Expand description
Headless runtime core.
Runtime projects session events into ephemeral short memory, resolves
workspace-scoped long memory, orchestrates model/tool turns, and normalizes
runner execution output into canonical protocol events. Session identity,
scheduling, event persistence,
and event sequencing belong to arabica-session.
Structs§
- Acceptance
Criterion - Acceptance
Result - Acceptance
Spec - Accounting
Coverage - Verification metrics regarding billing completeness.
- Accounting
Error - Archived
Memory - Workspace-scoped durable memory.
- Auto
Hydration Observation - One exact archive read automatically triggered by a repeated tool call.
- Available
Model Catalog - Catalog of available models for bounded proposal exploration.
- Blend
Evaluation Report - Blend
Model Evaluation - Blend
Policy Evaluation - Blend
Routing Policy - Deterministic per-run routing rules. Each configured alias must exist in the ModelProvider supplied to this runtime.
- Blend
Tool Route - A route selected after a completed tool batch.
- Call
Cost Breakdown - Fine-grained token and dollar breakdown for a single model call.
- Call
Cost Record - Detailed accounting record for one model call.
- Candidate
Aggregate Report - Comprehensive aggregate evaluation metrics for one policy candidate.
- Category
Weights - Predeclared category weights used to aggregate multi-category trial scores.
- Context
Evaluation Data - Context
Evaluation Group - Context
Evaluation Result - Context
Item Metrics - Context
Policy - No policy preserves legacy exposure. An explicit policy selects only its sets and includes; source disables and exclusions always take precedence.
- Context
Report - Core
Runtime - Criterion
Evaluation - Outcome of evaluating an individual criterion.
- Decay
Rule - Deterministic
Compaction Projection - Deterministic, provider-free evidence used by the Tier-A PGC/FBGC benchmark. Production projection code is reused directly; this wrapper only supplies a frozen fallback economic envelope and exposes diagnostics.
- Deterministic
Compaction Timing - Timings for the benchmark-only deterministic compaction audit path.
- Evaluation
Config - Evaluation
Error - Evaluation
Evidence - No prompts, arguments, provider envelopes, results, or failure messages. Run IDs are namespaced by workspace/session to avoid cross-session joins.
- Evaluation
Identity - Composite evaluation identity tying checker version, task, and scoring profile.
- Evaluation
Plugin Metadata - Evaluation
Registry - Event
Batch - A stable runtime grouping over immutable events.
- Event
Memory Traits - Relationship metadata used for precise decay. A completion affects only an older event with the same relation key, such as one matching tool call.
- Event
Ttl - Event
Visibility Decision - Explainable decision for one immutable source event.
- Evidence
Acceptance Scorer - File
Archive Store - Atomic filesystem archive. Each pointer is stored as one JSON envelope
beneath
root, preserving the normalized memory-id hierarchy. - Gate
Check Evaluation - Generated
Policy Record - A policy version candidate or active definition generated by background or offline evaluation.
- Historical
Trace Config - Configuration for historical offline trace rescoring.
- KeyAdmission
Policy - Hard limits for historical
LOAD_KEYmaterialisation.Noneis unbounded. Limits apply to UTF-8 key-content bytes, excluding fullLOAD_ALLitems. - KeyAdmission
Summary - Live
Session Config - Configuration for real live interactive session evaluation.
- Long
Memory Error - Long
Memory Manager - Model
Evaluation Data - Model
Evaluation Group - Model
Evaluation Result - Model
Price Rule - A pricing rule for a specific model or model family. Rates are specified per million tokens (standard LLM pricing unit).
- Observed
Tool - Paired
Comparison - Paired comparison metrics between a candidate and baseline across matched tasks.
- Permission
Decision - A decision about one call.
- Permission
Request - One call awaiting a decision, sent to the host’s approver.
- Pointer
GcAdmission Observation - One cache-aware PointerGC admission decision made before a Provider call.
- Policy
Comparison Summary - Policy
Comparison Trial - Policy
Proposal - A bounded candidate proposal derived from evaluated evidence.
- Price
Card - Versioned price card mapping models/aliases to billing rules in a given currency.
- Proposal
Evidence Basis - Provenance and metrics that justified generating this candidate.
- Recommendation
Gates Config - Predeclared gates required to recommend promotion of a candidate over baseline.
- Recommendation
Report - RunAcceptance
- RunCancellation
- A cancellation signal for one run, shared between the host and Runtime.
- RunControl
- Everything a host may attach to one run.
- RunCost
Accounting - Complete, reproducible run cost and time accounting report.
- RunEvaluation
- RunEvidence
Snapshot - A bounded projection of already-persisted facts. No live file reads, prompts, command arguments, provider envelopes, or raw tool output enter this snapshot.
- RunTiming
- Runtime
Error - Runtime
Session - Scoring
Observation - Observed runtime metrics and evaluator conditions for scoring a task run.
- Scoring
Profile - Frozen scoring profile with formula version, component weights, and references.
- Scoring
Weights - Relative weights for scoring formula components summing to 1.0.
- Short
Memory Materialization - Complete output of short-memory materialisation. The source event slice is never modified; decisions and batches are returned for replay diagnostics.
- Short
Memory Policy - Declarative, replayable policy for event visibility and batch loading.
- Short
Memory Projector - Pure projection of a Session’s immutable event history for the next model turn. TTL and batch policies alter prompt visibility, never the audit log.
- Skill
Definition - A host-loaded immutable instruction bundle. Resource keys are relative names; hosts validate paths and capture content before handing it to Runtime.
