dsh-tool-normalizer

In DeepSeek Harness autonomous agent loops, every failed tool call triggers a full model round-trip: the error returns, the model re-reads the entire conversation context, and tries again. In a large workspace that retry resubmits roughly 180k input tokens, so even a modest 5% failure rate quietly doubles your token spend on affected turns. dsh-tool-normalizer sits on the tools/execute pipeline and repairs the failure before it ever reaches the model, so your agent stops tripping over its own tool calls.

What it actually does is concrete: it auto-corrects run_code schema mismatches (command to code, missing description fields, accidental Markdown fences), repairs syntax breakage from Python-style triple-quoted strings and unescaped backticks, bridges direct tool calls the model hallucinates in Code-Mode back through the host dispatcher, resolves relative paths to absolute ones, and performs a scoped read-then-retry when the file mutation policy blocks an edit. For errors that genuinely cannot be auto-fixed, it appends one actionable hint to the original text so the model can correct itself in the same round-trip. Every interception, heal, and pass-through is logged to structured JSONL events and surfaced in a live diagnostics dashboard embedded in the Settings panel—KPI cards, per-tool and per-category rankings, and a filterable before/after trace table.

It is built for developers and teams running DeepSeek Harness workspaces where both agent reliability and token cost matter: no changes to your calling conventions are needed, the plugin works silently in the background, and the dashboard lets you verify exactly how many failures were healed and how many tokens were saved.