tournament

First Read ${CLAUDE_PLUGIN_ROOT}/skills/model-routing/SKILL.md and apply its routing, foreground agent, validation, quoting, and background-job rules. The rules below are step-specific.

Run one stored plan through independent Codex-routed or Claude-routed contestants in isolated worktrees. Codex contestants run as concurrent detached jobs, while Claude contestants run one at a time in the foreground. The main Claude session owns preflight, containment, evidence collection, independent reviews, winner selection, hand-back, cleanup, and final verification.

Raw slash-command arguments: $ARGUMENTS

Arguments

After reading the routing skill, parse all arguments before loading state:

• --implementer <model> is repeatable and optional. Declaration order defines contestant labels c1, c2, and c3; duplicate models are legal and produce independent samples of the same model. Zero occurrences select the conditional default lineup below. With one, stop, explain that a single contestant is not a tournament, and name /stereo:implement as the single-implementer command. Two or three occurrences define the lineup explicitly. With four or more, stop and state that tournaments are capped at 3 contestants.
• Accept claude:sonnet, claude:opus, claude:haiku, claude:fable, and claude:inherit as Claude contestants. Accept a Codex selection with or without the codex: prefix. Reject claude:session because Claude writes must stay inside the contained stereo:implementer agent; reject unknown claude:* values and codex:claude:*. A Claude contestant runs as one foreground stereo:implementer invocation in its own worktree, so multiple Claude contestants run sequentially.
• --implementer-effort <none|minimal|low|medium|high|xhigh|max> is repeatable with positional pairing. It is legal only when --implementer is present, every contestant is Codex-routed, and its occurrence count equals the --implementer count exactly; the k-th effort pairs with the k-th contestant. Reject a partial list and name the exact implementer and effort counts seen. Reject the flag entirely when --implementer is absent, name the effective default lineup, and point to --effort. When any contestant is Claude-routed, reject the flag, name every such label and selection, say that Claude has no effort dial, and point to --effort as the shared Codex effort.
• --implementation-reviewer <model> selects one shared implementation reviewer. Accept it once and resolve it as explicit flag > workspace implementationReviewer default > claude:fable. claude:session is legal for this role.
• --implementation-reviewer-effort <none|minimal|low|medium|high|xhigh|max> overrides effort for a Codex-routed reviewer and is illegal for a Claude-routed reviewer.
• --effort <none|minimal|low|medium|high|xhigh|max> is the command-wide default for Codex roles without a role effort flag. When no active role is Codex-routed, a command-wide --effort is inert: accept it, report it as inert, and never translate it into a Claude-side control.
• --slot <name> selects the stored plan slot and defaults to default. Define <slotArg> as --slot <slot> for a non-default slot and omit it entirely for default.
• --resume re-enters the recorded incomplete tournament after a crashed, compacted, or closed Claude session. The record owns the lineup, reviewer, every effort, and plan slot. Reject --implementer, --implementer-effort, --implementation-reviewer, --implementation-reviewer-effort, --effort, and --slot with --resume; name the recorded values and tell the user to run without --resume to start over. Reviewer flags remain legal on /stereo:implement --resume, but are rejected here because every contestant must face the same recorded reviewer or the comparison is meaningless.

When --implementer is absent entirely, use the conditional default lineup. Set c1 to the workspace implementer model when its config --json entry has a null invalidReason and resolves to a Codex route; otherwise use the fallback c1 = codex:sol. Set c2 = claude:opus always. Report which default source supplied c1 before launch.