star-auto

Drive the workflow toward a goal

Read .agents/commands/star-auto.md from the current project root and follow it as the authoritative procedure. Write the user-facing wording in the language resolved under conventions §7.6: Chinese when .env sets STAR_LANG=zh, or when it is unset and the conversation is in Chinese. .agents/commands/star-auto.md is the only procedure; its decisions do not change with the language.

Adapt its invocation spelling and model routing for Codex:

• $star-auto <goal> [stop=<stop line>] [involve=<level>] is this command.
• $star-<name> <argument> is the spelling where the shared file writes /star-<name> <argument>.

For every run that this command starts, resolve its tier and any mode exception under conventions §10.8. Read the corresponding STAR_PLAN_MODEL, STAR_EXEC_MODEL, or STAR_READ_MODEL value once with the opening .env load. In a comma-separated value, take codex:<model> first, otherwise an untagged model; ignore entries tagged for another harness. Split a recognized final @<depth> from the model, retaining an unrecognized @ suffix as part of the model name. An empty result names no model or depth.

When the resolved model is non-empty, this Codex runtime can name a subagent model, and either that model is not an alias of the active session's actual model or the entry carries a supported per-dispatch depth, start the run with spawn_agent: pass the model explicitly and, when the tool and selected model accept it, pass the depth as reasoning_effort. A usable explicit depth triggers this fresh run even when the model is unchanged. Pass fork_turns: "none" whenever the runtime exposes that field, including for blind reads: this runtime does not allow a model or effort override with a full fork. Use the same rule for both unmarked and † skills. If a dispatch rejects the requested depth, first confirm it started no work; then retry without reasoning_effort only when the model itself differs, otherwise retain the shared local route, and state once that the depth was not applied. With no routing difference, retain the shared procedure: load and follow an unmarked project star-* skill here, or dispatch the † skill's subagent as the shared file requires, without a model or effort override.

Every model- or effort-routed subagent brief is self-contained: tell it to read the selected skill's project SKILL.md in full; include the original skill invocation, its resolved tier=<name> and involve=<level> tokens, auto=unattended when this invocation carries it, and the language resolved from STAR_LANG or the dialogue. Include the active session's actual provenance as a model id only. Never pass the parent session's model-resolver command or treat its output as the child's model provenance: the child resolves and records its own provenance from its own session context. Record the requested depth, or default, beside the tier and actual child model. These routing rules inherit the shared procedure's confirmation, STOP-line, sandbox, and approval limits; they authorize no additional action.

If .agents/commands/star-auto.md is missing, report that the project does not contain the STAR auto procedure instead of guessing from the plugin package.