cc-bro
Documentation

Getting started

InstallationYour first ticketWorkers and loginProject allowlist

Choose your crew

Worker rosterSettingsCapabilities and mediaCustom CLI templates

Working with tickets

SteeringWorktrees and reviewThe board

Reference

Command referenceUpgradingFAQ

cc-bro / v0.3.1 / Field guide · 07 / 14

Model capabilities and attachments

Choose a model by what it can actually receive through its CLI. Read capabilities.input in cc-bro roster --json before attaching media. A value of true means the model and cc-bro’s current CLI transport both support it; false means one side does not; unknown means support has not been established. A model’s vision, audio or video capability alone is not enough.

Send an image

cc-bro roster --json
cc-bro say visual-check "Describe the layout in this image." --worker codex --model MODEL_FROM_ROSTER --attach ./screen.png --bg --json
cc-bro wait visual-check 1
cc-bro done visual-check

Repeat --attach for multiple files. Paths are resolved from the sending terminal’s directory. Files are copied into cc-bro’s project data before acceptance, so a queued turn does not depend on the original file remaining unchanged. Up to eight attachments are accepted, each up to 20 MiB and all together up to 32 MiB. A follow-up with attachments queues for the next turn; its receipt says so. Unsupported or unknown input capabilities are rejected before the worker receives the prompt. A filename in ordinary prompt text is not a native attachment.

What each transport accepts

CLI / current transport Delivery and limits
Codex app-server Sends images as localImage input blocks. The installed schema also accepts localAudio; audio is enabled only when the selected model reports support. codex exec fallback uses repeated --image; it rejects audio instead of dropping it. Video and PDF have no direct input block in this transport.
pi RPC Sends images with image content blocks (type, base64 data, mimeType). Its input, contextWindow and maxTokens model fields supply exact metadata, without parsing rounded table values. The current RPC prompt type has no audio/video/PDF attachment field.
agy print / stream-json The documented headless input accepts text blocks only. Interactive clipboard support for images and videos does not establish a headless attachment route, so these are unavailable here.
cursor-agent print No direct attachment option is documented by the installed CLI. Tools may read images, and ACP is a separate integration; that is not proof of direct media delivery in the current print adapter. Media input is unknown and explicit attachments are rejected.
custom templates Existing templates pass text arguments or stdin. They do not declare a verified media route, so media remains unknown and explicit attachments are rejected.

Gemini CLI documents image, audio and PDF input through its file tools and @file context. OpenCode documents run --file. These are possible routes for future template adapters, not capabilities that cc-bro’s current generic text template claims. Do not infer video support from the model’s API alone.

Official OpenAI app-server input documentation, pi RPC documentation, agy headless input, Cursor CLI parameters, Gemini file tools, OpenCode CLI.

Limits and descriptions

Use cc-bro roster --worker WORKER --json --full for token limits and their sources. Compact output keeps capabilities.input; when grouped variants disagree, it reports unknown until you inspect the specific full variant.

CLI-reported limits take precedence over the catalog. If only the catalog supplies a value, its source is catalog-api-limit: an API ceiling, not a guarantee of this subscription’s effective limit. Missing values stay unknown. For example, pi’s Codex subscription can report a smaller window than the model’s API catalog. The roster keeps pi’s reported value.

Descriptions, grades and picks are advisory research, not capability detection or a quota guarantee. User notes and rules take precedence. See the roster for local overrides and freshness reminders.