Summary

• Add a dedicated system prompt for GLM models (glm.txt).
• Wire GLM models to use this new prompt instead of the generic Qwen/fallback prompt.
  Background
  Previously, SystemPrompt.provider did not have any GLM-specific handling.
  Any model whose ID did not match the GPT, Gemini, Claude, or Polaris branches fell through to the default:
  return [PROMPT_ANTHROPIC_WITHOUT_TODO] // imported from "./prompt/qwen.txt"
  As a result, GLM sessions were using the qwen.txt system prompt rather than a GLM-optimized one.
  Changes

1. New GLM prompt
   • Added packages/opencode/src/session/prompt/glm.txt.
   • This prompt:
     • Positions opencode as an interactive CLI assistant.
     • Emphasizes GLM-4.6’s reasoning and analytical strengths.
     • Reinforces security constraints (no malicious code).
     • Encourages structured workflows using TodoWrite, WebFetch, and the existing toolchain.
     • Defines concise communication style and guidance for help/feedback.
2. System prompt routing for GLM

   • Updated packages/opencode/src/session/system.ts:

     • Imported the new prompt:
       import PROMPT_GLM from "./prompt/glm.txt"

     • Adjusted SystemPrompt.provider to detect GLM models and return the GLM prompt

   • Matching is done on "glm" (case-insensitive) rather than a specific version, so all GLM variants benefit from the new prompt.
     Behavior / Impact

• GLM-based sessions now receive GLM-specific system instructions tailored to:
  • Software engineering workflows in the opencode environment.
  • Stronger task planning, research, and tool usage guidance.
• Other providers and models retain their existing prompts:
  • GPT-* and o1/o3 → beast.txt / codex.txt as before.
  • Gemini → gemini.txt
  • Claude → anthropic.txt
  • Polaris → polaris.txt
  • All other non-GLM, non-GPT/Gemini/Claude/Polaris models → qwen.txt (unchanged).

Testing

• bun turbo typecheck (runs typecheck across all packages) – pass.
• Manual verification recommended:
  • Start a session with a GLM model (model ID contains "glm").
    • Confirm that behavior aligns with the new GLM system prompt (e.g., references to TodoWrite, WebFetch, and GLM’s reasoning capabilities).
  • Start a session with non-GLM models and confirm their prompts are unchanged.