Skip to content
/
Phase B Batch 31

Batch 31 — Hooks/Governance Overflow (lint, audit, steering, verification)

Batch 31 — Hooks/Governance Overflow (lint, audit, steering, verification)

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
vibelint 2 npm-package vibelint none none no A
ctxlint 6 npm-package ctxlint none none no A
theauditor 550 unknown unknown unknown unknown unknown C
clauder 58 bash-script-bundle clauder (wrapper) web-dashboard sequential no A
zenable 0 unknown unknown unknown unknown unknown C
haft 1333 cli-tool haft terminal-tui sequential no A
sponsio 445 standalone-repo sponsio web-dashboard none no A
dashclaw 268 standalone-repo dashclaw web-dashboard none no A
pi-steering-hooks 5 npm-package none none none no A
spartan-ai-toolkit 72 npm-package installer none sequential no A

Intra-Batch Patterns

All 8 full-analyzed frameworks share the conviction that prompt-based behavioral rules are insufficient and must be supplemented by code-level enforcement. The batch cleaves into two camps: pre-session linters (vibelint, ctxlint — analyze before or after sessions, no live hooks) vs runtime guardrails (clauder, sponsio, dashclaw, pi-steering-hooks — intercept tool calls mid-session). Haft and Spartan are outliers: Haft governs the reasoning discipline before tool calls happen (decision contracts, not regex rules), while Spartan enforces engineering workflows through mandatory phase gates rather than tool-call interception. The intervention axis runs from post-session audit (vibelint, leftmost) through pre-session static lint (ctxlint) through pre-tool blocking (pi-steering-hooks, clauder, sponsio, dashclaw) to pre-reasoning governance (haft). None of the 8 full-analyzed frameworks use multi-model routing.

Most Interesting Finds

Sponsio — ships a Fuzzy LTL (Linear Temporal Logic) monitor for temporal ordering contracts (e.g., "compliance_approve must precede wire_transfer"), benchmark-validated on public datasets (ODCV-Bench, RedCode-Exec), at p50 0.139ms per check. This is formal methods applied to agent safety at production speed — qualitatively different from regex-based guardrails. The universal bundle being intentionally empty (with a code comment explaining the design evolution) is a rare example of documented architectural change-of-mind in production code.

DashClaw — introduces "durable finality": tracking terminal outcomes so a retried agent never silently double-executes. This addresses a failure mode (network interruption mid-action followed by retry) that none of the 11 seed frameworks acknowledge.

Items Written as Tier C

  • theauditor — README says "Product launch imminent," repo has only README + LICENSE + THIRD_PARTY_LICENSES.txt. 550 stars appear to be pre-launch waitlist. No source code, hooks, or architecture available.
  • zenable — assigned URL is https://zenable.io/ (website, not GitHub repo). No public source code found via gh api search.

Cross-References Discovered

  • DashClaw references OpenClaw explicitly in README and topics; provides an @dashclaw/openclaw-plugin — DashClaw extends/governs OpenClaw agents.
  • pi-steering-hooks targets @earendil-works/pi-mono (pi coding agent) — a niche agent not covered by any seed or other batch members.
  • Sponsio references Simon Willison's "Lethal Trifecta" security research as an explicit design input — rare cross-domain citation from web security research into agent safety frameworks.
  • clauder is attributed to blueraai/clauder in the batch list but the actual canonical repo is spacehendrix/clauder. The blueraai org does not have this repo.
  • haft was formerly quint-code — the install URL still points at the historical path; the binary is haft.
  • Spartan AI Toolkit supports the Claude Managed Agents TeamCreate/TeamDelete API — the only batch member explicitly referencing Anthropic's managed agents coordination primitives.
← Batch 30
Batch 32 →