Skip to content
/
Phase D Batch 19

Phase D Batch 19 — Observability, Audit, Migration & Ops Layers

Phase D Batch 19 — Observability, Audit, Migration & Ops Layers

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
openagentscontrol 4153 npm-package oac (thin wrap) none hierarchical no A
mycoder-ai 567 npm-package mycoder (own runtime) none parallel-fan-out no A
traceroot-ai 591 docker-image none web-dashboard (Next.js) background-daemon yes (BYOK) A
arbor-zed-gpui 14 standalone-repo none none none no C (misassigned slug)
asyncope-bringyour unknown unknown unknown unknown unknown unknown C (404)
agentops-boshu 369 claude-plugin ao (own runtime) none hierarchical+council yes (mixed Claude+Codex) A
agenttrace-luoyu 49 bash-script-bundle agenttrace (own runtime) terminal-tui (BubbleTea) none (observer) no A
liza 227 bash-script-bundle liza (own runtime, 35k LOC Go) terminal-tui (BubbleTea) hierarchical+adversarial yes (BYOM/5 CLIs) A
optio 967 docker-image none web-dashboard (Next.js:3100) hierarchical+event-driven yes (per-repo config) A
oh-my-codex-yeachan-omx duplicate C (duplicate)

Intra-Batch Patterns

The batch title "observability, audit, migration & ops layers" describes the theme accurately but with nuance: the batch contains frameworks across the full spectrum from active orchestrators (Liza, Optio) to passive observers (AgentTrace, TraceRoot). A strong pattern: 4 of the 7 valid frameworks (AgentTrace, AgentOps, Liza, Optio) ship dedicated Go or TypeScript CLI binaries with their own runtimes, not just plugin layers on top of Claude Code. Multi-provider ambition is pronounced: OAC, MyCoder, Liza, AgentOps, and Optio all explicitly support 3+ AI providers, suggesting the market is moving toward provider-agnostic orchestration. The "local-first" vs. "cloud-native" split is clean: AgentTrace + Liza + AgentOps are local-first; TraceRoot + Optio target cloud/Kubernetes deployment.

Most Interesting Finds

Liza — The only framework in the full surveyed set where agent role boundaries are enforced by compiled Go code (43+ validation rules, forbidden state machine transitions). The "55+ failure mode catalog with mechanical countermeasures" is published nowhere else. BYOM via provider CLI wrapping (no API keys needed) is architecturally distinctive. L4-level AI engineering maturity per an independent VP Engineering assessment.

AgentOps — The CDLC (Context Development Life Cycle) framing and "three ratchet rules" (no self-grade, fresh agent on failure, knowledge becomes constraints) are the most rigorous articulation of what prevents agent corpus rot. The multi-model council skill with adversarial 0-1000 cross-scoring in debate mode is the most sophisticated consensus primitive in the batch. Dogfooding stat (~3,867 cited decisions accumulated while building the tool itself) validates the approach empirically.

Items Written as Tier C

  1. arbor-zed-gpui — Slug collision. arbor-dev/arbor is a 2015 Go API gateway framework (last updated 2022, 14 stars) with zero connection to AI coding agents. Written 2-file stub only.
  2. asyncope-bringyour — Repository bringyour/migration-auditor returns HTTP 404. No public material. Written 2-file stub only.
  3. oh-my-codex-yeachan-omx — Flagged as duplicate of oh-my-codex-yeachan (Phase B Batch 23). Canonical analysis exists. Written 2-file stub with canonical: false.

Cross-References Discovered

  • OAC (openagentscontrol) explicitly compares against "Oh My OpenCode" (a yeachan variant) in its README comparison table — connects to the batch duplicate
  • AgentOps uses bd (beads) for issue tracking and gc (Gas City) for out-of-session scheduling — both are separate companion projects in the same ecosystem
  • AgentTrace ships a agenttrace-session-audit skill for Claude Code/Codex, making it simultaneously an observer and a Claude Code plugin
  • TraceRoot and AgentTrace are complementary: TraceRoot for cloud production monitoring, AgentTrace for local developer forensics — neither mentions the other but serve adjacent needs
  • Liza explicitly benchmarks against BMAD-METHOD in its README competitive matrix — the closest architectural comparison it names
  • Optio is conceptually positioned against Devin/Sweep (named in README) as the self-hosted alternative
  • OAC is built on OpenCode CLI (opencode.ai) — a dependency unique in the full catalog