dsh-routing-suite: a runtime injector plus reasoning-mode router preset for the DeepSeek Harness

The problem dsh-routing-suite targets: one flat reasoning mode for every task

An agent harness that runs every request through the same reasoning configuration spends planning effort on trivial prompts and under-plans on hard ones. The repository's preset component addresses that by routing each task into one of several behaviour bands. The README describes three bands plus an internal weak route: spec for plan-collective work, react for execution, mixed as a trap to be avoided, and weak where the model classifies itself. The router-standard preset also picks a persona by model, with Pro getting a spec sentence plus few-shot examples and Flash getting neutral plus classify.

The intended audience is narrow. Everything here assumes the DeepSeek Harness, referred to throughout as DSH, and the package.json declares peer dependencies on @deepseek-ai/dsh-tools, cordis and schemastery. If you do not run DSH, the injector has nothing to inject into. The repository is written for people who already operate DSH profiles and want finer control over how a session reasons, not for teams evaluating agent frameworks in general.

How the injector, the router preset and the graded protocol fit together

Three directories sit at the repository root: injector/, preset/ and graded/. The README states that all three are ordinary directories committed directly, not submodules, and that upstream standalone repositories remain for independent release. The injector supplies dev_* tools covering injection, hot reload, unload, side-mount promotion and route self-healing. The preset supplies the reasoning-mode router. The graded component layers a session-level two-tier task protocol on top: brainstorm-to-multiple-choice alignment, a north-star statement, a specification plan, injection by specification, a check-in cadence, group close-out and final verification, with six tools, three modes, a red-team gate and an audit endpoint.

Data flow through the router is the part with the most concrete detail. The v0.3.0 changelog says the first turn now routes for real: the plugin captures the first genuine user message via agent/inbox/claimed before assembly, so the first request is classified by task instead of unconditionally falling into weak. Proximity guidance moved to agent/pre-step, which the changelog says injects the guidance in the same request as the user message and removes an extra API call that previously doubled cost per turn. The preset README reports cache hit rates of 92 to 94 percent for that guidance. Those figures come from the project's own experiment reports, not from independent measurement.

Installing dsh-routing-suite and running a first routed session

The README gives a one-line install through the DSH plugin command, targeting the web profile. The package name is the GitHub repository reference, not an npm package.

dsh plugin --profile web add github:yjh051108/dsh-routing-suite

If dsh is not on your PATH, the manual instructions use npx with the quoted package name '@deepseek-ai/dsh'. The repository also ships a three-step install chain: clone the repository, run .\install.ps1, which the README says performs injector assembly, preset copy and layout self-check, then restart DSH. The manual path is more explicit about where files land: the injector is added as a plugin, and each preset directory is copied flat into the user's .dsh directory.

git clone https://github.com/yjh051108/dsh-routing-suite.git
cd dsh-routing-suite
dsh plugin --profile web add .\injector

After the injector is assembled, the README copies each preset directory into the agent-presets folder. The commands below show the router-standard copy; router-spec follows the same shape.

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-standard'
Copy-Item -Recurse .\preset\router-standard $target

The README warns explicitly against copying the whole preset directory, because that adds a nesting level and DSH only scans one level of subdirectories under .agent-presets. After a restart, a new session should offer Router Standard and Router Spec (experimental) as choices. The repository also ships an install.sh for non-Windows shells, though the README's worked examples are all PowerShell.

What the P1-P23 experiment reports claim, and what they do not

The preset documentation reports results across 23 numbered experiments. Among the claims: persona selection by model yields a discrimination gain of 5.0 for Pro and 5.7 for Flash; per-turn proximity guidance reaches 96 percent routing accuracy and 100 percent convergence; and a single-task three-anchor persona takes open-ended task completion from 0 percent to 100 percent. The graded component publishes a separate study on attention decay in long tasks, with a comparison of a protocol chain against a no-protocol chain, and a measurement script at graded/scripts/measure.mjs that the README says anyone can run against their own session JSONL, with numbers only in the output for desensitisation.

