Skip to content
/
Phase D Batch 28

Phase D Batch 28 — Notes

Phase D Batch 28 — Notes

Batch theme: Enterprise gateways, registries & governance runtimes — MCP registries, proxies, gateways, policy/governance runtimes

Agent: Phase D batch agent #28 of 33 Date: 2026-05-26 Frameworks: 10 Files written: 110 (11 per framework × 10)

Frameworks Covered

Slug Repo Stars Tier Primary Category
contextforge IBM/mcp-context-forge 3770 A MCP-gateway-with-admin-UI
archestra archestra-ai/archestra 3745 A K8s-native-MCP-orchestration
agentgateway agentgateway/agentgateway 2871 A universal-AI-gateway
atmosphere Atmosphere/atmosphere 3767 A Java-runtime-with-compliance
parlant emcie-co/parlant 18084 A conversation-behavior-engine
cai-cybersecurity aliasrobotics/cai 8753 A cybersecurity-pentest-framework
plano katanemo/plano 6543 A envoy-based-AI-gateway
gitagent-protocol open-gitagent/gitagent-protocol 2790 B agent-governance-standard
acpx openclaw/acpx 2759 B agent-to-agent-protocol-client
claude-plugin-design-joestump joestump/claude-plugin-sdd 11 C architecture-governance-plugin

Tier Counts

  • Tier A (production-ready, significant stars, runtime): 7 frameworks
  • Tier B (active, meaningful stars, notable gaps): 2 frameworks (gitagent-protocol, acpx)
  • Tier C (low stars but technically capable): 1 framework (claude-plugin-sdd)

Duplicates / Overlaps with Other Batches

None detected. All 10 frameworks are distinct from the 11 seed frameworks and from the other Phase D batches (batch numbers known: 1, 2, 8, 9, 11, 13, 17, 18, 22, 23, 24, 25, 26, 29, 30, 32).

Most Surprising Find

Atmosphere's Stop hook + drift-log enforcement (atmosphere/atmosphere): The framework ships a git hook script check-drift-log.sh that fires when Claude Code attempts to stop a session. The script reads the Claude Code session transcript, scans for 6 specific "drift-correction language" patterns (phrases that indicate the AI went off-spec), and blocks the stop if it detects these patterns without a corresponding entry in .harness/drift-log.md. Additionally, git commit hooks silently reject commits with AI co-author signatures in the message.

This is the only framework in the corpus that:

  1. Uses a Claude Code Stop hook as a compliance enforcement gate
  2. Reads the AI's own session transcript to detect behavioral drift
  3. Blocks the AI from stopping until it documents its drift in a human-readable log

The mechanism treats the AI as a contractor that must log deviations before clocking out. No other framework in any batch takes this approach.

Second-Most Surprising Find

acpx's prompt queueing via IPC: acpx implements a Unix socket/named pipe queue so prompts submitted to a busy agent session are queued and executed in order rather than rejected. This IPC-based cooperative scheduling has no equivalent in any other framework. Combined with --no-wait (fire-and-forget), it enables orchestrators to submit work without polling for agent availability.

Notable Patterns Across the Batch

audit_log implementations (batch focus)

Framework Audit Log Mechanism Storage
contextforge OTEL spans + SQLAlchemy events PostgreSQL/SQLite
archestra PostToolUse hook logs in-memory/file
agentgateway CEL-evaluated policy decisions structured JSON
atmosphere governance_audit_log table (JSONB) PostgreSQL
parlant Per-turn behavioral trace SQLite
cai-cybersecurity .cai/ JSONL session traces flat files
plano Agentic Signals™ (OTEL zero-code) OTEL backend
gitagent-protocol git commit history git repository
acpx ~/.acpx/sessions/*/history/ JSON flat files
claude-plugin-sdd CLAUDE.md + docs/ (no formal audit log) markdown files

Deployment topology patterns

  • K8s-native: archestra (Helm), agentgateway (Gateway API + K8s controller), plano (Envoy-based)
  • Self-hosted service: contextforge (uvicorn/Docker), atmosphere (JVM/brew), parlant (Docker/pip), cai-cybersecurity (pip/Docker)
  • Single binary: agentgateway (Rust), plano (brightstaff native binary)
  • npm package (definition/client): gitagent-protocol (opengap CLI), acpx (acpx CLI)
  • Claude Code plugin (no server): claude-plugin-sdd

