roster

roster — Who Does the Work

Prerequisite: this pillar is optional. If ~/.o9k/roster.json is missing, do not invent multi-agent flows — use dispatch path A (in-host RESULT subagents) only. Users who never enabled the roster at /o9k-init should never see these commands.

Model choice is config + code, not judgment. One primary role per task; the role's chain IS the fallback mechanism. Never pick a model by reasoning — that's how model-family favoritism happens.

Chain entries are CLI×model cells, not models alone:

• "cursor:grok-4.5-high" or { "cli": "hermes", "model": "deepseek-v4-pro" } pins the pair
• bare "claude-sonnet-5" still works → uses models[m].cli[0]

Effort (CLI-native)

Precedence (highest wins): chain-entry.effort > roles.<role>.effort > models.<id>.effort > omitted. Values are raw per-CLI strings — never mapped: Claude low|medium|high|xhigh|max (--effort), Codex low|medium|high|xhigh|max|ultra (-c model_reasoning_effort=…; per-model support — check codex debug models). When effort is unset, the {effort} template element and its preceding flag are dropped from argv. Effort set for a CLI whose cmd has no {effort} → stderr warning, spawn proceeds. Put effort on chain entries (CLI-pinned) unless the model has exactly one CLI. refresh --apply rewrites chain heads as cli:model strings → chain-entry effort on a replaced head is lost; use roles.<role>.effort for durable role depth.

"advisor": {
  "effort": "max",
  "chain": [
    { "cli": "claude", "model": "claude-fable-5", "effort": "max" },
    { "cli": "codex", "model": "gpt-5.6-sol", "effort": "ultra" },
    "claude:claude-opus"
  ]
}

All commands below: ROSTER="node <marketplace>/plugins/o9k-roster/scripts/roster.mjs" (in Claude Code: node "${CLAUDE_PLUGIN_ROOT}/scripts/roster.mjs" when this plugin is active). Cross-CLI mailbox runs: RUNS="node <marketplace>/plugins/o9k-roster/scripts/runs.mjs". No config yet → $ROSTER init, then tell the user to curate ~/.o9k/roster.json.