qstack-choose-model

/qstack-choose-model

Which model should do this, and how hard should it think? Answer from what is reachable and what is measured, never from a name typed into a skill.

Invocation

/qstack-choose-model [task text] [--points N] [--files a,b] [--gate] [--live-target] [--orchestrator <model>]

With no task text, the task is the one settled in the conversation. Every flag is optional metadata; when a board card is the task, its points, files and refs are the metadata and the caller passes them.

Output, always the same shape, in the response and as JSON when a caller asked for it:

{"tier":"strong","model":"claude-opus-5","reasoning":"high",
 "why":"5-point card with an integration seam; strong = top coding index among reachable non-orchestrator models (Artificial Analysis, 2026-09-20)",
 "catalogue":"codex model_provider http://127.0.0.1:8317/v1","scores":"Artificial Analysis Data API, 2026-09-20"}

The script

Everything mechanical lives in scripts/qstack-models, next to this file. Resolve it relative to this SKILL.md's real path; never hardcode a home.

Command	Prints
discover	the reachable model ids and which source produced them
score	the catalogue joined with coding index and price
tier --orchestrator M	the strong / standard / cheap map, M excluded
resolve NAME	the id NAME means, or the closest three when it means nothing
probe-brief ID	the one-line task that proves ID can be launched
aliases	what Claude Code's fable/opus/sonnet/haiku resolve to here, from ANTHROPIC_DEFAULT_*_MODEL
codex-agents --orchestrator M [--reviewer R] [--dir D] [--dry-run]	writes the tier map as custom agent files qstack_strong, qstack_standard, qstack_cheap, qstack_reviewer under ~/.codex/agents/

Where the catalogue comes from

First hit wins, and the output names it:

1. QSTACK_MODELS_URL, an override nobody has to set.
2. The base URL the CLI already trusts: Codex's model_providers.<current>.base_url in ~/.codex/config.toml, then ANTHROPIC_BASE_URL, then OPENAI_BASE_URL, each asked GET /v1/models. A CLIProxyAPI wired into either CLI is found here with no setup.
3. The harness's built-in list: Claude Code's Agent tool enum, detected from CLAUDECODE in the environment or QSTACK_HARNESS=claude-code. Codex has no fixed list of its own, so from Codex this step is empty.
4. Nothing usable: ask the user for the models that may work, one question, free text, then resolve each answer.