Updating Model Lists
Each project keeps its own copy of the supported AI code-agent model lists in aitasks/metadata/models_<agent>.json. Refresh these files when vendors release new coding-capable models, rename existing ones, or deprecate old variants.
Why refresh
The local aitasks/metadata/models_<agent>.json files drive:
ait codeagent— model selection per skill/operation.- Settings TUI — the Agent Defaults and Models tabs.
- Verified scores — per-skill, per-model satisfaction ratings.
Refreshing periodically keeps these surfaces accurate as vendor offerings change.
One-shot refresh of all agents
The /aitask-refresh-code-models skill walks through all four agents in a single run:
- claudecode, codex, geminicli — discovered via web research (
WebSearch+WebFetchagainst vendor documentation URLs). - opencode — discovered via CLI (
bash .aitask-scripts/aitask_opencode_models.sh), because the available OpenCode models depend on which providers the user has connected locally.
The skill preserves existing verified scores and never auto-removes models — deprecated entries are flagged for explicit user approval. Run it from the project root:
/aitask-refresh-code-models
OpenCode-only quick path
When you only need to refresh the OpenCode list (e.g., after connecting a new provider locally), call the helper directly:
bash .aitask-scripts/aitask_opencode_models.sh # discover and update
bash .aitask-scripts/aitask_opencode_models.sh --dry-run # show diff only
bash .aitask-scripts/aitask_opencode_models.sh --sync-seed
The helper preserves verified scores and the unavailable status marker on models the local CLI cannot currently see.
Adding a single known model
When you already know a model’s cli_id (e.g., a vendor just announced a specific variant), skip web research and use /aitask-add-model:
/aitask-add-model --agent claudecode --name opus4_7_1m --cli-id 'claude-opus-4-7[1m]' --notes "1M context"
Add --promote --promote-ops <ops> to also set the new model as the default for the specified ops. See the skill page for the full flag list and the manual-review file list emitted after a promote.
Where the files live
| File | Purpose | Branch |
|---|---|---|
aitasks/metadata/models_<agent>.json | Runtime model list — read by ait codeagent, Settings TUI, and stats. | Task-data branch (committed via ./ait git). |
seed/models_<agent>.json | Template for new projects bootstrapped with ait setup. | Source-repo branch (committed via plain git). Only present in the framework source repo. |
<agent> is one of: claudecode, codex, geminicli, opencode.
Commit conventions
The two file locations live on different branches, so they need separate commits:
- Metadata (
aitasks/metadata/...) —./ait git add+./ait git commit. See the Git branching model for why task data lives on a separate branch. - Seed (
seed/...) — plaingit add+git commit. Seed files exist only in the framework source repo and never need./ait git.
Both /aitask-refresh-code-models and /aitask-add-model handle this split automatically; the convention matters only if you are editing the JSON files by hand.
Related
/aitask-refresh-code-models— full refresh skill reference./aitask-add-model— single-model registration.- Settings TUI — Models tab (read-only view of the lists).
- Verified scores — how ratings accumulate.
ait codeagent— model picker that consumes these files.