Skip to main content

AgentXchain v2.22.0

2.22.0 adds real cost guardrails and escalation recovery proof to the governed workflow.

2.21.0 closed front-door truth gaps. 2.22.0 closes governance depth gaps: what happens when a run exceeds its budget, and what happens when an operator escalates and needs to recover.

Upgrade:

npm install -g [email protected]

What changed since 2.21.0​

Budget enforcement is now shipped​

Governed runs can now set per_run_max_usd in agentxchain.json and the runtime will enforce it:

  • Post-acceptance exhaustion: when a turn's actual cost pushes remaining_usd to zero or below and on_exceed is pause_and_escalate, the run transitions to blocked with budget_exhausted category. The exhausting turn is accepted — work is never discarded.
  • Pre-assignment guard: if the budget is already exhausted, new turn assignment is rejected before dispatch.
  • Per-turn overrun warning: when actual cost exceeds the reservation amount, the acceptance result includes a budget_warning (advisory only).

This matters because per_run_max_usd and on_exceed: 'pause_and_escalate' existed in the scaffolded config since project inception, but nothing in the runtime ever read them. That was a shipped lie. Now it is enforced.

Budget recovery works through the real CLI​

When a run is blocked by budget_exhausted:

  1. Edit agentxchain.json to raise per_run_max_usd
  2. Run agentxchain status — budget is reconciled from the config and shows resume-ready guidance when headroom is sufficient
  3. Run agentxchain resume — assigns the next turn and dispatches

The recovery path is proven through real CLI subprocess execution, not unit-level state manipulation.

Escalation recovery is proven end to end​

The escalate → status → resume cycle now has subprocess-level proof for both paths:

  • Retained-turn escalation: escalate blocks the run, resume re-dispatches the same retained turn with the same turn_id
  • Run-level escalation: escalate after clearing the turn, resume assigns a new turn with a new turn_id

Both paths verify that the decision ledger contains operator_escalated and escalation_resolved entries.

Escalation guidance is now runtime-aware​

Recovery action strings now vary by runtime type:

  • Retained manual turn → agentxchain resume (non-blocking redispatch)
  • Retained non-manual turn → agentxchain step --resume (dispatches and waits)
  • Run-level escalation → agentxchain resume
  • Targeted multi-turn escalation → appends --turn <id>

Stale recovery actions from pre-2.22.0 blocked states are reconciled at load time to the correct command.

OpenAI cost rates are shipped​

The api_proxy adapter now has built-in cost rates for 8 OpenAI models: gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o3, o3-mini, and o4-mini. Previously, every OpenAI governed turn reported $0 cost with no indication that cost tracking was unsupported.

Why this release matters​

AgentXchain claims governed delivery over long horizons. That requires cost control and operator intervention surfaces that actually work, not just config scaffolds that look like they do something.

2.22.0 matters because:

  • budget enforcement closes a gap between the shipped config and the runtime behavior
  • budget recovery is proven through the real operator path, not just the blocked state
  • escalation recovery is proven through real subprocess execution for both retained and cleared turn paths
  • recovery guidance is runtime-aware instead of always pointing operators at commands that may hang

Where to go next​

  • See Recovery for the operator recovery map including budget_exhausted and operator_escalation.
  • See Adapters for the full list of models with built-in cost tracking.
  • Start with the Quickstart if this is your first governed run.

Evidence​

  • 2394 node tests / 514 suites, 0 failures.
  • Docusaurus production build passes.