dsh-solo-agent

In multi-step coding tasks the main agent often plays both orchestrator and executor, running shells, reading and writing files, and searching code. The context window fills up fast, the model's attention bounces between orchestration and execution, and dispatch efficiency keeps dropping. dsh-solo-agent tackles this by injecting a Solo preset into DSH agent-presets: the main agent is strictly constrained to pure orchestration. Its visible tool list is reduced to ask_user_question, todo_write, skill, and four foreground subagent tools. All shell, filesystem, and web capabilities are mounted at the preset layer for subagents and hidden from the main agent via solo-bootstrap.mjs.

The dispatch loop revolves around four roles. subagent_explore performs read-only research and produces target tests, the test command, and a root cause. Up to five iterations follow: subagent_editor makes focused edits without self-testing, subagent_general runs tests and returns raw output, and Solo makes the next decision solely based on that raw output. For large or high-risk changes that pass tests, subagent_verify performs adversarial validation, returning PASS, FAIL, or PARTIAL. The role registry lives in lib/roles.js; adding a new role is just one extra record. The preset also enables full compaction (dsh-compaction-basic, dsh-compaction-tool-result-pruner, dsh-command-compact) and applies minimal anchoring to subagents with delegationDepth greater than zero, while the Solo main agent stays unanchored to maintain a global view.

This plugin suits developers who frequently run multi-step coding, debugging, or refactoring tasks and are sensitive to context-window growth. If your DSH agent workflows suffer from context bloat, cross-task context bleed, or uncontrolled dispatch rounds, dsh-solo-agent offers a structured fix: each role does only what it should, and the main agent only makes decisions.