Command Code over ACP (Agent Client Protocol)

Run recipe:

cyberine run --fleet command-code-acp "Read the repository and explain its entry points"

The opt-in command-code-acp worker runs command-code --no-auto-update acp on stealth/glyph-cluster:free. Install and log in to Command Code first. It stays out of automatic fleet discovery and keeps the existing command-code worker available. An explicit CYBERINE_BUILD_FLEET=command-code-acp also selects it; no new environment variables are introduced.

Extend recipe: use makeAcpAdapter from src/adapters/AcpAgentAdapter.ts with id, provider, command, optional args, modeId, modelId and permissionPolicy (worktree by default, or reject). A task model pin overrides modelId. For agents that load instruction files themselves, set selfReadsInstructionFilenames to avoid injecting those files twice. The default starts no bypass mode, disables client filesystem and terminal services, and allows permissions only for paths inside the task cwd after following symlinks. Requests with no determinable path are rejected. Host sandbox options use the existing CLI-worker contract.

Goose *_acp subscription bridges are refused at construction. Pi and OpenCode require an explicitly configured API-key provider and authMode: 'api-key'; subscription-OAuth configurations and ambiguous auth are refused. The caller must configure the binary to use that key provider. Cyberine never extracts subscription credentials.

Known Limitations and Deferred Work: ACP provides no per-turn token counts. Context size is not token usage; usage stays unknown unless the agent reports an optional USD cost. The spike's free-model prompt failed with 400 You have insufficient credits, so the live free-lane run and FLEET-BENCH diff remain pending the owner's Command Code account check. That refusal returns an error naming insufficient credits and parks this lane in the cooldown store. An offline fake-agent check is available as npm run selftest:acp-permission-gate; it needs no CLI login, network or model.