The `SafetySection` in `SystemPromptBuilder` always hardcoded
"Do not run destructive commands without asking" and "Do not bypass
oversight or approval mechanisms" regardless of the configured
autonomy level. This caused the gateway WebSocket path (web interface)
to instruct the LLM to simulate approval dialogs even when
`autonomy.level = "full"`.
PRs #3955/#3970/#3975 fixed the channel dispatch path
(`build_system_prompt_with_mode_and_autonomy`) but missed the
`Agent::from_config` → `SystemPromptBuilder` path used by
`gateway/ws.rs`.
Changes:
- Add `autonomy_level` field to `PromptContext`
- Rewrite `SafetySection::build()` to conditionally include/exclude
approval instructions based on autonomy level, matching the logic
already present in `build_system_prompt_with_mode_and_autonomy`
- Add `autonomy_level` field to `Agent` struct and `AgentBuilder`
- Pass `config.autonomy.level` through `Agent::from_config`
- Add tests for full/supervised autonomy safety section behavior
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
96f25ac701