Skip to content

armgit / avibe

Public Imported from GitHub
0

Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.

4.2 KB
Raw Blame

vibe agent models — query available models + reasoning efforts

Background

vibe agent create / update accept --model and --reasoning-effort as free-form strings with no choices, no validation, and no discovery path. An agent (or user) calling the CLI to configure another Agent has no way to learn which models or reasoning-effort levels a backend actually supports — it can only guess. The knowledge already exists in the codebase (Claude catalog, Codex built-in list + caches, OpenCode live providers incl. PR #461 custom providers / user models) but nothing surfaces it through the CLI.

Goal

Add vibe agent models so an agent can discover, for a given Agent (or backend), the available models and the reasoning-effort levels valid for each — including OpenCode custom providers and user-added models — and feed that back into create / update.

Design (approved)

Command, Agent-centric to match vibe agent show / update / remove:

vibe agent models <name>          # an Agent: resolve its backend + current model
  --backend {claude,codex,opencode}   # query a backend directly (pre-creation)
  --provider <id>                     # OpenCode-only filter (error on other backends)
  --model <id>                        # narrow to one model's reasoning efforts
  --json                              # no-op; JSON is the default envelope

Exactly one of <name> or --backend. Output kind: "agent_models":

{ ok, kind: "agent_models", agent, backend,
  current: { model, reasoning_effort, model_known, reasoning_effort_valid, valid } | null,
  default_model,
  providers: [ { id, name, custom } ],          # OpenCode only
  models:    [ { value, default?, provider?, source?, reasoning_efforts: [...] } ],
  source, live, notes }

Reasoning efforts are nested per model because they are a property of the model (Claude xhigh/max depend on the model; OpenCode varies by provider variant).

Architecture — reuse, don't rebuild

  • New backend-agnostic api.agent_model_options(backend, *, model, provider, cwd) wraps the three existing producers into one shape:
    • claude_models() → models + per-model reasoning options (already returns both)
    • codex_models() → extend to also return reasoning_options (build_codex_reasoning_options() already exists; just surface it)
    • opencode_options(cwd) → already overlays custom providers + user models (PR #461: _read_opencode_custom_provider_ids + _merge_opencode_user_models), so reusing it covers them for free. Annotate source: user|catalog (opencode_config._is_vibe_user_model) and custom providers (read_opencode_custom_providers). --provider filters via the existing allowed_providers concept. cwd resolved internally (runtime.default_cwd), never a CLI flag.
  • The same resolver can later back the existing /api/{claude,codex}/models endpoints so the Claude reasoning rules stop being duplicated in TS (RoutingConfigPanel.tsx). Out of scope for this PR (follow-up).

Validation policy — warn, don't reject (decision #1)

create / update keep accepting free-form values (catalogs are best-effort and can lag a new model). When a set value is not in the known set, attach a non-fatal warnings + hint pointing at vibe agent models <name>. Uses only the cheap file-based sources (claude/codex); OpenCode is skipped (its list is live) to keep create/update fast.

Scope (decision #2)

In: the query command, the resolver, codex_models() reasoning, the create/update warning. Out (follow-up PR): consolidating the duplicated TS reasoning rules onto the Python source.

Todo

  • api.agent_model_options + codex_models() reasoning
  • vibe agent models parser + cmd_agent_models
  • create/update warn-not-reject hint
  • tests (resolver, CLI, codex reasoning, warning)
  • docs + ruff + targeted pytest
  • reviewer subagent → PR (non-draft) → Codex review watch

Evidence layers

  • unit: codex reasoning, resolver per backend, warning helper
  • contract: cmd_agent_models name / --backend / --provider error envelope
  • scenario: n/a (no catalog entry for CLI agent mgmt yet)
  • manual: vibe agent models against claude/codex/opencode in an isolated home