Feature hasn't been suggested before.
Describe the enhancement you want to request
1. Custom instructions (free text, like Codex instructions). Today personalization only exists as files (instructions: string[] paths/URLs plus AGENTS.md/CLAUDE.md). There is no free-text field for role, machine tools, and way of working that applies to every new chat. Proposed: customInstructions?: string in opencode.json[c] (global + project, concatenated), injected into the system prompt of each new session after the env block and before AGENTS.md, with a settings editor.
2. Environment-aware shell context. The V2 <env> block only carries cwd, workspace root, git yes/no, Platform and date — the model never learns the effective shell. On Windows it then emits Linux-isms (ls, &&, grep) that break under cmd.exe/PowerShell 5.1. Proposed: add human-readable OS and resolved effective Shell (reusing the existing shell discovery) to <env> so generated commands match the runtime. Related: #42733, #41321, #22141, #35552.
Feature hasn't been suggested before.
Describe the enhancement you want to request
1. Custom instructions (free text, like Codex instructions). Today personalization only exists as files (
instructions: string[]paths/URLs plus AGENTS.md/CLAUDE.md). There is no free-text field for role, machine tools, and way of working that applies to every new chat. Proposed:customInstructions?: stringin opencode.json[c] (global + project, concatenated), injected into the system prompt of each new session after the env block and before AGENTS.md, with a settings editor.2. Environment-aware shell context. The V2
<env>block only carries cwd, workspace root, git yes/no,Platformand date — the model never learns the effective shell. On Windows it then emits Linux-isms (ls,&&,grep) that break under cmd.exe/PowerShell 5.1. Proposed: add human-readableOSand resolved effectiveShell(reusing the existing shell discovery) to<env>so generated commands match the runtime. Related: #42733, #41321, #22141, #35552.