Jev Context Compaction

Review the record before it leaves your control

The examples are synthetic input, not model answers. The original stays available throughout. A run sends the text you choose to this site’s server and the configured Jev provider.

{
  "task": "Fix the failing pagination test while keeping the public API stable.",
  "constraints": "Do not change authentication or production data. Preserve the test failure and the relevant source code.",
  "entries": [
    {
      "kind": "message",
      "id": "user1",
      "role": "user",
      "text": "The dashboard pagination test fails after a refactor. Please find the regression."
    },
    {
      "kind": "tool",
      "id": "baseline",
      "tool": "git status",
      "input": "git status --short",
      "result": "main clean; no uncommitted changes",
      "pinned": false
    },
    {
      "kind": "tool",
      "id": "search",
      "tool": "rg",
      "input": "rg 'pagination' src tests",
      "result": "src/query/paginate.ts:17 paginate(items, page)\ntests/pagination.test.ts:42 expected page 2 to begin with item 11",
      "pinned": false
    },
    {
      "kind": "message",
      "id": "assistant1",
      "role": "assistant",
      "text": "I found the shared pagination helper and will inspect its boundary handling."
    },
    {
      "kind": "tool",
      "id": "oldlog",
      "tool": "test runner",
      "input": "npm test -- old-suite",
      "result": "An older unrelated color-snapshot test passed. Its output is no longer relevant to the pagination regression.",
      "pinned": false
    },
    {
      "kind": "tool",
      "id": "source",
      "tool": "read file",
      "input": "sed -n '1,80p' src/query/paginate.ts",
      "result": "export function paginate(items, page, size) { const start = page * size; return items.slice(start, start + size); }\nThe UI uses one-based page numbers.",
      "pinned": true
    },
    {
      "kind": "tool",
      "id": "failure",
      "tool": "test runner",
      "input": "npm test -- pagination.test",
      "result": "FAIL pagination.test: page 2 expected item 11, received item 21. Current call stack points to src/query/paginate.ts:17.",
      "pinned": false
    }
  ]
}

{
  "task": "Fix the failing pagination test while keeping the public API stable.",
  "constraints": "Do not change authentication or production data. Preserve the test failure and the relevant source code.",
  "entries": [
    {
      "kind": "message",
      "id": "user1",
      "role": "user",
      "text": "The dashboard pagination test fails after a refactor. Please find the regression."
    },
    {
      "kind": "tool",
      "id": "baseline",
      "tool": "git status",
      "input": "git status --short",
      "result": "main clean; no uncommitted changes",
      "pinned": false
    },
    {
      "kind": "tool",
      "id": "search",
      "tool": "rg",
      "input": "rg 'pagination' src tests",
      "result": "src/query/paginate.ts:17 paginate(items, page)\ntests/pagination.test.ts:42 expected page 2 to begin with item 11",
      "pinned": false
    },
    {
      "kind": "message",
      "id": "assistant1",
      "role": "assistant",
      "text": "I found the shared pagination helper and will inspect its boundary handling."
    },
    {
      "kind": "tool",
      "id": "oldlog",
      "tool": "test runner",
      "input": "npm test -- old-suite",
      "result": "An older unrelated color-snapshot test passed. Its output is no longer relevant to the pagination regression.",
      "pinned": false
    },
    {
      "kind": "tool",
      "id": "source",
      "tool": "read file",
      "input": "sed -n '1,80p' src/query/paginate.ts",
      "result": "export function paginate(items, page, size) { const start = page * size; return items.slice(start, start + size); }\nThe UI uses one-based page numbers.",
      "pinned": true
    },
    {
      "kind": "tool",
      "id": "failure",
      "tool": "test runner",
      "input": "npm test -- pagination.test",
      "result": "FAIL pagination.test: page 2 expected item 11, received item 21. Current call stack points to src/query/paginate.ts:17.",
      "pinned": false
    }
  ]
}

What the local policy actually does

A shorter record is a proposal, not a guarantee

Before and After use the same JSON serialization for character counts. Expand a decision to inspect the full removed result. Restoring shows the complete original; exporting never writes to an agent conversation. Historical tool output can be irreplaceable or describe a side effect, so human review matters even when a tool could be called again.

Keeping the original intact does not make a pruned export lossless. Fewer characters do not establish better answers or lower total cost. The threshold and preview length are application policy; changing them only recomputes the result locally.

This browser tool and a host plugin solve different parts

What this demonstration can establish

TypeSafe documents Noul as a probability for a yes/no question and supports multiple questions over one State. This tool uses two independently judged questions per candidate. Noul has no separate Choice confidence. We do not treat a returned probability as proof that deleting evidence is safe.

The source review used fast-jev-compaction at e3f262a, Winnow at 51d80b9, and the Codex proof of concept at 7147a4b. The latter’s maintainer explicitly does not recommend it for real work. A published npm package exists for the upstream experiment, but its host compatibility and installation path were not independently exercised here; consult the upstream repository and current host documentation before trying it.

This is not a replacement for native model compaction or internal state handling, and it does not generate summaries or promise lossless compression. Community interest does not establish production quality. Model judgments for this new scenario have not been benchmarked here.