fault-injection

Adversarial agent loops for verifiable vibe researching. Claude Code plugin. Falsification-first, not production-first.

Local dev proxy. Named services, HTTPS, and failure injection — no ports, no mocks.

Distributed Mutual Exclusion Explorer (Variant 3): interactive Token Ring & Ricart–Agrawala visualiser with fault injection, scripted demos, and evidence export.

Deterministic 8-node chaos and soak testing for causal ordering, deduplication, replay, and recovery behavior.

Reproducible teardown of the openai-agents-js financial-research example: the manager emits a research report when every web search failed and when verification never passed. 6-case synthetic-orchestration harness, baseline 4/6, one-file patch 6/6, plus a field guide for probing your own agent.

Reproducible teardown of a scheduled maintenance agent in google/adk-python: every issue audit fails, the batch reports them all as successfully processed, and the run exits 0. Harness, one-file fail-closed patch, hashed RED to GREEN evidence. Reported upstream first.

Fault-injection experiments for tool-using agents under partial failure