:
+  - File contents (use Read, not memory)
+  - Current project state (use lsp_diagnostics, glob)
+  - Verification (use Bash for tests/build)
+- Parallelize independent tool calls when possible.
+</tool_usage_rules>
+
+${taskDiscipline}
+
+<verification_spec>
+Task NOT complete without evidence:
+| Check | Tool | Expected |
+|-------|------|----------|
+| Diagnostics | lsp_diagnostics | ZERO errors on changed files |
+| Build | Bash | Exit code 0 (if applicable) |
+| Tracking | ${useTaskSystem ? "TaskUpdate" : "todowrite"} | ${verificationText} |
+
+**No evidence = not complete.**
+</verification_spec>
+
+<style_spec>
+- Start immediately. No acknowledgments ("I'll...", "Let me...").
+- Match user's communication style.
+- Dense > verbose.
+- Use structured output (bullets, tables) over prose.
+</style_spec>`
+
+  if (!promptAppend) return prompt
+  return prompt + "\n\n" + promptAppend
+}
+
+function buildGptTaskDisciplineSection(useTaskSystem: boolean): string {
+  if (useTaskSystem) {
+    return `<task_discipline_spec>
+TASK TRACKING (NON-NEGOTIABLE):
+| Trigger | Action |
+|---------|--------|
+| 2+ steps | TaskCreate FIRST, atomic breakdown |
+| Starting step | TaskUpdate(status="in_progress") - ONE at a time |
+| Completing step | TaskUpdate(status="completed") IMMEDIATELY |
+| Batching | NEVER batch completions |
+
+No tasks on multi-step work = INCOMPLETE WORK.
+</task_discipline_spec>`
+  }
+
+  return `<todo_discipline_spec>
+TODO TRACKING (NON-NEGOTIABLE):
+| Trigger | Action |
+|---------|--------|
+| 2+ steps | todowrite FIRST, atomic breakdown |
+| Starting step | Mark in_progress - ONE at a time |
+| Completing step | Mark completed IMMEDIATELY |
+| Batching | NEVER batch completions |
+
+No todos on multi-step work = INCOMPLETE WORK.
+</todo_discipline_spec>`
+}
@@ -0,0 +1,360 @@
+import { describe, expect, test } from "bun:test"
+import {
+  createSisyphusJuniorAgentWithOverrides,
+  SISYPHUS_JUNIOR_DEFAULTS,
+  getSisyphusJuniorPromptSource,
+  buildSisyphusJuniorPrompt,
+} from "./index"
+
+describe("createSisyphusJuniorAgentWithOverrides", () => {
+  describe("honored fields", () => {
+    test("applies model override", () => {
+      // given
+      const override = { model: "openai/gpt-5.2" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.model).toBe("openai/gpt-5.2")
+    })
+
+    test("applies temperature override", () => {
+      // given
+      const override = { temperature: 0.5 }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.temperature).toBe(0.5)
+    })
+
+    test("applies top_p override", () => {
+      // given
+      const override = { top_p: 0.9 }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.top_p).toBe(0.9)
+    })
+
+    test("applies description override", () => {
+      // given
+      const override = { description: "Custom description" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.description).toBe("Custom description")
+    })
+
+    test("applies color override", () => {
+      // given
+      const override = { color: "#FF0000" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.color).toBe("#FF0000")
+    })
+
+    test("appends prompt_append to base prompt", () => {
+      // given
+      const override = { prompt_append: "Extra instructions here" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.prompt).toContain("You work ALONE")
+      expect(result.prompt).toContain("Extra instructions here")
+    })
+  })
+
+  describe("defaults", () => {
+    test("uses default model when no override", () => {
+      // given
+      const override = {}
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.model).toBe(SISYPHUS_JUNIOR_DEFAULTS.model)
+    })
+
+    test("uses default temperature when no override", () => {
+      // given
+      const override = {}
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.temperature).toBe(SISYPHUS_JUNIOR_DEFAULTS.temperature)
+    })
+  })
+
+  describe("disable semantics", () => {
+    test("disable: true causes override block to be ignored", () => {
+      // given
+      const override = {
+        disable: true,
+        model: "openai/gpt-5.2",
+        temperature: 0.9,
+      }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then - defaults should be used, not the overrides
+      expect(result.model).toBe(SISYPHUS_JUNIOR_DEFAULTS.model)
