dsh-loglens

When debugging inside DeepSeek Harness, logs are often the hardest part to digest: pasting everything into context blows the window, while ad-hoc truncation loses the timeline and repetition patterns. dsh-loglens provides a bounded, cursor-aware log inspection layer that turns raw log streams into structured, incrementally readable signals a model can focus on turn by turn.

The plugin exposes four model-facing tools. logs_read parses a bounded window from the current cursor, normalizes severity, collapses repeated lines, and fingerprints exceptions to produce a compact snapshot. logs_raw retrieves original cached events only after an explicit event-id or fingerprint lookup. logs_list reports configured sources and cursor state, and logs_reset_cursor restarts incremental reading for a single source. File sources use byte offsets with inode checks; shell-command sources are administrator-authored and memory-capped. The model only ever sees a source id, never an arbitrary path or command. In the Harness Web conversation UI, the snapshot renders as stats, grouped issues, recent events, and expandable stack frames, while raw content stays behind logs_raw for on-demand retrieval.

This plugin fits developers who do daily debugging, alert triage, or code review within DeepSeek Harness. If you want the model to drill into logs incrementally over a bounded window rather than swallow megabytes of text in one shot, dsh-loglens makes log inspection repeatable, auditable, and firmly bounded.