fix(agents): route worker models by available profiles

[guustavomv]

Summary

• resolve semantic agent profiles (fast, review, coding, testing, and balanced) against the authenticated Verboo /models catalog
• preserve the full catalog model ID, including plan prefixes such as max/qwen3.6-27b
• make tool-specified Claude aliases (haiku, sonnet, and opus) inherit the parent model on non-Claude-native providers
• support model: profile:<name> in forked and inline skills while keeping existing external-provider routing compatible
• route the built-in read-only Explore agent through the fast profile and bound it to four turns
• document profile configuration and add routing diagnostics

Motivation and runtime evidence

The planner itself was running on gpt-5.6-sol, but delegated calls could fail or take a long time because the harness forwarded model identifiers that the active provider did not expose.

Observed in real Verboo transcripts before this fix:

• worker-tests sent the unqualified ID qwen3.6-27b; the authenticated catalog exposed max/qwen3.6-27b, so the call failed before editing any file
• after switching the skill to a semantic value, an old runtime forwarded profile:balanced literally instead of resolving it
• the built-in Plan agent attempted sonnet and then opus; both finished with 0 tool uses, 0 tokens, and 0 seconds because those aliases were unavailable on the current provider

Observed with the catalog-aware Explore route:

• the main planner remained gpt-5.6-sol
• Explore requests expressed as haiku executed with the available fast model max/deepseek-v4-flash

This change makes the same behavior general across account plans: profiles choose the best preferred model that is actually returned by /models, then preserve its canonical ID. If no candidate is available, the agent inherits the parent model unless first-available is explicitly configured.

Impact

• user-facing impact: plans no longer stall on unavailable Claude aliases or fail because a plan-qualified Verboo model prefix was omitted; worker and reviewer skills can use portable semantic profiles
• developer/maintainer impact: routing decisions are centralized, schema-validated, covered by focused tests, and visible in debug logs

Testing

• bun run build
• bun run smoke
• focused tests: bun test src/utils/model/agent.test.ts src/services/api/agentRouting.test.ts src/query/model.test.ts src/tools/SkillTool/SkillTool.test.ts src/skills/loadSkillsDir.test.ts (64 pass, 0 fail)

Notes

• provider/model path tested: Verboo/Codex planner gpt-5.6-sol, catalog-qualified worker IDs, and non-Claude-native alias fallback
• screenshots attached (if UI changed): not applicable; this changes runtime routing and diagnostics
• follow-up work or known limitations: the process must be restarted after rebuilding so an already-running CLI loads the new bundle; personal ~/.verboo skill/settings changes are intentionally not part of this repository PR

[ImPedro29]

Obrigado pela contribuição, @guustavomv. O #20 já foi mesclado e incorpora os pontos centrais deste PR: IDs canônicos com prefixo de plano, perfis semânticos fast/review/coding/testing/balanced, suporte a profile: em skills inline e forked, fallback seguro dos aliases Claude e diagnósticos de roteamento. A solução mesclada também mantém os papéis anunciados pelo router como autoritativos e acrescenta limites de execução, timeout com resultado parcial e garantia de pareamento tool_use/tool_result. Por isso, este PR ficou substituído pelo #20 e será encerrado. Valeu por levantar o problema e pela base das melhorias de catálogo e perfis.