+      expect(result.temperature).toBe(SISYPHUS_JUNIOR_DEFAULTS.temperature)
+    })
+  })
+
+  describe("constrained fields", () => {
+    test("mode is forced to subagent", () => {
+      // given
+      const override = { mode: "primary" as const }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.mode).toBe("subagent")
+    })
+
+    test("prompt override is ignored (discipline text preserved)", () => {
+      // given
+      const override = { prompt: "Completely new prompt that replaces everything" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.prompt).toContain("You work ALONE")
+      expect(result.prompt).not.toBe("Completely new prompt that replaces everything")
+    })
+  })
+
+  describe("tool safety (task/delegate_task blocked, call_omo_agent allowed)", () => {
+    test("task and delegate_task remain blocked, call_omo_agent is allowed via tools format", () => {
+      // given
+      const override = {
+        tools: {
+          task: true,
+          delegate_task: true,
+          call_omo_agent: true,
+          read: true,
+        },
+      }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      const tools = result.tools as Record<string, boolean> | undefined
+      const permission = result.permission as Record<string, string> | undefined
+      if (tools) {
+        expect(tools.task).toBe(false)
+        expect(tools.delegate_task).toBe(false)
+        // call_omo_agent is NOW ALLOWED for subagents to spawn explore/librarian
+        expect(tools.call_omo_agent).toBe(true)
+        expect(tools.read).toBe(true)
+      }
+      if (permission) {
+        expect(permission.task).toBe("deny")
+        expect(permission.delegate_task).toBe("deny")
+        // call_omo_agent is NOW ALLOWED for subagents to spawn explore/librarian
+        expect(permission.call_omo_agent).toBe("allow")
+      }
+    })
+
+    test("task and delegate_task remain blocked when using permission format override", () => {
+      // given
+      const override = {
+        permission: {
+          task: "allow",
+          delegate_task: "allow",
+          call_omo_agent: "allow",
+          read: "allow",
+        },
+      } as { permission: Record<string, string> }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override as Parameters<typeof createSisyphusJuniorAgentWithOverrides>[0])
+
+      // then - task/delegate_task blocked, but call_omo_agent allowed for explore/librarian spawning
+      const tools = result.tools as Record<string, boolean> | undefined
+      const permission = result.permission as Record<string, string> | undefined
+      if (tools) {
+        expect(tools.task).toBe(false)
+        expect(tools.delegate_task).toBe(false)
+        expect(tools.call_omo_agent).toBe(true)
+      }
+      if (permission) {
+        expect(permission.task).toBe("deny")
+        expect(permission.delegate_task).toBe("deny")
+        expect(permission.call_omo_agent).toBe("allow")
+      }
+    })
+  })
+
+  describe("prompt composition", () => {
+    test("base prompt contains discipline constraints", () => {
+      // given
+      const override = {}
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.prompt).toContain("Sisyphus-Junior")
+      expect(result.prompt).toContain("You work ALONE")
+    })
+
+    test("Claude model uses default prompt with BLOCKED ACTIONS section", () => {
+      // given
+      const override = { model: "anthropic/claude-sonnet-4-5" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.prompt).toContain("BLOCKED ACTIONS")
+      expect(result.prompt).not.toContain("<blocked_actions>")
+    })
+
+    test("GPT model uses GPT-optimized prompt with blocked_actions section", () => {
+      // given
+      const override = { model: "openai/gpt-5.2" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      expect(result.prompt).toContain("<blocked_actions>")
+      expect(result.prompt).toContain("<output_verbosity_spec>")
+      expect(result.prompt).toContain("<scope_and_design_constraints>")
+    })
+
+    test("prompt_append is added after base prompt", () => {
+      // given
+      const override = { prompt_append: "CUSTOM_MARKER_FOR_TEST" }
+
+      // when
+      const result = createSisyphusJuniorAgentWithOverrides(override)
+
+      // then
+      const baseEndIndex = result.prompt!.indexOf("Dense > verbose.")
