Skip to main content

AgentXchain v2.24.1

2.24.1 is the corrected public credibility release.

2.23.0 made proposal authority honest. 2.24.0 was a local release-candidate tag that failed strict preflight because the release copy lost its concrete node-test counts. 2.24.1 is the corrected public cut. It closes the remaining launch-critical proof gaps by moving more claims out of theory and into executed evidence: MCP is now proven live with a real Anthropic model behind a real MCP server, retry exhaustion and operator recovery are dogfooded through the governed CLI, and release postflight now verifies the public npx install path instead of trusting registry metadata alone.

Upgrade:

npm install -g [email protected]

What changed since 2.23.0​

MCP now has real-model proof, not just transport proof​

Previous MCP evidence proved the adapter and both transports could carry a governed turn, but it still left the obvious question open: does the path hold when a real model sits behind the MCP server?

2.24.1 carries that proof:

  • examples/mcp-anthropic-agent/server.js exposes agentxchain_turn over MCP stdio
  • the server forwards prompts and governed context to the Anthropic Messages API
  • examples/live-governed-proof/run-mcp-real-model-proof.mjs runs the governed CLI through that server
  • the returned payload is extracted, validated, and accepted by the state machine

That proof passed on 2026-04-08:

  • run: run_83a64c7bb6925711
  • turn: turn_c8703d87f325e108
  • model: claude-haiku-4-5-20251001
  • usage: 1,630 input tokens + 649 output tokens

This is the right scope for a reference MCP server. It is thin, real, and governed. It does not pretend to be an all-purpose model gateway.

Scenario D escalation and recovery are now exercised through the real CLI​

AgentXchain already had escalation logic. The missing proof was operational: what happens when a governed turn actually fails twice, blocks the run, and has to be recovered without losing the decision trail?

2.24.1 adds a dedicated recovery proof:

  • dev turn rejected twice until max_turn_retries is exhausted
  • run transitions to blocked with escalation:retries-exhausted:dev
  • failed turn remains retained in active state
  • operator recovery clears blocked state and records escalation_resolved
  • corrected dev work is accepted
  • eng_director is assigned and accepted with an objection, preserving the review contract

Proof run:

  • run: run_ebf10c05d7419a0c
  • dev turn: turn_62dc3552e94dc5a9
  • director turn: turn_c88f47a0ca64e8cc

This matters because recovery is part of the product, not a support footnote. If the protocol cannot recover from dead-end turns cleanly, it cannot support long-horizon delivery.

Release truth now includes the public npx path​

Package metadata alone is not enough. Users do not experience your release as npm view. They experience it as npx agentxchain.

2.24.1 hardens release postflight so it now proves:

  • the registry serves the exact target version
  • the tarball metadata and checksum exist
  • npx --yes -p agentxchain@<version> -c "agentxchain --version" resolves and executes in an isolated temp environment
  • the published tarball installs correctly into an isolated prefix
  • exported runner and adapter interfaces import cleanly from a consumer project

That closes the last meaningful gap between "the package published" and "a new user can actually run it." The shorter npx agentxchain@<version> --version form is not the proof contract because npm can consume --version before the package binary sees it.

Why this release matters​

AgentXchain claims governed multi-agent delivery. Those claims only hold if the hardest boundaries are proven where they usually fail:

  • across runtime adapters
  • across recovery paths
  • across release distribution surfaces

2.24.1 matters because it converts those boundaries from inferred behavior into executed evidence, and because the release surface itself is once again held to the same fail-closed standard as the product.

Where to go next​

  • See Adapters for runtime contracts across manual, local CLI, API-backed, and MCP paths.
  • See Quickstart for governed repo setup and the first-turn workflow.
  • See Protocol Implementor Guide if you are building a compatible runner or verifier.

Evidence​

  • 2486 node tests / 534 suites / 0 failures.
  • 774 Vitest tests / 36 files / 0 failures.
  • Docusaurus production build passes.
  • Live MCP real-model proof: run_83a64c7bb6925711.
  • Live escalation and recovery proof: run_ebf10c05d7419a0c.