RunRequest: add effort (low|medium|high|xhigh|max)

Mirrors the system_mode wiring: optional field on the wire,
maps to 'claude -p --effort <level>'. None lets the CLI default
decide. Right for prose-craft tasks (skald) where the author
persona benefits from extra thinking budget.

clawdforge/runner.py

@@ -34,6 +34,7 @@ class Runner:
         model: str | None = None,
         system: str | None = None,
         system_mode: str | None = "append",
+        effort: str | None = None,
         files: list[str] | None = None,
         timeout_secs: int | None = None,
     ) -> RunResult:
@@ -60,6 +61,8 @@ class Runner:
                 cmd += ["--system-prompt", system]
             else:
                 cmd += ["--append-system-prompt", system]
+        if effort:
+            cmd += ["--effort", effort]
         if files:
             for f in files:
                 cmd += ["--files", f]

clawdforge/server.py

@@ -122,6 +122,12 @@ class RunRequest(BaseModel):
     # Default is "append" for back-compat — every caller before this
     # field landed used append semantics implicitly.
     system_mode: Literal["append", "replace"] | None = "append"
+    # claude -p --effort <level>. Controls extended-thinking budget on the
+    # reasoning models. None = let claude default decide (today: medium).
+    # Right for prose-craft tasks where the model benefits from extra
+    # thinking; e.g. skald passes "max" on gen + cleanup so author personas
+    # get the budget they deserve.
+    effort: Literal["low", "medium", "high", "xhigh", "max"] | None = None
     files: list[str] | None = None
     timeout_secs: int | None = Field(default=None, ge=5, le=600)
 
@@ -203,6 +209,7 @@ def run(
         model=body.model,
         system=body.system,
         system_mode=body.system_mode,
+        effort=body.effort,
         files=file_paths or None,
         timeout_secs=body.timeout_secs,
     )

clients/rust/src/lib.rs

@@ -53,6 +53,6 @@ pub use client::{Client, ClientBuilder};
 pub use error::Error;
 pub use session::{Session, SessionList, SessionOptions, SessionState, TurnEvent, TurnResult};
 pub use types::{
-    AppToken, AppTokenInfo, FileToken, Healthz, RunFailure, RunRequest, RunResult,
+    AppToken, AppTokenInfo, Effort, FileToken, Healthz, RunFailure, RunRequest, RunResult,
     SystemMode, TokenCreateRequest, TokenList,
 };

clients/rust/src/types.rs

@@ -26,6 +26,30 @@ impl Default for SystemMode {
     }
 }
 
+/// Maps to `claude -p --effort <level>`. Controls the extended-thinking
+/// budget on the reasoning models. `None` on [`RunRequest::effort`] means
+/// "let the CLI default decide" (today: medium).
+///
+/// Use [`Effort::Max`] for prose-craft tasks (fiction generation,
+/// long-form rewrite) where the author persona benefits from extra
+/// thinking. Leave default for tool-use / structured-JSON tasks.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum Effort {
+    /// `--effort low` — minimum thinking budget.
+    Low,
+    /// `--effort medium` — claude's default if no flag is passed.
+    Medium,
+    /// `--effort high`.
+    High,
+    /// `--effort xhigh`.
+    Xhigh,
+    /// `--effort max` — maximum thinking budget. Right for long-form
+    /// prose generation where the author persona benefits from extra
+    /// thinking before each paragraph.
+    Max,
+}
+
 /// `GET /healthz` response body.
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct Healthz {
@@ -80,6 +104,11 @@ pub struct RunRequest {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub system_mode: Option<SystemMode>,
 
+    /// Maps to `claude -p --effort <level>`. `None` lets the CLI default
+    /// decide. Use [`Effort::Max`] on prose-craft / long-form generation.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub effort: Option<Effort>,
+
     /// File tokens previously returned from [`Client::upload_file`].
     ///
     /// [`Client::upload_file`]: crate::Client::upload_file