run-agent

Run Callee agents

Use callee when available. Otherwise use the pinned fallback npx --yes @baldaworks/[email protected] for every command in the task.

Discover and select

For each fresh task, inspect the versioned catalog:

callee agent list --json

Resolve a naturally named agent against its exact agent ID or unambiguous description. Inspect the selected tree and every required parameter before execution:

callee agent view "<agent-id>" --json

The selected ID may identify a Role, Sequential, or Loop. Treat all kinds as the same run boundary. Do not invent a separate workflow command.

Execute

Every run requires a real controlling PTY, even when --message supplies the initial prompt. Keep terminal interaction separate from stdout and stderr.

Verify /dev/tty in the same shell invocation before launching Callee; a host tool's tty option alone may not create a controlling terminal. If test -r /dev/tty && test -w /dev/tty fails on Linux, allocate one with util-linux script:

script -qefc 'callee agent run "<agent-id>" --message "<task>" > /tmp/callee-artifact.txt 2> /tmp/callee-diagnostics.txt' /dev/null

On BSD/macOS, use script -q /dev/null /bin/sh -c '<callee command with the same redirections>'. Keep /dev/tty attached for prompts, use unique temporary output paths, inspect the wrapper's exit status, and read the artifact and diagnostics files separately. Do not first launch Callee without a verified controlling terminal; that produces an intentional failed run rather than a useful capability probe.

callee agent run "<agent-id>" \
  --message "<task>" \
  --param "<effective-node-id>.<name>=<value>"

Use --param-file "<effective-node-id>.<name>=<path>" for exact multiline values. Supply known parameters shown by agent view; let Callee collect the rest just in time through the PTY.

Read questions and permission requests from the terminal and answer through the same terminal. A Role inside a workflow may enter REPL mode; its stderr lifecycle has one entering repl / exiting repl pair, with every await turn inside that pair. Do not send quit, exit, /done, or a synthetic completion marker; the Role selects control through Callee's injected protocol.

Wait for automatic root completion. The sole successful root artifact is written to stdout only after provider cleanup succeeds. Info lifecycle events and diagnostics are written to stderr, so use the exit status rather than stderr emptiness to determine success. Treat empty stdout on failure as intentional.

Callee v1alpha1 does not define Parallel; do not imply parallel workflow semantics or merge PTYs.

Report results

Return the final stdout artifact and a concise capability trace. Then read the structured lifecycle events from stderr and include these execution metrics in the final response:

• From the final agent run finished event, report agent_duration, agent_wait_duration, agent_token_usage, and every emitted numeric agent_*_tokens field.
• From every Role agent finished event, report its effective ID and visit together with role_provider, role_model, role_reasoning, role_token_usage, every emitted numeric role_*_tokens field, and role_duration plus role_wait_duration when present.

Keep repeated and nested Role visits separate. Preserve complete, partial, or unavailable token status and do not invent numeric or duration fields that the event omitted. On failure, report the exit status and any metrics emitted before termination. Do not expose provider session IDs, internal handles, or raw terminal transcripts.

Setup

callee setup <codex|claude|grok|copilot|opencode|cursor>

Do not add Gemini, a server transport, a Callee thread store, or handle binding.