Observability/safety primitives

  • OTEL: contextforge (full), plano (zero-code Agentic Signals), agentgateway (export targets)
  • CEL policy engine: agentgateway (cel-fork/celx crates)
  • AWS Cedar / OPA Rego: atmosphere (governance modules)
  • Multi-guardrail stack: agentgateway (regex + OpenAI moderation + Bedrock Guardrails + Model Armor)
  • PII plugins: contextforge (40+ plugins including PII redaction)
  • ARQ reasoning blueprints: parlant (arXiv 2503.03669) — test-validated behavioral correctness
  • Dual-LLM sanitizer: archestra (sub-agent sanitizes inputs before primary LLM)
  • Prompt injection notes: agentgateway (A2A indirect prompt injection mitigation)

Key Differentiators per Framework

contextforge

IBM-backed Python gateway with the widest plugin ecosystem (40+) in this batch: TOON, circuit-breaker, PII, content-moderation, rate-limit, retry. Unique: HTMX+Alpine.js admin UI embedded in a Python service (no npm/React). Tool authorization log in SQLAlchemy.

archestra

Lethal Trifecta awareness (sensitive data + prompt injection + exfiltration = Simon Willison's concept) built into architecture. Dual-LLM sanitizer sub-agent. Cost optimizer claiming 96% reduction. PostToolUse hook for Biome formatting (unique in corpus). AGPL-3.0 with CNCF/LF membership.

agentgateway

Only framework supporting all three: LLM Gateway + MCP Gateway + A2A Gateway. Rust single binary with embedded Next.js UI at :15000. CEL policy engine with cel-fork/celx crates. Most complete "universal AI gateway" in the batch.

atmosphere

Java/JVM 2008-vintage; v4.0.48 most battle-tested. Stop hook behavioral drift detection (most surprising find). 12 runtime adapters with OpenSSF SKILLCARD.yaml signed manifests. Cedar/Rego policy, HITL durable approvals, PostgreSQL audit log.

parlant

Highest stars (18084) in batch by large margin. Per-turn contextual matching engine — only items semantically matching the current turn are assembled into the LLM call. ARQ reasoning blueprints (published research). Strict/Canned output mode. Production at JPMorgan Chase.

cai-cybersecurity

Only offensive security framework in corpus. 15 code-class security agents. Professional Edition ($350/month, alias1 model, "no refusals"). retester → HackerOne production pipeline. NOASSERTION license. 4 CLIs including cai-gif (offensive tool presentations as GIFs).

plano

Built by Envoy core contributors. WASM filters embedded in data plane. Plano-Orchestrator 4B LLM for routing. Agentic Signals™ zero-code OTEL is a commercial differentiator. 8 developer skills in .claude/skills/.

gitagent-protocol

git IS the audit trail — not a workaround, a design principle. SOD conflict matrix in agent.yaml with AJV validation + CI blocking. 12 framework export adapters (only framework-neutral definition standard in batch). Financial compliance (FINRA/Fed/SEC) as first-class.

acpx

ACP (Agent Client Protocol) client — third protocol category alongside MCP (tools) and git (state). 17 built-in agent adapters across 8 vendors. IPC prompt queueing. Designed for agent-as-primary-user. Dynamic sub-agent role assignment via --system-prompt.

claude-plugin-sdd

Only framework implementing a complete development methodology (SDD) as a plugin. Adversarial Grumpy Engineer in scrum planning. Security-by-default specs with opt-out requiring justification. Conflict-marker hard gate before PR review. Artifact dependency DAG with impact analysis. 18 slash commands.

Files

All 110 files written to /Users/yigitkonur/research/spec-driven-dev/spec-system/<slug>/ (00-summary.md through 09-uniqueness.md + METRICS.yaml).

Paths

/Users/yigitkonur/research/spec-driven-dev/spec-system/contextforge/
/Users/yigitkonur/research/spec-driven-dev/spec-system/archestra/
/Users/yigitkonur/research/spec-driven-dev/spec-system/agentgateway/
/Users/yigitkonur/research/spec-driven-dev/spec-system/atmosphere/
/Users/yigitkonur/research/spec-driven-dev/spec-system/parlant/
/Users/yigitkonur/research/spec-driven-dev/spec-system/cai-cybersecurity/
/Users/yigitkonur/research/spec-driven-dev/spec-system/plano/
/Users/yigitkonur/research/spec-driven-dev/spec-system/gitagent-protocol/
/Users/yigitkonur/research/spec-driven-dev/spec-system/acpx/
/Users/yigitkonur/research/spec-driven-dev/spec-system/claude-plugin-design-joestump/