dsh-failbook

One of the most expensive and frustrating failure modes in agent workflows is the same mistake retried over and over with slightly different arguments: a missing binary returns 127 again, a sandbox policy denial is hammered five times, a broken npm dependency is reinstalled N times. The built-in repeat-tool-reminder in DeepSeek Harness only catches identical-argument consecutive repeats, so as soon as the model tweaks a parameter and hits the same root cause, the guard goes blind. dsh-failbook closes that gap by clustering failures by their signature—command-not-found, sandbox EACCES, npm ERR!—regardless of how the arguments differ. Failures are persisted across sessions via ctx.storageDomain, and once a bucket crosses a configurable threshold within a rolling window, a short advisory (root cause plus a suggested fix) is injected into the next agent request, breaking the "same mistake, new parameters" loop.

The plugin ships with a Settings → Failbook panel that surfaces top failure signatures with counts, recent timestamps, and one-click mute or clear actions. Detection is deliberately conservative: only unambiguous signals (structured errors, non-zero exits ≥ 2, sandbox markers, well-known error text) are recorded, cancellations are never counted, and false-positive buckets can be muted or excluded per tool. Stored entries are truncated previews only; full command output never leaves the host, and the management API is loopback-bound. If you run long multi-step agent pipelines in DeepSeek Harness and are tired of watching the same error burn tokens in a retry loop, dsh-failbook plugs in with essentially zero configuration and no pipeline changes.