Evidence record / unreal-agent
A dated record of one change, kept so the writing that cites it can be checked. Compiled from the sources listed below by the research run, not written for reading. The judgment lives in the issues that cite it, below.
2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-app
v0.3.1 ships an interactive TUI that runs Bash as you in the current directory with no approval step, defaulting to your ChatGPT login. The project now ships a terminal client: “A minimal terminal client for Harness. Run it from your workspace” (README L3). The Homebrew cask installs it as unreal-agent, uat and unreal-agent-tui. Its workspace is whatever directory you launch it from (os.Getwd(), run.go L190). The model gets Bash, run through your $SHELL in that directory (L245), plus ViewImage, plus skill use when .harness/skills/*/SKILL.md exists. Each command runs as soon as the model asks for it, as your user, and the child process inherits the whole environment, because the shell operation sets no environment of its own (operation/shell.go L522-L530; primitives/process.go L608). There is no approval, permission or policy code in the TUI or in the harness packages it uses. The system prompt is one line asking for a progress update before long-running tools (L254); unlike the runner’s prompt, it claims no sandbox. With no provider flag the TUI defaults to openai-codex (L86). It finds your Codex auth file and asks “Use your existing Codex login?”, offering gpt-6.1-sol (L45) at high reasoning effort (L103). It saves the answer to ~/.config/unreal-agent/preferences.json and reuses it on later launches without asking again (README L20-L29; dialog text in startup.go L127-L130). The Codex client identifies itself as originator: unreal-agent and User-Agent: unreal-agent. It accepts only https://chatgpt.com/backend-api/codex or a loopback base URL, refuses redirects, and reads the token once at startup, so an expired token still means “Renew credentials with Codex, then restart the agent” (README L73; client.go L48-L60 and L90-L103). Every launch starts a new session (store.Create, L203), and the TUI has no resume. Unlike the runner, it does not read the workspace .env (item 3).
Channel: tagged-release. Half: both. Date: 2026-10-06.
Operator consequence: At intake the profile said to avoid running this directly on a workstation; that is now the vendor’s default path. The advice therefore has to become a setup rule: run unreal-agent in a VM or a container whose mounts, credentials and network access you have restricted, because every command the model writes runs immediately as you, with your credentials in its environment. If you accept the Codex prompt, a third-party harness spends your ChatGPT plan on every later launch. unreal-agent -setup changes the choice, and -provider openai (or any other provider) keeps the subscription out. Human attention moves from per-command review, which this tool never asks for, to one decision per session: which box it runs in. That matches the vendor’s stated view that approvals belong in “deterministic environment or sandbox constraints, outside the harness” (launch post, item 9), but the TUI ships no such box.
Receipt
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/cmd/unreal-agent-tui/README.md
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/cmd/unreal-agent-tui/run.go#L84-L104
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/cmd/unreal-agent-tui/run.go#L190-L262
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/cmd/unreal-agent-tui/startup.go#L127-L131
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/harness/operation/shell.go#L519-L530
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/harness/primitives/process.go#L606-L609
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/harness/llm/clients/openaicodex/client.go
- https://github.com/unreallabsai/unreal-agent/blob/v0.3.1/CONTRIBUTING.md
Finding metadata
Run: 2026-10-13-profile-refresh-2026-09-21_2026-10-13-frontier-v0
Finding ID: 2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-app
Profile citations
- unreal-agent / claim / terminal-authority
- unreal-agent / claim / codex-default-no-refresh
Source links
Primary links, including exact changelog lines when available.
- tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / cmd/unreal-agent-tui/README.mdtagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / cmd/unreal-agent-tui/run.go#L84-L104tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / cmd/unreal-agent-tui/run.go#L190-L262tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / cmd/unreal-agent-tui/startup.go#L127-L131tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / harness/operation/shell.go#L519-L530tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / harness/primitives/process.go#L606-L609tagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / harness/llm/clients/openaicodex/client.gotagged commit file2026-10-13-unreal-agent-v0-3-1-ships-an-interactive-tui-that-runs-bash-as-you-in-the-current-directory-with-no-appunreallabsai/unreal-agent / CONTRIBUTING.md
Versioned source: run artifact