pi-no-spin

A pi extension that detects when the LLM spins — repeating the same string over and over — interrupts the generation, and reminds the model to stop — so you don't burn tokens while the model churns out the same segment on repeat.

Why it exists: open-source / self-hosted models (DeepSeek, Qwen, Llama, GLM, …) have a non-negligible chance of falling into output loops — repeating the same sentence, code block, or token sequence continuously without making progress (deepseek-reasoner and quantized local models are especially prone). pi-no-spin cuts these off in real time.

How it works

1. Listen — the extension hooks into pi's message_update event, which fires token-by-token while the assistant streams a response.
2. Detect — it accumulates the text and runs a periodicity check on the tail: if the last period × threshold characters satisfy text[i] === text[i + period] for all i (a periodic run), a repeated segment of length period has been output at least threshold times in a row. Period is scanned from minUnit (default 4) up to maxUnit (default 300) characters.
   • Period detection is used instead of block-aligned matching, so it still fires while the model is mid-way through streaming the next repeat (a partial segment) — no false negatives due to alignment jitter.
   • Pure-whitespace repetition (indentation / blank lines) is not treated as a loop.
   • Repetition of separators (─ and .) is not treated as a loop either — those are visual dividers / dot leaders, e.g. a long ───── rule or a ..... line, not spinning.
3. Interrupt — on detection the extension calls ctx.abort() to stop generation immediately, shows a ⛔ notification, then queues a follow-up user message telling the model it is stuck in a loop and should re-assess instead of repeating.

Install

pi install git:github.com/wilhg/pi-no-spin

or with SSH shorthand:

pi install git:git@github.com:wilhg/pi-no-spin.git

Then reload in a running pi session:

/reload

Try without installing

pi -e git:github.com/wilhg/pi-no-spin

Usage

Detection is on by default. Configure it with the /nospin command:

/nospin                Show status and current configuration
/nospin on|off         Enable / disable detection
/nospin threshold N    Reps of the same segment that count as a loop (default 7)
/nospin min N          Minimum segment length to consider (default 4)
/nospin max N          Maximum segment length to consider (default 300)

Examples:

/nospin                    →  pi-no-spin: 🟢 enabled | threshold=7 | minUnit=4 | maxUnit=300 | cooldown=10000ms
/nospin threshold 15       →  require 15 consecutive repeats
/nospin max 500            →  detect segments up to 500 chars
/nospin off                →  disable

Configuration is persisted per-session (pi.appendEntry) and restored automatically on the next session_start.

Quick verification

Tell the model something like:

Please repeat the same sentence 20 times in a row.

You should see it interrupted around the 7th repetition — a ⛔ notification appears and the model receives a reminder to stop repeating.

Options & defaults

Option	Default	Meaning
threshold	7	Consecutive reps that count as a loop
minUnit	4	Minimum segment length (chars)
maxUnit	300	Maximum segment length (chars)
cooldownMs	10000	Cooldown between triggers (ms)
enabled	true	Detection on by default

Known limitations

• Non-streaming whole-message arrival: if an entire response arrives at once (non-streaming provider) and the tail of that message happens to contain a few non-repeating closing characters, the tail-anchored periodic window can miss the run. In practice this never matters: with streaming, detection fires at the exact moment the 7th repetition completes — before the model gets a chance to emit any closing text.
• Exact periodicity: the check requires characters at distance period to be exactly equal. A model that repeats a segment with slight mutations each time will not trigger (returning the same string byte-for-byte is the common failure mode this targets).
• Tool-call loops: repeated identical tool calls are a different loop signature and are not covered (yet).

Development

npm test     # runs tests/pi-no-spin.test.mts (Node 22.6+ / 23+, type stripping)

npm test does not require pi or a model — the detection algorithm is pure and unit-tested.

License

MIT © wilhg