Skip to main content

AgentXchain v2.29.0

2.29.0 is the remote-agent proof and automation-correctness release.

2.28.0 closed security and plugin credibility debt. This release closes a different truth gap: whether AgentXchain can front a real remote model through a governed adapter surface without silently repairing the model's output, and whether automation can trust the CLI to fail closed on retained validation errors.

remote_agent is now a real shipped connector surface​

The new remote_agent adapter lets a governed role dispatch over synchronous HTTP POST to an external service.

What shipped:

  • Runtime config validation for url, headers, and timeouts
  • Secret-header redaction for authorization, x-api-key, cookie, and proxy-authorization
  • Public CLI integration through agentxchain step
  • Runnable example under examples/remote-agent-bridge/

That example is not hand-wavy sample code. It includes:

  • server.js for deterministic proof
  • run-proof.mjs for end-to-end CLI lifecycle proof
  • model-backed-server.js for a real Anthropic-backed bridge
  • run-repeated-proof.mjs for repeated reliability measurement

Real-model proof through the governed adapter boundary​

AgentXchain now proves a real model can satisfy the governed turn-result contract through the remote_agent bridge:

  • Dev turn: model returns non-empty proposed_changes[], proposal materializes, and proposal apply succeeds
  • QA turn: model returns a review artifact with at least one objection
  • Both turns pass the same acceptance pipeline used by the CLI in normal governed execution

The fence-strip concession is now stated honestly​

The wrong invariant is "the model must never emit markdown fences."

The correct invariant is:

  • no field-level repair
  • no semantic fixups
  • only logged removal of one outer markdown-fence pair when the enclosed JSON is otherwise valid

Fence-free raw JSON is still the preferred transport shape and the prompt still requests it. But the release-critical proof boundary is "no field-level repair," not "perfect fence-free transport."

Repeated reliability proof​

Single-shot success is weak evidence. The bridge now ships a repeatable proof harness:

  • 5 independent governed lifecycles
  • 5/5 PASS (100%)
  • No retries
  • 10/10 logged outer-fence strips
  • No field-level repair

This does not prove universal reliability across prompts, providers, or higher sample sizes. It does prove the contract is reliably teachable to Claude Haiku from one system prompt under the shipped proof conditions.

Automation correctness: validation failures exit non-zero​

agentxchain step previously had a real automation bug: a staged turn result could fail validation, be retained for correction, and still leave the command exiting 0.

That is not acceptable for CI or shell automation. step now exits 1 when validation fails and the staged result is retained. Scripts can fail closed without guessing from console text.

Evidence​

  • 2752 node tests / 582 suites / 0 failures
  • node --test cli/test/e2e-remote-agent-proposed-authoring.test.js
  • node --test cli/test/remote-agent-model-backed-proof.test.js
  • node --test cli/test/remote-agent-repeated-proof.test.js
  • cd cli && npm test
  • cd website-v2 && npm run build