Claude Code Coding Agent

Claude Code Coding Agent

Use Claude Code only after the user or active workflow explicitly selects it as the external executor.

Minimal Workflow

1. Set <launcher> to the requested Claude Code binary, absolute path, alias, or wrapper. Preserve a provided wrapper; it may inject model, authentication, or permission settings, including bypass permissions. Treat wrappers as opaque: discover their contract only by invoking <launcher> --version or <launcher> --help; do not run inspection commands that may print an alias or function definition, open or print wrapper source, or dump its environment because it may contain credentials.
2. Run <launcher> --help before composing version-sensitive flags. Use other subcommand help only when the task needs it.
3. Choose the model and thinking effort deliberately, following the guidance below and the choices currently exposed by the launcher.
4. Use monitor mode by default for any task that may take time. Use final mode only when the task is clearly trivial and short.
5. Run from the intended repository or workspace, pass a bounded task contract, and wait for the external process to finish.
6. Inspect the resulting diff, tests, and final answer before claiming success.

Model and Effort

Honor explicit model or effort choices. Otherwise inspect current help before launching:

• For routine, bounded work, prefer a balanced model and moderate effort.
• For deep review, ambiguous debugging, cross-module design, or other high-risk work, prefer a frontier model and high or maximum supported effort.
• Use the highest tier only when its quality benefit justifies the extra latency or cost, and after checking whether that tier has additional behavior.

Do not hardcode model names or effort levels from this skill; the launcher's current help is authoritative.

Monitor Mode (Default)

Put the bounded task contract in a task-specific TASK_PROMPT variable and pass it through stdin. This remains unambiguous when options such as --tools accept multiple values and would otherwise consume a trailing positional prompt.

Use Claude Code's semantic JSON stream without requesting partial messages:

printf '%s' "$TASK_PROMPT" | <launcher> --print --output-format stream-json --verbose

For a read-only review, when current help supports these tools, keep the prompt on stdin and restrict capabilities explicitly:

printf '%s' "$TASK_PROMPT" | <launcher> --print --output-format stream-json --verbose --tools Read Grep Glob

Never append a bare positional prompt after --tools, --allowedTools, --add-dir, or another variadic option.

Start the command with the host's long-running process facility. Keep the task ID returned by the host, then use the host's wait, poll, or resume facility to read only newly available output while the process runs.

Reduce the stream to small liveness signals such as:

running — process alive
working — Read completed
running — no new semantic event; process alive
completed

Ignore raw thinking/reasoning and token deltas. Do not add --include-partial-messages for ordinary monitoring. Treat the terminal result event together with process exit as completion evidence; do not kill a live process merely because it has produced no recent semantic event.

Final Mode

For a clearly trivial, short task, wait for one final response:

printf '%s' "$TASK_PROMPT" | <launcher> --print

Task Boundaries

• State explicitly whether the task may edit files. Keep review and explanation prompts read-only and findings-first.
• Follow the current help and wrapper contract for permissions and tool access.
• Do not silently create worktrees, commit, push, deploy, or widen task scope.
• Put optional captures in the system temporary directory. Cleanup is optional.

For event mapping, terminal-state handling, wrapper details, and current capability discovery, read references/monitoring.md.