Skip to content

runtime.md

Latest commit

 

History

History
103 lines (79 loc) · 5.73 KB

runtime.md

File metadata and controls

103 lines (79 loc) · 5.73 KB
Raw

Agent Runtimes

Logos supports three runtime families from the Agent panel:

  • Logos: the built-in OpenAI Responses runtime with workspace tools, permissions, resumable local sessions, cancellation, and protocol tracing.
  • Claude: the existing Claude Agent SDK integration.
  • ACP: manually configured agents and agents discovered through the canonical ACP Registry.

Logos Authentication

Open Settings > Model APIs & ChatGPT and choose one method:

  • Connect ChatGPT opens the ChatGPT Plus/Pro PKCE login flow in the browser.
  • Use API key stores an OpenAI API key for direct Responses API access.

OAuth tokens and API keys are encrypted with Electron safeStorage and written to the operating system's per-user Logos data directory. They are not stored in settings.json or renderer local storage. Logging out removes the encrypted file.

agent.openaiBaseUrl applies only to API-key requests. ChatGPT subscription requests use the Codex Responses endpoint selected by the OAuth transport.

OpenAI Models

The built-in runtime defaults to gpt-5.6-sol and also exposes the GPT-5.6 Terra and Luna tiers. Each tier supports none, low, medium, high, xhigh, and max reasoning effort. With Auto selected, GPT-5.6 uses medium effort.

Each GPT-5.6 tier has a Fast mode that requests priority processing. Pro mode uses OpenAI's Pro reasoning mode and is available with API-key authentication; the ChatGPT Codex endpoint does not support Pro mode or the bare gpt-5.6 alias. Logos therefore uses the explicit gpt-5.6-sol model ID as its cross-authentication default.

Workspace Tools

The Logos runtime provides Read, Glob, Grep, Write, Bash, Skill, MCP, DAP, the legacy DAP_REPL shortcut, and the loop-control tool Finish. The exact tool contract and the mode-specific system prompt are shown at the top of every Logos thread before its first prompt. A turn remains active until the model explicitly calls Finish; an assistant-only planning message therefore cannot end a task before tool work begins. Paths are constrained to the active workspace, including real-path checks that reject symbolic-link escapes. Write follows the selected permission mode. Every Bash command, MCP connection/tool call, mutating DAP action, and DAP_REPL expression requires one-time approval, including in bypass mode. Threads are bound to the workspace where they started and cannot silently continue in another folder.

Bash runs /bin/bash on macOS/Linux and an installed bash.exe on Windows. Windows users can provide Bash through Git for Windows or WSL; Logos does not reinterpret Bash syntax through cmd.exe or PowerShell.

Grep accepts JavaScript regular expressions and an optional file glob. Skill discovers Agent Skills under .agents/skills, .claude/skills, and .logos/skills in the project, plus enabled user setting sources. Skill reads are restricted to the selected skill directory.

The built-in runtime reads MCP server definitions from workspace .mcp.json. Both the standard mcpServers key and VS Code-style servers key are accepted and merged, with mcpServers winning duplicate names. Stdio and Streamable HTTP transports are supported. Reading the configuration does not launch a server; a server is connected only after the user approves list_tools or call_tool. See MCP project configuration for the complete schema, external-client file mappings, and automatic setup offered after opening a new folder.

DAP controls the same debug sessions as the workbench UI. It can discover and start launch.json configurations, list/stop/restart sessions, continue, pause and step, replace source breakpoints, inspect threads/stacks/scopes/variables/source, evaluate expressions, and send advanced raw DAP requests. When exactly one session is active it is selected automatically; otherwise the tool requires an explicit session id. Read-only inspection is available in Plan mode. Process control, breakpoint changes, evaluation, launch, and raw requests require one-time approval. DAP_REPL remains as a compatibility shortcut for DAP evaluation.

Third-party Agents can use the same control plane through the project logos-debug MCP server. Its stdio proxy discovers an authenticated loopback bridge published by a running Logos process; it only connects when that Logos window has the same canonical workspace open. Discovery files are private to the current OS user and contain a per-launch random token. See Debugging with launch.json.

ACP Registry

Logos refreshes the canonical registry with ETag revalidation and a five-minute cache. The last valid snapshot remains available when offline.

Binary distributions are downloaded lazily into the Logos user-data directory. Logos verifies SHA-256 when supplied, rejects unsafe archive paths and links, and supports raw binaries, ZIP, tar.gz, tgz, tar.bz2, and tbz2.

npx and uvx distributions are shown as available only when the corresponding package runner can be found on the host. Package specifications and arguments are passed as separate argv values; no shell command is constructed.

Registry ACP processes receive a restricted inherited environment so credentials from the Logos process are not automatically exposed to third-party agents. ACP agents are still native local processes and should only be run when their publisher is trusted.

Agent Debug

Each active thread exposes a collapsed Agent Debug timeline. It includes Logos request lifecycle, model output items, tool calls/results, errors, and ACP lifecycle, session updates, and stderr. Trace payloads are capped in the renderer and are not persisted with conversation history.