What Problem This Solves

Closes #77165.

New Control UI sessions otherwise fall back to first-message snippets or session ids, which makes a busy sidebar difficult to scan. Small internal model tasks also had no shared lower-cost model setting, so title generation had to reuse ad hoc operation-specific routing or the agent's primary model.

This supersedes the older implementation attempt in #77225. The broader periodic re-titling proposal in #99583 remains out of scope.

Why This Change Was Made

• Add agents.defaults.utilityModel plus agents.list[].utilityModel as a shared model route for short internal tasks. Explicit operation overrides win, then the per-agent utility model, the default utility model, and the agent primary model.
• Generate one asynchronous dashboard-session title from the first non-command message and persist it as displayName, which the existing sidebar/session picker already consumes.
• Preserve explicit labels, display names, subjects, and origin labels. Historical sessions, replaced/reset sessions, duplicate concurrent sends, and title failures do not overwrite metadata or block chat.
• Route the existing Telegram DM topic-label and Discord generated thread-title paths through the same utility model selection.
• Keep utility-only providers discoverable at Gateway startup and include utility refs in configured-model catalog scans.

User Impact

New dashboard sessions get a concise 3-6 word sidebar title after their first message. Existing sessions are not backfilled, explicit/manual names are never replaced, and title generation runs off the chat critical path.

Users can set a smaller model globally or per agent:

{
  agents: {
    defaults: {
      utilityModel: "openai/gpt-5.4-mini",
    },
  },
}

Utility tasks are separate model calls. Dashboard titles send at most the first 1,000 characters of the first non-command message to the selected provider; when utilityModel is unset, the agent primary model is used. The docs call out the cost and data-handling boundary.

Evidence

• Rebased onto main at 0221544190; PR head 3657c373c6.
• Fresh Codex autoreview after fixes: no accepted/actionable findings, overall correct (0.90).
• Focused exact-head tests: 325 passed across model-ref discovery, utility-model selection, label generation, dashboard-title persistence/races, chat send integration, provider activation, and Discord title generation.
• Local generated-surface checks: base config schema, config docs baseline, plugin SDK API baseline, docs formatting/MDX/links/map, and git diff --check passed.
• Live macOS Gateway proof on feature head 37b2e3edeb (later PR-head changes only refresh generated fingerprints and rebase the same feature tree):
  • primary model remained openai/gpt-5.5
  • temporary utility model codex/gpt-5.4-mini was independently available
  • first message Plan a safe migration of a family photo archive to an encrypted SSD with rollback checkpoints. persisted sidebar displayName: "Encrypted SSD Photo Migration"
  • synthetic session deleted, prior absent utility-model config restored, Gateway healthy, live checkout clean
• Exact-head Testbox changed-surface gate passed: https://github.com/openclaw/openclaw/actions/runs/28719635907 (tbx_01kwqf6aytwmxx01rj6yhqf03c).
• Exact-head Testbox Plugin SDK API fingerprint passed: https://github.com/openclaw/openclaw/actions/runs/28719764582 (tbx_01kwqfg50fm0g7810ep79rqrby).
• Docs, schema, and config-doc gates passed before the mechanical SDK fingerprint refresh: https://github.com/openclaw/openclaw/actions/runs/28719321621 (tbx_01kwqefahhag92a4mqw67zp20g).

Thanks @Juliangsm for the feature request and @zhangguiping-xydt for the original PR work.