- Sqlite
Archive Store - SQLite archive with an idempotent primary-key contract.
- Task
Resource Budget - Resource budget limits for a task trial.
- Task
Score - Evaluated task score and components.
- Task
Spec - Versioned task specification with criteria, weights, and resource budgets.
- Task
Trial Score - A score record for one candidate on a specific task and repetition.
- Time
Accounting - Bounded time metrics separating active work from user/permission approval wait.
- Tool
Argument Permission Rule - Tool
Permission Gate - The permission gate a host attaches to a run.
- Tool
Permission Policy - Which rule applies to each tool, keyed by tool name.
- Trajectory
Decision Observation - Trajectory
Routing Feedback - Trajectory-level feedback associating run scores with routing decisions.
- Workspace
Gate Config - Configuration for workspace integrity and quality gates on real projects.
- Workspace
Gate Result - Result of evaluating real workspace integrity gates.
Enums§
- Acceptance
Check - Acceptance
Status - Artifact
Evidence - Blend
Tool Matcher - Call
Category - Behavioral role of a model invocation.
- Decay
Match - Evaluation
Diagnostic - Evaluation
Mode - Unified evaluation mode distinguishing live sessions, workspace gates, preset benchmarks, and historical audits.
- History
Projection - How Runtime represents runs OTHER than the one currently executing.
- KeyAdmission
Decision - Explainable outcome of applying the key budget to one batch.
- Long
Memory Error Kind - Memory
Class - Small retention taxonomy used by Runtime. Protocol events remain detailed audit facts and are projected into one of these policy classes.
- Pointer
GcAdmission Policy - Selects whether PointerGC is admitted by the production profitability gate or by the preregistered mechanism-qualification gate. The qualification mode is explicit evidence-generation configuration; it must not be used for economic-effect estimates.
- Policy
Lifecycle Status - Lifecycle state for an immutable routing policy definition.
- Policy
Parameter Delta - A specific parameter modification within a proposed policy candidate.
- Recommendation
Decision - Runtime
Archive Store - Runtime
Compaction Strategy - Runtime-owned context compaction mode.
- Runtime
Error Kind - Task
Score Status - Scoring outcome status distinguishing completed, pending, incomplete, and fault states.
- Tool
Permission Rule - What the gate does with a call to one tool.
Constants§
- AUTO_
COMPLETION_ REQUIRED_ MESSAGE - CREATE_
PLAN_ TOOL_ NAME - RUNTIME_
COMPLETE_ TOOL_ NAME - UPDATE_
PLAN_ TOOL_ NAME
Traits§
- Context
Evaluation Strategy - Long
Memory Store - Durable backing contract for recoverable short-memory pointers.
- Model
Evaluation Strategy - Pointer
GcObservation Sink - Optional best-effort telemetry sink for incremental PointerGC decisions.
- RunAcceptance
Scorer - Runtime
Engine - Runtime
Event Log - Skill
Source - Hosts own discovery and I/O; Runtime only pins the returned immutable bundle.
Functions§
- account_
run_ cost_ and_ time - Compute complete cost and time accounting from an event history and price card. Duplicate/replayed observations are identified and suppressed from billing.
- aggregate_
candidate_ scores - Aggregate candidate scores using predeclared category weights.
- calculate_
task_ score - Pure deterministic calculation of a versioned task score.
- compare_
policy_ trials - Compare actual trials on identical task/specification/repetition sets. Reject duplicate runs, mixed policy versions, and incomparable cohorts.
- compute_
paired_ comparison - Compute matched paired comparisons and cluster-robust standard errors across repetitions.
- context_
report - evaluate_
blend_ history - Rebuild call and trajectory measurements from canonical persisted events.
- evaluate_
live_ session - Evaluate a real running session from its evidence snapshot and live configuration.
- evaluate_
live_ session_ from_ history - Fully evaluate a live session directly from its event history, including automatic snapshot extraction and price-card cost accounting.
- evaluate_
recommendation - Evaluate candidate against baseline across predeclared recommendation gates.
- evolve_
blend_ policy - Propose an evolved routing policy candidate given historical evaluation evidence and reports. This derives an updated candidate version without mutating currently active or running session policies.
- exact_
transcript_ entries - Reconstruct short memory for every run OTHER than
current_run_idby replaying each run’s own typed model exchange and tool-call Events, instead of expanding TTL/batch-projected event batches. This is whatcrate::HistoryProjection::ExactTranscriptsubstitutes for theshort_memoryhalf ofShortMemoryProjector::materialize_for_model_step’s output; the current run’s own exchange is unaffected and continues to be reconstructed separately byexact_run_continuation. - execute_
controlled_ test - Execute a controlled test command in an isolated, bounded environment.
- execute_
workspace_ gate - Execute workspace integrity checks on a real project workspace.
- generate_
bounded_ proposals - Generate bounded candidate policy proposals informed by evaluation evidence.
- mcp_
context_ identity - JSON tuple encoding avoids ambiguity from separators or sanitized names.
- mcp_
server_ context_ id - memory_
read_ definition - memory_
search_ definition - plan_
tool_ definitions - project_
compaction_ for_ benchmark - run_
evidence_ snapshots - Build snapshots only for terminal runs, stopping at their first terminal event. Later conversation turns cannot change the evidence being graded.
- runtime_
complete_ tool_ definition - score_
run_ evaluation - Score a completed run evaluation against a task specification and profile.
- tool_
context_ identity