HorusJiang/dsh-jev-tools

dsh-jev-tools is a native plugin for DeepSeek Harness that wires TypeSafe's System One decision model Jev into long-running agent sessions. The package declares a dsh.bundle field in its package.json, which the DSH loader reads; installation is performed with dsh plugin --profile web add dsh-jev-tools, after which the package name is appended to dsh.profile.bundles for that profile. Plain npm install dsh-jev-tools only drops the tarball into node_modules and never runs a single byte because npm ignores that custom field, so the documented DSH install path is the only one that mounts the plugin. The product exists exclusively for DSH and does not target Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode, Cline, or any other agent runtime.

The plugin surfaces five capabilities: per-segment pruning of oversized read, grep, glob, web_fetch, and web_search outputs once they exceed 2000 tokens; prompt-injection screening of fetched pages with reminders rather than blocks; skill picking on the first prompt assembly of a turn when the catalog has at least 15 entries, scoring against the latest user message (falling back to the most recent three when it is too short); the jev_ask tool for typed multi-choice questions; and jev_gate, a delivery gate the model calls before declaring completion, which is the only capability that inverts the project's otherwise pervasive fail-open posture so that ambiguous paths collapse to escalate instead of silent pass-through. A prune.shadow dry-run mode runs the judging and bookkeeping but mutates nothing, and slash command /jev-status reports enablement, key source, endpoint, counts, ledger path, and skip reasons such as task-too-vague, too-small, budget-turn, no-saving, no-skills, catalog-too-small, and unauthorized. It targets DSH power users who care about context hygiene and output bloat in long-horizon workflows and who already hold a TYPESAFE_API_KEY from <https://console.typesafe.ai/keys>.

Dependencies are minimal, tests run via node --test with no framework, and the license is MIT; the project requires Node 24 or newer and publishes through npm trusted publishing in two deliberate steps so a compromised workflow cannot ship unilaterally. Without an API key the plugin is fully inert, makes zero network requests, and the five capabilities can each be toggled independently in the settings page with immediate effect. The default endpoint is https://api.typesafe.ai with the path /v1/systemone appended by the plugin, so self-hosted Jev-compatible hosts or gateway-fronted deployments must override baseUrl. Limits include text-only input, a 64k token ceiling per request, no image or audio handling, no reasoning or arithmetic, and probabilities that are good for ranking but uncalibrated, so they must not be used as accuracy thresholds. Edits under lib/ require restarting dsh web to re-import ESM modules when the profile points at the development directory, and the bundled ledger persists per-judgment metadata to $DSH_HOME/storages/dsh_jev_tools/ keeping the most recent 1000 entries in both memory and disk with a separate counter line for cumulative input tokens.