Read those numbers as the project's own published measurements. The repository does not describe an external replication, and the comparison table in graded/docs/COMPARE.md covers three sessions of the same 3D task type, which is a small base. The measurement script is the most useful part of this bundle for a sceptical reader: it lets you recompute the reported metrics against your own sessions rather than trusting the published tables. That is a stronger position than most preset collections take, but it still means the headline percentages rest on the author's task selection.

Version pinning and the DSH developer preview constraint

The README states a DSH target range of >=0.1.0-rc.6 <0.2.0, and notes that the project has tracked rc.8, 0.1.1-rc.2 and 0.1.2-alpha.1. It also states plainly that DSH is in developer preview and that the vendor has said breaking changes will occur. That is the single largest operational risk in adopting this suite: the router hooks into agent/inbox/claimed and agent/pre-step, which are harness lifecycle points, and a harness change to either one can break routing without any change in this repository. The preset/CHANGELOG.md is where version tracking is recorded, so that file, not the top-level README, is the thing to read before upgrading DSH.

The package.json sets engines.node to >=22 and declares peer dependencies on @deepseek-ai/dsh-tools >=0.0.1-rc <2, cordis >=4.0.0-rc <5 and schemastery ^3.18.0. Note the licence split: the README states MIT for the repository, while package.json declares BSD-3-Clause for the distribution entry point. If you need a single unambiguous licence answer for a compliance review, that discrepancy has to be resolved with the maintainer rather than assumed.

Where this is the wrong tool, and what to use instead

If your sessions are short and single-purpose, the router has little to classify. The proximity guidance, the persona selection and the three-anchor mechanism all target long-horizon tasks where reasoning mode drifts; on a one-shot question they add prompt overhead for no behavioural gain. The graded protocol is heavier still, with its brainstorm, north-star and check-in stages, and it will feel like ceremony on small edits.

A real alternative in the same space is the anchored-standard preset from xiaobright, which the README credits for the anchoring mechanism that this suite builds on. The difference in approach is scope. An anchored preset fixes a reasoning posture for the session and stays there; dsh-routing-suite adds a classification layer in front of the posture, so the mode can differ per task, and it adds a runtime injector so the plugin set itself can change without a restart. If you want one stable posture and no routing logic, the anchored preset is the smaller dependency. If you want the mode to switch with the task and you are willing to track a developer-preview harness, this repository is the more capable option.

Maintenance, upgrade cost and licence notes

The last push to the default branch was on 2026-09-04, and the repository is not archived. The most recent releases are 0.0.1-rc1 and 0.0.1-exp for the graded component, both dated 2026-09-02, with v0.1.0 of the suite dated 2026-08-15. The graded version carries a release-candidate suffix, which is worth weighing before you put it in a shared profile.

Upgrade cost is dominated by the DSH target range. Because the plugin hooks harness lifecycle events, a DSH upgrade inside the tracked range still warrants a look at preset/CHANGELOG.md. The README documents the install path but does not document a rollback procedure for a partially completed install, so keep the previous .agent-presets directories before overwriting them. On licensing: the README says MIT and package.json says BSD-3-Clause for the distribution entry, and package.json also marks the package private. Treat that as a question for the maintainer, not something to resolve by reading the two files.

Editorial conclusion

Adopt it if you already run DSH 0.1.0-rc.6 or later and want per-task reasoning-mode routing plus hot-reloadable plugin injection; skip it if you are not on DSH, since the peer dependencies and the install command both assume the harness is present. Before installing, verify that your DSH version falls inside >=0.1.0-rc.6 <0.2.0, that Node is at least 22, and that the preset directories are copied flat into ~/.dsh/.agent-presets rather than as a nested folder.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. Releases
5. yjh051108/dsh-routing-suite on GitHub