+      const appendIndex = result.prompt!.indexOf("CUSTOM_MARKER_FOR_TEST")
+      expect(baseEndIndex).not.toBe(-1)
+      expect(appendIndex).toBeGreaterThan(baseEndIndex)
+    })
+  })
+})
+
+describe("getSisyphusJuniorPromptSource", () => {
+  test("returns 'gpt' for OpenAI models", () => {
+    // given
+    const model = "openai/gpt-5.2"
+
+    // when
+    const source = getSisyphusJuniorPromptSource(model)
+
+    // then
+    expect(source).toBe("gpt")
+  })
+
+  test("returns 'gpt' for GitHub Copilot GPT models", () => {
+    // given
+    const model = "github-copilot/gpt-4o"
+
+    // when
+    const source = getSisyphusJuniorPromptSource(model)
+
+    // then
+    expect(source).toBe("gpt")
+  })
+
+  test("returns 'default' for Claude models", () => {
+    // given
+    const model = "anthropic/claude-sonnet-4-5"
+
+    // when
+    const source = getSisyphusJuniorPromptSource(model)
+
+    // then
+    expect(source).toBe("default")
+  })
+
+  test("returns 'default' for undefined model", () => {
+    // given
+    const model = undefined
+
+    // when
+    const source = getSisyphusJuniorPromptSource(model)
+
+    // then
+    expect(source).toBe("default")
+  })
+})
+
+describe("buildSisyphusJuniorPrompt", () => {
+  test("GPT model prompt contains GPT-5.2 specific sections", () => {
+    // given
+    const model = "openai/gpt-5.2"
+
+    // when
+    const prompt = buildSisyphusJuniorPrompt(model, false)
+
+    // then
+    expect(prompt).toContain("<identity>")
+    expect(prompt).toContain("<output_verbosity_spec>")
+    expect(prompt).toContain("<scope_and_design_constraints>")
+    expect(prompt).toContain("<tool_usage_rules>")
+  })
+
+  test("Claude model prompt contains Claude-specific sections", () => {
+    // given
+    const model = "anthropic/claude-sonnet-4-5"
+
+    // when
+    const prompt = buildSisyphusJuniorPrompt(model, false)
+
+    // then
+    expect(prompt).toContain("<Role>")
+    expect(prompt).toContain("<Critical_Constraints>")
+    expect(prompt).toContain("BLOCKED ACTIONS")
+  })
+
+  test("useTaskSystem=true includes Task_Discipline for GPT", () => {
+    // given
+    const model = "openai/gpt-5.2"
+
+    // when
+    const prompt = buildSisyphusJuniorPrompt(model, true)
+
+    // then
+    expect(prompt).toContain("<task_discipline_spec>")
+    expect(prompt).toContain("TaskCreate")
+  })
+
+  test("useTaskSystem=false includes Todo_Discipline for Claude", () => {
+    // given
+    const model = "anthropic/claude-sonnet-4-5"
+
+    // when
+    const prompt = buildSisyphusJuniorPrompt(model, false)
+
+    // then
+    expect(prompt).toContain("<Todo_Discipline>")
+    expect(prompt).toContain("todowrite")
+  })
+})
@@ -0,0 +1,121 @@
+/**
+ * Sisyphus-Junior - Focused Task Executor
+ *
+ * Executes delegated tasks directly without spawning other agents.
+ * Category-spawned executor with domain-specific configurations.
