dsh-tool-args-unwrap

When running small or mid-sized models through long sessions with heavy tool-calling, models occasionally emit shape-malformed arguments — most typically an extra {"arguments":} wrapper around the real payload. The host persists that wrapper verbatim into tool/call and assistant/message, then replays the entire history back next turn. The model sees its own bad example and imitates it, deepening recursively (observed 1→38 layers in the wild). Restarting the session "fixes" it not by resetting model state but by removing the bad exemplars.

dsh-tool-args-unwrap intercepts and corrects arguments before they are persisted, repairing execution, persistence, and history replay in one pass. Since v0.3.0 it also sanitizes stale malformed records replayed in old sessions (clone-on-write in memory; persisted data is never touched), so the session heals on the very next turn. It covers four defect families observed across 16,535 real tool/call records and judges every rewrite against the host's own validator — the acceptance rule is strictly fewer violations, never merely "required keys present." When in doubt, the block passes through unchanged.

Ideal for DSH developers who run small or mid-sized models through long sessions or high-frequency tool-calling loops. If you have hit the frustrating "the agent says it can't break free and only a restart helps" cycle, this plugin should save you a lot of manual debugging time.