dsh-preset-anchored-standard
Run the following command in DeepSeek Harness:
dsh plugin install ruby1304/dsh-preset-anchored-standard
Paste the following prompt into your AI chat to install this plugin:
Run dsh plugin install ruby1304/dsh-preset-anchored-standard inside your DeepSeek Harness terminal to install this preset (source repository: https://github.com/ruby1304/dsh-preset-anchored-standard)
About this plugin
Trajectory-evaluation runs in DeepSeek Harness often need the very first request pinned to a minimal tool pair (persistent bash plus str_replace_editor) with auto-injected workspace and skill context stripped, so the model baseline under a constrained catalogue is measurable. Yet from request two onward, the full Standard catalogue job_output, run_in_background, and friends is exactly what you need. dsh-preset-anchored-standard bridges that gap: it anchors request one on the Minimal pair, watches for the first durable promotion signal (a tool call or the first assistant message), and exposes the complete Standard catalogue from request two on.
Beyond the structural switch, the preset addresses a subtle behavioural trap. A model that opens with only a shell tends to keep the foreground-sleep-and-pgrep-polling habit taught by the bootstrap bash description, even after the catalogue silently expands. We saw a delegated session poll a benchmark with sleep 240 loops for roughly ninety minutes while job_output, job_list, and job_kill were already available. So the preset injects a one-shot user-role notice at the promotion step, telling the model the full catalogue is live and that waits longer than about two minutes belong in background jobs with completion notices rather than foreground polling. The notice fires only for sessions the process actually saw bootstrapping; a session resumed after promotion is not interrupted.
Ideal for anyone running dsh trajectory evaluations, benchmark reproductions, or agent behavioural comparisons that need a minimal anchor on request one without giving up Standard capabilities for the rest of the session. All knobs (bootstrapTools, promoteOn, suppressedContextSources, bootstrapMaxTokens, promotionNotice) live in the tool-bootstrap row of agent.cordis.yml. The preset composes only shipped dsh plugins plus its own local script, targets dsh 0.1.0-rc.5, and has no third-party runtime dependencies.
Use Cases
- Trajectory evaluation requiring a minimal tool baseline on request one and full Standard capabilities thereafter
- Comparing foreground polling versus background job behavior under a constrained tool surface
- Benchmark reproduction where the first request must be pinned to a specific tool pair
Best For
- Researchers running agent behavioral ablation studies
- Evaluation engineers who need fine-grained control over dsh session tool catalogs
- Developers designing or reproducing benchmark trajectories
Related Plugins
A method pack that makes AI coding agents plan against your real baseline, prove completion with fresh evidence, and reduce reworks and unsafe changes.
Turns the DeepSeek Harness session into a captain that builds a durable sub-agent team, splits goals into dependency-aware tasks, and coordinates work via direct messages and a live Web UI.
Gives coding agents design judgment, letting Claude Code, Cursor, and 70+ agents generate and iterate high-quality UI, presentations, and graphics right from the terminal.
Run the Pi ecosystem's plugins on DeepSeek Harness, unmodified, via a compatibility layer that implements Pi's public extension ABI on DSH's native services.