Dsh Jev Tools

It keeps about half the text versus a fixed head-plus-tail quota, so who keeps more depends on payload size. The docs say pick-accuracy versus keep-volume were never measured separately, so treat savings as situational.

Pruning sorts by relevance with deterministic head, tail, and high-confidence retention and fails open to passthrough. The jev_gate delivery check is inverted: unclear inputs, refuted claims, and backend failures all escalate.

Tool outputs, fetched page text, task text, skill names, and your jev_ask or jev_gate payloads go to the configured System One endpoint. Keys stay in headers, the ledger stays on disk, and unselected results never leave.

Install by profile bundle, not plain npm install, and run prune.shadow first to log what would have been cut without changing anything. Check /jev-status for skip reasons, call counts, ledger path, and spend.

It stays inert. The plugin mounts normally, every capability is off, and no network request is made. Three ways to set the key are supported with no restart needed: the official SDK's TYPESAFE_API_KEY, the settings card, which writes through the DSH credential store without echoing the value, or an environment variable or .env resolved in the order process env, project env, user env, .env, then managed storage.

sessionCallLimit defaults to 200 judgments per session across every capability, including explicit jev_ask and jev_gate calls. prune.perTurnLimit defaults to 3 and is shared by pruning, screening and skill suggestion; the README measures an uncapped worst turn at 27 judgments and about 8.1 seconds against 0.9 seconds once capped. Input is billed at $0.042 per million tokens and output is free, so the ledger's input tokens are the cost, shown in dollars by /jev-status.

Pruning only fires on read, grep, glob, web_fetch and web_search results above 2000 estimated tokens, and pwsh is deliberately excluded because the irrelevant parts of terminal output are often what you need. Screening covers only web_fetch and web_search text above 300 tokens and warns at a probability of 0.75 without blocking or rewriting anything. Skill suggestion needs a catalog of at least 15 entries and a confidence of 0.3 or above.