AgentXchain v2.46.0

v2.46.0 ships the declarative policy engine — a first-class governance surface for turn-acceptance rules. Operators can now define policies that constrain agent behavior during governed execution, with proper escalation, recovery, and operator feedback at every layer.

Declarative Policy Engine

Policies are defined in agentxchain.json under the policies key. Each policy is a named set of rules evaluated at turn acceptance.

Five built-in rules

Rule	Description
require_status	Require specific turn status (e.g., only accepted, needs_revision)
max_consecutive_same_role	Cap consecutive turns from the same role
require_challenge	Require challenge content in turn artifacts
max_cost_per_turn	Enforce per-turn cost ceiling
min_artifacts	Require a minimum number of artifacts per turn

Three actions

Action	Behavior
warn	Log a warning, accept the turn
block	Reject the turn, do not persist state
escalate	Persist blocked state with structured recovery descriptor

Phase and role scoping

Rules can be scoped to specific phases and/or roles:

{
  "policies": {
    "qa-must-challenge": {
      "rules": [
        {
          "rule": "require_challenge",
          "action": "block",
          "phases": ["qa"],
          "roles": ["qa"]
        }
      ]
    }
  }
}

Runtime-aware escalation recovery

Policy escalations now produce the same structured blocked_reason as every other blocked-state writer in the system. This means:

• agentxchain status shows policy_escalation with the violating policy ID and the correct recovery command
• Recovery is runtime-aware: retained manual turns surface agentxchain resume; retained non-manual turns surface agentxchain step --resume
• The dashboard shows policy-blocked states with actionable recovery guidance
• Run history records policy-blocked transitions alongside all other terminal outcomes

Previously, policy escalation wrote a plain-string blocked_reason, which caused the recovery system to classify it as unknown_block.

Cost enforcement fix

max_cost_per_turn now reads turnResult.cost.usd as the primary cost field and falls back to legacy turnResult.cost.total_usd for compatibility. Previously it only read cost.total_usd, which could silently fail to enforce cost caps on runs using the current cost schema.

Operator CLI guidance

agentxchain accept-turn now renders policy-specific output for both policy_violation (block) and policy_escalation (escalate) outcomes:

• Violating policy IDs and rule names
• Violation messages with detail
• Typed recovery descriptor with owner and exact recovery command
• Turn retention status

Policy docs surface

A dedicated Policies page documents configuration, built-in rules, actions, phase/role scoping, and the recovery path. Policies appear in the docs sidebar, llms.txt, and sitemap.xml.

VS Code Marketplace readiness

Added a dedicated test guard for marketplace readiness: valid package.json, publisher field, .vscodeignore, and README. Publisher creation remains a human task.

Evidence

• 3308 tests / 698 suites / 0 failures
• Policy evaluator: 38 tests across 13 suites
• Policy escalation recovery: 7 unit + 2 runtime integration + 1 subprocess E2E
• Docs policy content: 12 guards
• Docusaurus build: clean