5.6. Final harness audit

This page records the production profile of the harness: what is on by default, what is optional, what is experimental. It describes the implementation, not a promise of model quality.

The agent harness in one picture: layers, roles and boundaries

Fig. 5.6 The whole harness on one sheet — the layers, what is on by default, what is optional, what is experimental, and where the security boundary runs. Architecture shows how the request classes route through it.

Every box in the picture can be checked against the tree.

5.6.1. Production maturity

CORE / DEFAULT ON

  • request classification and routing (TaskController, answer_scope)

  • the coding core and SpearHost, its control plane

  • the evidence plane: canonical evidence, source epochs, the implementation verdict and project validation

  • the MIXED orchestration: constraint packet, structural coverage, applicability, source predicates and bound conformance checks

  • AgentRuntime, the normative runtime (GENERAL requests run on the coding core, routed by task_controller.py)

  • WorkingState and ContextEngine

  • transactional semantic compaction

  • workspace knowledge (KnowledgeStore)

  • ToolRegistry / ToolRouter / ResultStore

  • SessionStore and cancellation

  • VerificationPolicy and CheckpointManager

  • BudgetManager, the failure and retry policy (failure_policy.py: FailureKind, RetryPolicy, RetryDecision) and ProgressMonitor

  • role-aware tool exposure

OPTIONAL

  • deterministic planning when PlanningPolicy identifies a complex task

  • runtime tracing, when SPEAR_TRACE turns it on (tracing.py)

  • external capabilities, when a deployment registers MCP providers (capability_gateway.py)

  • compatibility code-block mutation fallback

  • legacy chat_once and execute_tool APIs

EXPERIMENTAL / DEFAULT OFF

  • Explorer

  • Reviewer

  • reviewer-driven repair

OPERATOR-ONLY, OUTSIDE THE TURN

  • the training subsystem — capture, curation, governance, readiness, frozen bundles, and the launcher. Nothing in it is model-visible and freezing runs no training; its own readiness vocabulary tops out at READY_FOR_EXPERIMENT (Training and fine-tuning).

Explorer and Reviewer remain isolated, read-only roles and can be enabled by an explicit TaskRequest or benchmark configuration. They are not constructed, budgeted or given child sessions on the normal default path, and no quality benefit is claimed for them.

5.6.2. Ownership invariants

WorkingState is task truth; SessionStore is resumable runtime state; CheckpointManager is filesystem recovery; ResultStore owns large tool evidence; KnowledgeStore owns the workspace’s knowledge; ContextEngine composes the normative runtime’s context; and BudgetManager owns agent-level budgets. Legacy projections remain only where CLI/history compatibility or weaker local models require them.

5.6.3. Security and provider boundary

All command execution still passes ToolRouter’s authorization and the CommandPolicy/CommandRunner/Bubblewrap boundary — the coding core’s included, through SpearHost. The two host commands, spear-capability and spear-knowledge, are answered by SpearHost and never reach a shell; the MCP providers behind the first run on the host, outside the sandbox, under the gateway’s read/write policy. Read-only child roles receive no checkpoint, memory-write, web or mutating-tool capability. Lower runtime modules do not import the CLI, and provider-specific protocol details remain inside backend adapters.

5.6.4. Known limitations and future candidates

The code-block fallback, chat_once/enforce_ctx_budget compatibility path and string execute_tool facade remain maintenance debt because tests and legacy callers still reach them.

ToolSpec.execution_modes is enforced by ToolRouter: a tool called in a mode its spec does not declare is denied (execution_mode_denied), and a mutating spec must declare its modes and may not declare safe. It does not filter what the registry exposes, and it adds to, rather than replaces, the authorization in CapabilityPolicy and CommandPolicy (Security model). Explorer/Reviewer contract reliability, long-context benchmark scoring and repair quality need more model evidence. ToolSearch, parallel agents, model routing, new retrieval, and new planning or review capabilities are deliberately not implemented.