LLMRequestHash class abstract

Computes a stable hash of an LLM request, used as the cache key for recording / replay.

Stability requirements:

• The hash must be deterministic given the same logical request.
• It must ignore non-deterministic fields (timestamps, UUIDs, ids).
• It must be sensitive to anything that would change the model's output (messages, tools, model id, sampling parameters).

Non-determinism caveat: when trialsPerRun > 1, every trial sends the same request (same messages, same tools, same model). Without extra signal the hash collapses across trials and the cache forces every trial to receive an identical response — which silently destroys the non-determinism that pass^k / pass@k are supposed to measure (Anthropic Step 6).

To preserve per-trial variation, callers can pass trialSalt (a stable per-trial identifier such as runName/taskId#trialIndex). Each trial then computes a distinct hash and the recording store keeps a separate cache entry per trial. A null salt means "no per-trial axis" — appropriate for one-off requests, judges, or any other call that should be cached once across all trials.

Implementers

• Sha256LLMRequestHash

Constructors

LLMRequestHash()

Properties

hashCode → int

The hash code for this object.

no setter, inherited

runtimeType → Type

A representation of the runtime type of the object.

no setter, inherited

Methods

compute({required List<LLMMessage> messages, required List<Tool>? tools, required ModelConfig modelConfig, bool? jsonOutput, String? trialSalt}) → String

Compute the hash for a request.

noSuchMethod(Invocation invocation) → dynamic

Invoked when a nonexistent method or property is accessed.

inherited

toString() → String

A string representation of this object.

inherited

Operators

operator ==(Object other) → bool

The equality operator.

inherited