@benvargas/pi-model-sort

Sorts pi's model picker by last usage and starts fresh sessions on your most recently used model.

• /model picker — both "Scope: all" and "Scope: scoped" views, including fuzzy-search results — is sorted by recency: current model first → most recently used → provider/id alphabetical
• Ctrl+P / Ctrl+Shift+P scoped cycling follows last-used order (scoped models come from enabledModels or --models)
• Fresh starts and /new begin on your most recently used model instead of enabledModels[0] or the hardcoded provider default
• Continued sessions (pi -c, --session, /resume, forks) keep the model saved in the session file, and that restored model is recorded as last-used (fresh /new sessions and /reload are excluded from that recording)
• Remembers the thinking level you last used on each model and restores it on every switch, clamped to what each model supports
• No configuration needed — tracking starts on first use and degrades to the default alphabetical order with no history

Forked from monotykamary/pi-model-sort (MIT, v0.3.2). See THIRD_PARTY_NOTICES.md for the full list of fork changes and attribution.

Install

pi install npm:@benvargas/pi-model-sort

Or try without installing:

pi -e npm:@benvargas/pi-model-sort

Usage

The extension works automatically — there are no commands to learn.

/model                    # Most recently used models appear at the top
Ctrl+P / Ctrl+Shift+P     # Cycle through scoped models in last-used order
pi                        # Fresh starts use MRU
pi -c                     # Continuations keep the session's model

How It Works

• Tracking uses pi's documented extension events: /model switches (model_select) and thinking-level changes (thinking_level_select) are timestamped into ~/.pi/agent/extensions/pi-model-sort.json. Continued sessions restore their model during construction without emitting model_select (pi 0.84.3), so the extension records the restored model at session_start itself.
• Sorting has no SDK hook, so the extension wraps (monkey-patches) internal methods: ModelSelectorComponent.sortModels, its scoped loader and filterModels, and AgentSession._cycleScopedModel for scoped cycling. All original methods are preserved and restored on shutdown/reload; the patches survive modelRegistry.refresh().
• The MRU startup override calls pi.setModel() on session_start for fresh starts and /new only. A continued session is detected by projecting its branch through pi's own context-message rules (message, custom_message, non-empty branch_summary, and compaction entries — exactly what buildSessionContext() counts) — pi seeds every new session with model_change + thinking_level_change entries before session_start fires, so raw branch length cannot distinguish fresh from continued.

Known limitations on pi 0.84.x

• --list-models output is not sorted: the CLI lists models before extensions load and re-sorts by provider/id internally.
• /scoped-models and unscoped Ctrl+P cycling are not re-sorted: they read ModelRuntime snapshots directly, not the extension-facing ModelRegistry facade this extension wraps. (Scoped cycling — the common case with enabledModels set — is sorted via the _cycleScopedModel wrapper.)

Configuration

Usage history lives in ~/.pi/agent/extensions/pi-model-sort.json:

{
  "lastUsed": {
    "provider/modelId": 1717000000000
  },
  "thinking": {
    "provider/modelId": "high"
  }
}

No manual editing is needed. To clear usage history, delete the file and /reload.

Notes

• Ctrl+P cycling does not update last-used timestamps — doing so would create a sort feedback loop (each cycle step re-sorts the selected model to the top, making the cycle toggle forever between the top two). Manual selections and session restores still update it.
• Patches are coupled to pi internals; this package is maintained against current pi releases (peer requirement >=0.84.0, the version it is tested with).

Uninstall

pi remove npm:@benvargas/pi-model-sort