> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryvoss.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# October 25, 2026

> Go terminal client, honest confidence, opt-in Jev judgments, swarms, and a durable protocol for Voss ADE.

A lot shipped since June. The terminal client was rebuilt in Go, confidence numbers now say where
they came from, and `voss serve` became the product contract that SDKs and the separate Voss ADE
desktop app build on.

## Go terminal client

Bare `voss` now opens a full-screen Go client that runs the harness through `voss serve`. It has
the slash and `@` file palettes, permission prompts, session resume with a searchable picker,
`/model` and `/auth` pickers that persist, memory and recall commands, code lookup with
`/symbol`, `/refs`, and `/refresh`, and read-only `/probable`, `/btrace`, and `/vdiff` panels.

The Python Textual app has been removed. `voss chat` and `VOSS_USE_TUI=0 voss` give the plain
line-by-line chat, which is also the fallback when the Go binary is missing. The npm package
bundles the Go client.

## Confidence you can trust

A plain `ask()` reply used to get a made-up confidence of 0.9, so gates on it always passed. Every
`ProbableValue` now records its source: a Jev judgment, the model's self-report, a value you
supplied, or unknown. Unknown never passes a gate, the compiler warns (ANLY004) on gates over
`ask()` values, and footers and inspectors print `unknown` instead of a number. For a real
probability, use `judge(...).to_probable()`.

## Jev judgments

Judgments are opt-in per project and always advisory:

* `voss judge review` checks work against acceptance criteria and labels each one supported,
  contradicted, unclear, or no answer.
* With `judgments.review: shadow`, swarm candidates get the same check in the background.
* With `judgments.code_recall`, pre-turn code recall can be reranked. The locked evaluation
  passed every release gate, and `voss judge activate code_recall` turns it on.
* Receipts appear in `voss audit`. `VOSS_JUDGMENTS=off` stops every call.

`voss eval --ab` runs the task suite twice with one setting changed and compares the results.

## Swarms

A swarm runs a coordinator, builders, and a reviewer through `voss serve`. Builders can be native
agents or headless Claude Code and Codex sessions, each in its own worktree with owned files.
Finished tasks become candidates for review.

## Protocol and SDKs

The `voss serve` REST + SSE protocol is now the product contract, checked against committed
OpenAPI and event schemas in CI. The Rust, TypeScript, and Go SDKs launch `voss serve` by
executable and need no source checkout. New in the protocol:

* Capability discovery at `GET /capabilities`.
* A durable task journal with cursor replay and idempotent command IDs.
* Managed OAuth-only execution, with per-harness availability and preflight.
* Execution previews: plan a task graph, accept its exact revision, then run it in dependency
  order.
* Recorded task graphs with per-attempt output, and acceptance evidence for review.

## Voss ADE is its own product

The desktop app moved to its own repository. It integrates with Voss only through the released
executable, a compatible SDK, and the protocol.

## Also new

* MCP servers and skills are discovered from your Claude and Codex config (`voss mcp list`,
  `voss skills`).
* `AGENTS.md` and `CLAUDE.md` are injected into every native run within a budget
  (`voss instructions show`).
* `voss audit` renders one complete read-only report per run.
* `voss shell-init` prints shell integration for terminal hosts.
* Retained memory can use a Laravel memory service instead of the local store.
* Install with pip from GitHub. The `voss` package on PyPI is an unrelated project.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.