Struct PromptContext 

pub struct PromptContext<'a> {
    pub workspace_dir: &'a Path,
    pub agent_workspace_dir: &'a Path,
    pub model_name: &'a str,
    pub tools: &'a [Box<dyn Tool>],
    pub skills: &'a [Skill],
    pub skills_prompt_mode: SkillsPromptInjectionMode,
    pub identity_config: Option<&'a IdentityConfig>,
    pub dispatcher_instructions: &'a str,
    pub sends_native_tool_specs: bool,
    pub security_summary: Option<String>,
    pub autonomy_level: AutonomyLevel,
    pub channel_targets: Option<String>,
}

Fields

workspace_dir: &'a Path

agent_workspace_dir: &'a Path

Per-agent persona workspace (where SOUL.md / IDENTITY.md / USER.md / AGENTS.md live). Separate from workspace_dir, which is the security sandbox root and can be overridden per session by an IDE-supplied cwd. Channel-driven runs typically pass the same path for both; gateway and ACP sessions pass the agent’s own dir here while letting workspace_dir follow the session cwd.

model_name: &'a str

tools: &'a [Box<dyn Tool>]

skills: &'a [Skill]

skills_prompt_mode: SkillsPromptInjectionMode

identity_config: Option<&'a IdentityConfig>

dispatcher_instructions: &'a str

sends_native_tool_specs: bool

True when the provider request carries native tool specs. In that mode the prompt must not duplicate the same tool catalog in prose.

security_summary: Option<String>

Pre-rendered security policy summary for inclusion in the Safety prompt section. When present, the LLM sees the concrete constraints (allowed commands, forbidden paths, autonomy level) so it can plan tool calls without trial-and-error. See issue #2404.

autonomy_level: AutonomyLevel

Autonomy level from config. Controls whether the safety section includes “ask before acting” instructions. Full autonomy omits them so the model executes tools directly without simulating approval.

channel_targets: Option<String>

Pre-rendered channel target section injected so the agent knows where to deliver outbound messages via channel_send.