The Terminal

LAR: A Production-Grade Claim We Could Not Import

lar-agent-resilience advertised Agent, ModelRouter, and Checkpoint. The package __init__ was empty. The code imported lar.* and there was no lar package. Six tests passed — on a skill demo. We made the README true.

Paula Rossi

LAR: A Production-Grade Claim We Could Not Import

The GitHub description said: “A production-grade, open-source reference implementation for resilient agent design on Linux.”

The auditor’s first import said: ModuleNotFoundError.

PR: smfworks/lar-agent-resilience#2

Original state

pytest -v --tb=short
# 6 passed in 0.17s

Those six tests exercise skills/resilience-skill/scripts/failover.py only. src/ was untested.

README promised:

from agent_resilience import Agent, ModelRouter, Checkpoint

src/agent_resilience/__init__.py contained a comment. Runtime files did from lar.config import .... There is no lar package. tools.py shadowed tools/, so builtin tools were unimportable. pyproject.toml named the project local-agent-runtime, had no build-system, no scripts. ExecTool used shell=True with python / git / curl on the default allowlist.

P0 count from the audit: 8. The production-grade claim did not hold.

Decisions

  • One import path: agent_resilience. Rewrite every from lar..
  • Prefer deletion over finishing unused surface. Observatory, TUI, benchmark, memory, dashboard HTML went away rather than shipping more unwired code.
  • Promote the skill router into the library. Two ModelRouters is how READMEs rot.
  • Public Agent that failsover against a fake LLM. If you cannot test model death without Ollama, you cannot claim resilience.
  • Exec: shell=False, argv allowlist, drop python/git/curl defaults.
  • Fetch: https only; deny file:// and private/loopback.
  • No OpenClaw / clawhub expansion. Lab policy. The existing skill becomes a thin wrapper.

What landed

Package agent-resilience 0.2.0. Real exports: Agent, ModelRouter, Checkpoint. lar console script. Build-system and packages.find. Linux CI on 3.11 and 3.12.

Identity no longer TypeErrors on camelCase / string timestamps. Checkpoint save/load is tested. Tool safety tests reject pipes, file://, and loopback fetch.

Implementer verification on this Windows host: 30 passed in 0.30s, ruff clean.

Independent GitHub Actions: CI run 31684553251 — test (3.11) and test (3.12) completed/success. GitGuardian also passed. This is the only one of the four hardening PRs whose Actions run I could confirm via the API at publish time.

Testing approach

New tests are failure-mode tests, not happy-path tourism:

  • packaging import
  • identity HMAC / timestamps
  • checkpoint resume
  • agent failover with an injectable fake LLM
  • exec/fetch jail

The original six skill tests now import the library module.

Lessons

  1. pytest green is not a product. Six tests on a demo script plus a README written for a future API is a design note.
  2. Import the documented names in CI. from package import AdvertisedSymbol is the cheapest contract test in software.
  3. tools.py next to tools/ is a footgun. Python will pick the module. Your package data will not.
  4. Delete unused dashboards. Unwired Observatory code is not optionality. It is a second lie.
  5. shell=True + a generous allowlist is RCE the first time an agent is pointed at untrusted text.

Remaining limitations

  • Consolidation “recovery” is completion, not agency gain. Do not market it as such.
  • Live Ollama path is untested here; failover is proven with a fake backend.
  • The cron-misfire CircuitBreaker is still not the model-failover path. Naming debt remains.
  • Linux-first; Windows is not a required CI target.

We did not make LAR legendary. We made it importable, tested at the failure modes it claims, and honest. That is what production-grade actually means.