@tinysquid/pi-auto-session-name

A cheap model turns your first prompt into a short, searchable name for every pi coding agent session

After the first turn, this extension takes your opening prompt, asks a small, configurable model to summarize it into a 3–8 word title, and sets it as the session name.

Without it, pi's /resume picker falls back to showing each session's raw first prompt — fine for a one-liner, useless when the opening message is a pasted stack trace, a long spec, or a skill invocation, since you only see a truncated fragment:

TypeError: Cannot read properties of undefi…   42 3h
kitchen-sink Review the draft README and tell…  7 2d

With it, every session gets a short, searchable title instead:

Fix WebSocket reconnect race                   42 3h
Debug flaky Vitest snapshot                     7 2d

How it works

┌─────────────────┐  first prompt   ┌────────────────┐
│  user's opening │ ──────────────▶ │  naming model  │
│  prompt         │                 │  (configurable)│
└─────────────────┘                 └───────┬────────┘
                                            │  3–8 word title
                    ┌───────────────────────▼────────┐
                    │ pi.setSessionName("Fix         │
                    │   WebSocket reconnect race")   │
                    └────────────────────────────────┘

• When it fires — in the interactive TUI only, once per session, on agent_end after the first turn, and only if the session has no name yet. Skill invocations at session start are ignored; /reload re-runs the extension but the existing-name guard prevents a second naming run.
• When it doesn't fire — one-shot modes (-p / --print, --mode json, --mode rpc, and implicitly piped runs), --no-session, and any session that already has a name. That last guard is what makes explicit --name <x> respected: the user's name is never replaced.
• Why interactive only — one-shot runs are scripted; naming burns a model call per invocation and the sessions are usually throwaway. Named sessions only matter for the /resume picker, which people use interactively.
• Why a name — named sessions are searchable in the /resume picker (and survive Ctrl+N's named-only filter), and the terminal title becomes pi - <session name> - <cwd> instead of just pi - <cwd>.
• Model pick — uses the model pinned in the config file. Without a config, it picks the cheapest available model by input token cost (respecting session model scoping via enabledModels / --models), falling back to the active session model.
• Non-blocking — naming runs in the background so the prompt is never delayed (it never delays anything: it only runs in the interactive TUI).
• One name per fresh session — /new and /fork inside a TUI run reset the guard, so each fresh session gets its own naming run from its own first prompt.

Install

pi install npm:@tinysquid/pi-auto-session-name

For local development:

ln -s $(pwd)/auto-session-name/auto-session-name.ts ~/.pi/agent/extensions/auto-session-name.ts

That's it — start a session, and after the first agent turn it gets a name. No commands, no flags.

Configuration

Config file ~/.pi/agent/auto-session-name.json pins the naming model and request options:

{
  "provider": "google",
  "model": "gemma-4-26b-a4b-it",
  "temperature": 0.2,
  "thinking": "off"
}

• provider, model — pin the naming model (required keys).
• temperature — sampling temperature passed to the naming request as-is. Must be between 0 and 2; values outside that range (or non-numbers) are treated as malformed and repaired away. Omitted by default (provider default applies). Providers may drop it where they forbid it (Anthropic ignores it when thinking is enabled, and only accepts 0–1).
• thinking — reasoning level for the naming request: "off" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max". Defaults to "off". The level is clamped to what the model supports (model-specific level mappings are respected).

The file is read at naming time, so edits apply on the next session — no reload required. When a file is generated or repaired (see below), the rewrite is schema-normalizing: only provider, model, temperature, and thinking are kept, so any extra keys are dropped.

Self-healing config

The config file repairs itself, with three levels of intervention:

Missing keys are simply left unset (defaults apply).

[!TIP] The generated file is a starting point for your edits — pin a tiny, fast model like Gemma or a mini-tier model to keep naming cheap and instant.

[!NOTE] A configured model that is unknown or has no configured credentials falls back to the default pick (with a warning).

Behavior notes

• Quitting or switching sessions within the ~1–2s naming window loses the name (cosmetic; a warning is logged).
• Everything the config file touches happens inside a naming run, so file generation and self-healing only occur in interactive sessions — a purely scripted (-p) usage never creates the config file.
• Failures and config problems are logged to the terminal ([auto-session-name] ...), never surfaced in the UI.
• The naming call passes the session id, which opencode gateways require: without it they reject the call with 400 MissingSessionID and the session silently stays unnamed (pi-ai derives x-opencode-session from it since 0.87.1).
• Titles are sanitized: surrounding quotes added by chatty models are stripped, and an empty response leaves the session unnamed rather than setting a garbage name.