Coding agents¶

cafleet supports three coding-agent binaries inside member panes: claude (Claude Code), codex (OpenAI Codex CLI), and opencode (opencode.ai). The backend is selected at fleet-create and member-create time via --coding-agent {claude,codex,opencode}; the default is claude.

Backend resolution¶

Backend	Spawn command	Notes
`claude`	`claude --permission-mode dontAsk --name <member-name> <prompt>`	Pane title derives from `--name` via Claude Code's terminal-title escape.
`codex`	`codex --ask-for-approval never --sandbox workspace-write <prompt>`	No `--name` analog; pane title is whatever `codex` emits by default. Operators locate panes via `cafleet member list`.
`opencode`	`opencode --agent cafleet --prompt <prompt>`	No `--name` analog. Long-lived TUI bound to the `cafleet` agent definition at `~/.opencode/agents/cafleet.md`, written on first spawn if absent. Safety floor matches Claude Code's `dontAsk` posture (catch-all-allow + specific-deny ruleset enforced by opencode's own permission system) — NOT Codex's kernel sandbox.

All three backends honor a leading-! shell shortcut on the coding agent's input line, so cafleet member exec keystrokes ! <command> + Enter into any pane shape and the command runs natively. The --coding-agent value is one of claude / codex / opencode and is recorded in the placement's coding_agent column.

Mixed-backend teams are allowed: a single Director may spawn claude, codex, and opencode members in the same fleet with no broker-level differences.

For cafleet fleet create, the --coding-agent flag is operator-declared metadata only — cafleet does not spawn the root Director's coding-agent process and cannot auto-detect what is already running, so the operator declares which binary is in the pane. For cafleet member create, the flag both selects which backend is spawned AND is recorded as placement metadata.

Model selection¶

cafleet member create accepts an optional --model <string> that is forwarded to the spawned backend binary via that binary's --model flag (e.g. --model sonnet for claude, --model gpt-5.4-mini for codex, --model anthropic/claude-sonnet-4-6 or --model opencode/big-pickle for opencode). When the flag is omitted, no model tokens are emitted and each binary uses its own default model.

Validation is per-backend: claude and codex pass any string through verbatim — the binary itself rejects unknown models, so newly released models work without a cafleet release. opencode requires the <provider-id>/<model-id> format and rejects anything else at create time with exit 2, before any agent registration or tmux side effect.

The value is spawn-time only: it is NOT recorded in agent_placements, requires no migration, and does not appear in cafleet member list output.

Known asymmetries (intentional non-goals)¶

Pane title. Only the claude spawn argv carries --name, so codex and opencode panes do not display the member name in their pane title. Use cafleet member list to find a specific member's pane id; the pane_id column is ground truth for all three backends.
Bash-disable parity. Neither codex nor opencode has a --disallowedTools analog, but this does not affect the routing protocol — see Bash routing.
Sandbox isolation. Only codex provides OS-level (kernel-enforced) isolation via --sandbox workspace-write. Both claude and opencode rely on deny-list-only safety floors; operators who need kernel-enforced isolation should use the codex backend.

Operational details for codex members — including the codex CLI version pin, install pointer, and verification recipe — live at Codex members. The equivalent for opencode lives at Opencode members, including the preset materialization protocol and the refresh recipe for upgrading the preset after a CAFleet release.