#35 Live model lists: offer each agent CLI's own models in the picker, since the curated ones have already gone stale

[Anonymous]

Originally created by: imshaikot

The agent picker's model select offers a curated list per agent, from models in src/lib/agents/catalog.ts, whose comment says "the CLIs have no list command". Four of them now do, and the curated lists have already gone stale. Measured on 25 Sep 2026:

• Antigravity: neither gemini-3-pro nor gemini-3-flash is among the 14 models agy models (1.2.11) lists, which are spelled like gemini-3.8-flash-high.
• Cursor CLI: 3 of the 4 curated ids (claude-opus-4-8, gpt-5, sonnet-4-thinking) are missing from the 241 that cursor-agent models prints.
• Claude Code: claude-opus-5 and claude-fable-5 are a version behind; its own aliases opus and fable resolve to claude-opus-5-5 and claude-fable-5-1.

Picking one of those gives a run that fails on an unknown model. And a model released after a Browsentic build can only be reached with browsentic agent model, never from the picker.

What each CLI can tell us

Agent	Source	Signed out
Codex	its own cache, ~/.codex/models_cache.json: models[].slug where visibility is list. No process needed	no file until Codex has run once
Antigravity	agy models: id<TAB>label lines, ~5 s over the network	exit 1, Please sign in to view available models
Grok Build	grok models: * id (default) and - id lines	exit 0, with You are not authenticated and a built-in list the account may not have
Cursor CLI	cursor-agent models: id - label lines under a heading; accepts bracket overrides such as claude-opus-4-8[effort=high]	exit 1, Authentication required
OpenCode	opencode models (provider/model)	not yet recorded
Claude Code, Vibe, Qwen Code	no command	—

What has to stay true

A live list is only worth having if every way it can fail leaves the picker no worse than today:

1. Never empty, never waiting. The popup gets whatever list is already known; reading one happens in the background. When a read fails, the last list that read cleanly stays, or the curated one when there is none.
2. Never an argument injection. Whatever a CLI prints goes back to a CLI as --model <id>. An id that starts with a dash, or holds whitespace or control characters, must never reach argv, from a listing, from the panel, or from a hand-edited config.json. Cursor's bracketed ids must still pass.
3. Signed out is no list. Grok's exit-0 output in particular must not be taken as the account's models.
4. Bounded. A listing that hangs, floods its output or opens a login prompt cannot hold up the daemon, or leave a process behind.
5. What the user pinned is theirs. A pinned model the CLI no longer lists stays selected, and is shown as such.
6. The curated lists are the floor, so they have to be right on their own. Claude Code's should be its aliases, which cannot go stale.

Done means

• An optional lister on the Runner contract (a command or a file, plus a parser), implemented for Codex, Antigravity, Grok Build and Cursor CLI.
• The daemon keeps the lists in a cache it re-reads on a timer, backs off after a failure, and pushes agentInfo again when one changes. RunnerStatus carries the list and where it came from, which bumps the socket protocol.
• The picker offers the live list, with a filter once there are too many for a dropdown, and says where the list came from or why the read failed. Recheck reads it again.
• browsentic agent models <name> [--refresh].
• Parsers pinned by fixtures recorded from the real CLIs, signed in and signed out, and a test that every curated id is one its CLI listed.
• Docs: the agents guide, the configuration reference, troubleshooting, and state on disk.

Not in scope: the macOS app's picker, which keeps building from catalog.models (it decodes the new field and ignores it); and an OpenCode lister, until a real opencode models run has been recorded.

[Anonymous]

Ticket changed by: imshaikot

• status: open --> closed