Structure system prompt into composable sections

[TheGreatAxios]

Summary

• Split prompt construction into composable pure section builders
• Added prompt composition coverage for agent and chat prompts

Verification

• bun run typecheck
• bun run build
• bun test (blocked: tests/unit/run-agent.test.ts exits after OPENAI_COMPATIBLE_API_KEY env var is required)

[linear-code]

CL-811 Structure system prompt into composable sections with tests

Objective

Refactor the system prompt from a monolithic string into composable sections that can be tested independently.

Requirements

• Break prompt into sections: tool-call discipline, submit rules, budget rules, plan rules
• Each section is a pure function that returns a string
• buildSystemPrompt() composes all sections into the final prompt
• buildChatSystemPrompt() composes only the conversational sections
• Test: prompt contains all required sections
• Test: prompt contains all available tool names
• Test: chat prompt does not contain submit rules

Review in Linear

[TheGreatAxios]

Code Review: System Prompt Refactoring

Summary: Breaks monolithic system prompt into composable sections. Clean separation of tool-call, submit, budget, and plan rules.

Strengths:

• Each section is a pure function returning a string — excellent testability
• buildSystemPrompt() and buildChatSystemPrompt() compose correctly
• Comprehensive test coverage: sections, composition, tool names
• Commit messages accurately describe the refactor
• No behavioral changes — functionality preserved

Items to Consider:

• The prompt sections are tested independently, but verify the composed prompt still functions correctly in the agent loop (end-to-end test?)
• Some sections (e.g., budget rules) reference specific file sizes/limits — hardcoded or configurable? Document the values
• Tool list is dynamic via parameter — good. Verify all callers pass the full tool list

GPG Signature: ✅ Verified