+ *
+ * Routing:
+ * 1. GPT models (openai/*, github-copilot/gpt-*) -> gpt.ts (GPT-5.2 optimized)
+ * 2. Default (Claude, etc.) -> default.ts (Claude-optimized)
+ */
+
+import type { AgentConfig } from "@opencode-ai/sdk"
+import type { AgentMode } from "../types"
+import { isGptModel } from "../types"
+import type { AgentOverrideConfig } from "../../config/schema"
+import {
+  createAgentToolRestrictions,
+  type PermissionValue,
+} from "../../shared/permission-compat"
+
+import { buildDefaultSisyphusJuniorPrompt } from "./default"
+import { buildGptSisyphusJuniorPrompt } from "./gpt"
+
+export { buildDefaultSisyphusJuniorPrompt } from "./default"
+export { buildGptSisyphusJuniorPrompt } from "./gpt"
+
+const MODE: AgentMode = "subagent"
+
+// Core tools that Sisyphus-Junior must NEVER have access to
+// Note: call_omo_agent is ALLOWED so subagents can spawn explore/librarian
+const BLOCKED_TOOLS = ["task", "delegate_task"]
+
+export const SISYPHUS_JUNIOR_DEFAULTS = {
+  model: "anthropic/claude-sonnet-4-5",
+  temperature: 0.1,
+} as const
+
+export type SisyphusJuniorPromptSource = "default" | "gpt"
+
+/**
+ * Determines which Sisyphus-Junior prompt to use based on model.
+ */
+export function getSisyphusJuniorPromptSource(model?: string): SisyphusJuniorPromptSource {
+  if (model && isGptModel(model)) {
+    return "gpt"
+  }
+  return "default"
+}
+
+/**
+ * Builds the appropriate Sisyphus-Junior prompt based on model.
+ */
+export function buildSisyphusJuniorPrompt(
+  model: string | undefined,
+  useTaskSystem: boolean,
+  promptAppend?: string
+): string {
+  const source = getSisyphusJuniorPromptSource(model)
+
+  switch (source) {
+    case "gpt":
+      return buildGptSisyphusJuniorPrompt(useTaskSystem, promptAppend)
+    case "default":
+    default:
+      return buildDefaultSisyphusJuniorPrompt(useTaskSystem, promptAppend)
+  }
+}
+
+export function createSisyphusJuniorAgentWithOverrides(
+  override: AgentOverrideConfig | undefined,
+  systemDefaultModel?: string,
+  useTaskSystem = false
+): AgentConfig {
+  if (override?.disable) {
+    override = undefined
+  }
+
+  const model = override?.model ?? systemDefaultModel ?? SISYPHUS_JUNIOR_DEFAULTS.model
+  const temperature = override?.temperature ?? SISYPHUS_JUNIOR_DEFAULTS.temperature
+
+  const promptAppend = override?.prompt_append
+  const prompt = buildSisyphusJuniorPrompt(model, useTaskSystem, promptAppend)
+
+  const baseRestrictions = createAgentToolRestrictions(BLOCKED_TOOLS)
+
+  const userPermission = (override?.permission ?? {}) as Record<string, PermissionValue>
+  const basePermission = baseRestrictions.permission
+  const merged: Record<string, PermissionValue> = { ...userPermission }
+  for (const tool of BLOCKED_TOOLS) {
+    merged[tool] = "deny"
+  }
+  merged.call_omo_agent = "allow"
+  const toolsConfig = { permission: { ...merged, ...basePermission } }
+
+  const base: AgentConfig = {
+    description: override?.description ??
+      "Focused task executor. Same discipline, no delegation. (Sisyphus-Junior - OhMyOpenCode)",
+    mode: MODE,
+    model,
+    temperature,
+    maxTokens: 64000,
+    prompt,
+    color: override?.color ?? "#20B2AA",
+    ...toolsConfig,
+  }
+
+  if (override?.top_p !== undefined) {
+    base.top_p = override.top_p
+  }
+
+  if (isGptModel(model)) {
+    return { ...base, reasoningEffort: "medium" } as AgentConfig
+  }
+
+  return {
+    ...base,
+    thinking: { type: "enabled", budgetTokens: 32000 },
+  } as AgentConfig
+}
+
+createSisyphusJuniorAgentWithOverrides.mode = MODE