Skip to main content

AgentXchain v2.66.0

v2.66.0 gives governed agents cumulative decision memory and closes a compression model gap that left workflow_artifacts in an incoherent state.

Publication note: the v2.66.0 GitHub tag exists, but the npm/Homebrew publish path failed during tagged preflight. Use the latest published npm/Homebrew release instead of [email protected].

What Changed​

  • Cumulative decision ledger dispatch: agents now see all agent-authored decisions from decision-ledger.jsonl as a "Decision History" markdown table in CONTEXT.md — not just the last turn's decisions. This prevents decision relitigation, consistency drift, and wasted turns rediscovering settled constraints.
  • Decision history filtering: only entries with an id field (agent-authored) are included. System entries (gate failures, escalations, policy actions) are excluded. Maximum 50 entries with truncation note.
  • workflow_artifacts compression coherence: workflow_artifacts is now explicitly in the 9-step bounded compression order, dropped after decision_history and before gate_required_files. Previously it had required: false but was absent from COMPRESSION_STEPS, effectively behaving as sticky despite being marked compressible.
  • Sticky context preservation: project_goal and inherited_run_context are preserved as sticky sections through full preflight compression — they cannot be silently dropped under token pressure.
  • Compression order (9 steps): budget → phase_gate_status → decision_history → workflow_artifacts → gate_required_files → last_turn_objections → last_turn_decisions → truncate last_turn_summary → drop last_turn_summary.

Why It Matters​

VISION.md lists "explicit decision history" as a core governance property. Before this release, agents only saw the last turn's decisions — an agent in turn 10 had no visibility into decisions from turns 1-8. This caused relitigation, contradictions, and wasted governance cycles.

The compression model fix is equally important: a section marked compressible but never targeted for compression creates silent model incoherence. Under extreme token pressure, workflow_artifacts survived all compression steps as if it were sticky. Now the compression model is internally consistent — every non-sticky section has an explicit place in the bounded drop order.

Evidence​

  • 56 tests / 14 suites / 0 failures (combined context-compressor, context-section-parser, token-budget, decision-history, and workflow-kit runtime context suite)
  • Decision history dispatch: 10 dedicated tests covering presence, truncation, system-entry filtering, missing/empty ledger, column format, chronological order, malformed JSON handling, pipe escaping, and section positioning
  • Compression model: 11 tests covering exact step order, sticky section preservation, truncation, and COMPRESSION_STEPS export validation
  • Token budget: 16 tests covering fit/compress/overflow paths with the updated section model
  • cd website-v2 && npm run build → clean production build

Install / Upgrade​

Use the latest published release from npm or Homebrew.