Insane Jev Uses You Need To Use Right Now

Highlights

• TypeSafe's Jev is a high-speed decision model that chooses from pre-defined options rather than generating text token-by-token.

• Jev charges only for input tokens while returning decision outputs for free, drastically lowering API costs.

• Replacing default Claude Code compaction with the Fast Jev Compaction plugin yields context summaries in under a second.

• Using Jev as a file-ranking engine scores and orders 35 project files in 1.6 seconds to eliminate context window bloat.

• An automated hook sends edits to Jev and halts file creation if Jev determines with at least 80% confidence that a rule was violated.

Timeline

1. 00:00 - 03:13

   Core Mechanics and API Setup of the Jev Decision Model

   Unlike generative language models like Claude or GPT that produce answers token-by-token, Jev evaluates all given options simultaneously. This architectural difference provides sub-second evaluation times and eliminates generation costs. Access is managed via API gateways like OpenRouter or Vercel AI Gateway due to paused direct signups on TypeSafe.

   • Jev operates as a non-text decision model that selects between binary choices, lists, or score ranges.
   • Output includes a confidence value indicating how sure the model was in its choice.
   • Input tokens incur charges, but returned decision responses are completely free.
2. 03:13 - 04:56

   Fast Jev Compaction for Agent Context Windows

   Standard summarization consumes significant time and tokens as logs grow. Fast Jev Compaction evaluates previous conversation turns to filter out irrelevant history in under a second. The 25% threshold guardrail ensures API calls target sessions where compaction yields significant execution gains.

   • Standard compaction sends entire conversation logs back to full models for summarization.
   • The Fast Jev Compaction plugin replaces long summary prompts with rapid importance decisions.
   • Sessions with less than 25% context usage bypass Jev and default to standard compaction.
3. 04:56 - 07:46

   Enforcing Test Coverage and Streamlining Skill Selection

   Agents often fail to track documentation rules or struggle with context clutter when managing dozens of skills. By using Jev as a verification layer, missing tests get flagged immediately without relying on the main model to self-audit. For skill management, Jev reads system skill metadata and injects only the single relevant skill for a given user prompt.

   • A post-edit hook asks Jev whether access control rule changes have corresponding tests.
   • The skill picker hook matches prompts to skill descriptions before passing tasks to Claude.
   • Pre-selecting skills prevents agents from evaluating every available skill inside the main context.
4. 07:46 - 09:58

   Accelerating File Discovery and Pre-Filtering Code Reviews

   Replacing default sub-agents like Explore with a Jev-based skill prevents expensive models from filling context windows with unnecessary file reads. In code review workflows, Jev screens changes against core compliance questions. If all answers return negative, the system executes a brief review pass; positive or uncertain responses trigger full model inspection.

   • A custom skill pairs standard keyword search with Jev batch scoring to locate project files.
   • Jev ranks candidate files 20 at a time and evaluated 35 files in 1.6 seconds.
   • A preliminary review gate uses seven yes/no questions to determine if code requires full audit.
5. 09:58 - 12:38

   Automated Browser Interaction and Rule Enforcement

   Browser testing agents utilize Jev to decide which links or buttons to click next based on generated page element lists, leaving final verification to the main model. To prevent agents from straying from specification files, pre-edit hooks pass planned code changes to Jev alongside project rules. Attempted violations, such as client-side direct data reads, get blocked immediately with exact error feedback.

   • Jev selects interactive UI elements during automated browser testing to reach target states.
   • Pre-edit hooks evaluate file modifications against project rules defined in documentation.
   • Edits flagged with at least 80% certainty of violating rules are blocked before file creation.