dsh-behavior-enhancer

In complex agent workflows, once the model fans out parallel tool calls and one fails, nothing forces the system to converge. The parallel pool keeps expanding, errors cascade, and corrupted files have no rollback path. dsh-behavior-enhancer addresses the often-overlooked layer beneath output quality: it does not change what the model says, but constrains how it calls tools, so that failures naturally serialize and write operations become safely reversible.

Three independent mechanisms work together. behaviorPrompt injects a static soft-constraint section into the system prompt (read before write, validate before batch, fail-and-converge, serialize same-type edits) to give the model a behavioral baseline. parallelConvergence watches every tool result; any failure immediately drops the parallel pool ceiling to 1, restores the user-set value after three consecutive successes, and proactively downgrades when context pressure crosses a threshold, preventing failure from spreading. postWriteCheck snapshots each target file to a fixed directory (five snapshots per file) before write or edit executes, then runs a lightweight heuristic check after the write and automatically rolls the file back if the check fails. The three paths are fully independent—if one degrades, DSH core flow is unaffected.

Ideal for developers running multi-step, multi-file agent tasks in DSH. The plugin has zero npm dependencies, makes no network requests, spawns no subprocesses, and every module is fail-open. It complements dsh-token-optimizer (which handles content compression and trimming) by handling call discipline and failure convergence. Both can be installed independently or combined to form a complete stability layer.