Struct ExternalLoopConfig 

pub struct ExternalLoopConfig {
    pub max_turns: Option<u32>,
    pub timeout_secs: Option<u64>,
    pub repair_invokes: u32,
    pub transient_retries: u32,
    pub model: Option<String>,
    pub deadline: Arc<SessionDeadline>,
    pub baseline_captures: BaselineCaptures,
}

Tuning for external delegation.

Fields

max_turns: Option<u32>

Per-invocation model-turn cap (maps to the CLI’s --max-turns).

timeout_secs: Option<u64>

Per-invocation wall-clock budget (runner clamps to 1h).

repair_invokes: u32

Fresh repair invocations after a red first pass (the stream-json protocol has no session resume yet, so repairs re-state the task plus the failing-check output).

This is the hypothesis budget: each one buys another attempt at being right. It is deliberately not spent on transport failures — see transient_retries.

Defaults to 2 because recurrence escalation needs it. Round 1 establishes a failure signature, round 2 is the first that can repeat it, and only round 3 can be told it did. At 1 the loop still detects the repeat, but the session ends before the feedback carrying that news reaches anyone — and since rpc is the only construction site and takes ..Default::default(), a default of 1 made the escalation unreachable in every shipped configuration.

The cost is smaller than it looks: worst-case invocations are max_hypotheses + transient_retries, so this moves 3 -> 4, about +33%, and only on sessions that are already failing.

transient_retries: u32

Re-invocations after the CLI process itself died mid-run (timeout or I/O) with the contract still red.

Separate from repair_invokes because it buys a different thing: an availability retry, not a new hypothesis. Sharing one counter means a single flaky timeout eats a replan the coder needed for an actual hypothesis — the difference between a session that recovers and one that silently gives up on a task it was about to finish.

model: Option<String>

Pin the external CLI’s backbone (coder.start’s model). None = the CLI’s own default.

The paired A/B only measures the harness when both arms share a backbone; this is the external half of that invariant (the native half is NativeLoopConfig.model). Before this existed the pin reached only the native loop, so “both arms on gpt-5.5” was an unverified assumption.

deadline: Arc<SessionDeadline>

The session’s absolute deadline, SHARED with every other rung of the fallback ladder. See super::budget for why this is a handle and not a value.

baseline_captures: BaselineCaptures

The session-start baseline captures differential checks compare against (car#1067). Empty when the contract declares none; differential checks fail closed under an empty map.

Trait Implementations

impl Clone for ExternalLoopConfig

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ExternalLoopConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ExternalLoopConfig

fn default() -> Self

Returns the “default value” for a type.