Configuration Reference

The defaults.voice.glossary setting provides a list of terms that bias Whisper’s recognition toward correct spellings of domain-specific vocabulary. Terms are passed as an initial prompt to the Whisper model.

In addition to the global glossary, each project can provide a .cc-deck/voice-glossary.txt file in its root directory. This file uses one term per line, with blank lines and lines starting with # treated as comments.

When the attended session switches to a project that has a glossary file, the relay merges the global and project terms automatically. Global terms come first, project terms last. Duplicate terms are removed case-insensitively, with the project’s casing taking precedence.

Whisper’s prompt window is limited to approximately 224 tokens (roughly 800 characters or 50 terms). The relay logs a warning when the merged glossary exceeds this limit but still passes the full list. Whisper truncates from the beginning, so project-specific terms at the end are preserved.

Project glossary files are cached per directory for the lifetime of the relay process. Changes to glossary files require a relay restart to take effect.

Badge icons appear in the sidebar, which expects single-column characters. Emoji and East Asian Wide characters render as 2 columns in most terminals, breaking alignment.

Safe icons include characters like ⋮ (U+22EE VERTICAL ELLIPSIS) or │ (U+2502 BOX DRAWINGS LIGHT VERTICAL).

Badge values may include a color prefix in #RRGGBB:icon format. The hex portion must be exactly 6 valid hexadecimal characters.

When a credential’s required env var is not set in the host environment, the provider is skipped with a warning. The workspace still starts, but the corresponding service will not be available.

Use /cc-deck.capture to detect available credentials automatically. Use /cc-deck.capture --all to auto-accept all proposals without prompting.

The agents field in the build manifest specifies which AI coding agents the workspace supports. Each agent has its own set of backend domain groups that are automatically included in the network policy.

When the agents field is omitted, the manifest defaults to ["claude"] for backward compatibility. Agent adapters declare their required domain groups (e.g., Claude declares anthropic, OpenCode and Codex declare openai), and the policy assembly resolves those groups to concrete endpoints automatically.

The network section of the build manifest controls domain-level egress filtering for containerized sessions.

The build refresh command assembles openshell/policy.yaml deterministically from YAML component files. Each component file defines a set of network endpoints, match conditions, and optional binary restrictions.

The cc-deck build init command renders fixed Containerfile layers into snippet files at <target>/snippets/. These snippets are pre-rendered Go templates with all paths and variables resolved. The /cc-deck.build Claude Code command assembles the final Containerfile by copying snippets verbatim and generating only the variable parts (tool installs, plugin commands, settings) between them.

Snippet files:

The Zellij binary installed by --install-zellij defaults to the latest GitHub release. Pin a specific version with cc-deck config plugin install --install-zellij --zellij-version 0.45.1. Zellij 0.45.0 is the minimum: earlier releases could load the background controller plugin twice on startup.

Define a new group with a domains list:

Use extends: builtin to merge additional domains with a built-in group of the same name. The user-defined domains are appended to the built-in domains rather than replacing them.

This produces a python group containing all built-in Python domains plus pypi.internal.corp. Without extends: builtin, the user-defined group replaces the built-in group entirely.

Use includes to compose a group from other groups. Included groups are resolved recursively.

This produces a group containing all domains from python, golang, and github, plus the literal domain artifacts.internal.corp.

How a recording decides where one speaker stops and the next one starts. Accepted values are pause (default) and voice (case-insensitive). With pause, a new block starts after a silence longer than recording.pause_break, or where Whisper marks a speaker change with a leading dash. With voice, the tinydiarize model detects a change of voice through whisper-cli, even without a pause; this mode recognizes English only and needs the small.en-tdrz model. Pressing s in the idle voice TUI switches the mode for the current session. Any other value produces a configuration error in category voice, and the relay starts with pause.

Whether the relay skips non-speech audio before transcription (default true). When enabled and installed, whisper-server and whisper-cli run with the Silero voice activity model (ggml-silero-v6.2.0.bin in the model directory), which keeps keyboard noise or breathing from turning into phantom words such as "Thank you.". cc-deck ws voice --setup installs the model. Set the value to false to transcribe all audio that passes the VAD threshold, as earlier versions did.

VAD parameters applied while a transcript recording is active. These override the main dictation VAD settings for the duration of the recording and are restored when recording stops. All fields are optional. An omitted field uses the recording default from the table below, not the dictation setting.

Values outside these ranges produce a configuration warning in category voice, and the relay uses the default for that field.