Proposal: differentiate as the local / small-context / unreliable-endpoint auto-compactor

[lkraider]

Summary

capyup and aashishd/pi-auto-compact-plugin currently overlap heavily, and aashishd has taken the "clean, native-only, heavily-tested, pinned-to-pi" lane. Rather than converge on feature parity, capyup could own the lane aashishd deliberately does not serve — which is exactly where #2 (Qwen3-122b / LM Studio) and #1 (flaky endpoint) live.

Thesis

Be the auto-compactor for small-window, local, and unreliable models, where pi's native overflow recovery is insufficient and a smarter emergency path is genuinely needed.

Concrete direction

1. Compaction that can't overflow itself. The core fix for Inconsistent and errors frequently #2 / Error: Compaction failed: Turn prefix summarization failed: Connection error. #1: hierarchical / chunked summarization so the summarization request is guaranteed to fit the window even when the live context has already blown past it. This is the biggest differentiator — native-only compactors assume the summarization call fits; on a 32–40K local window it often doesn't.

2. Provider-agnostic, calibrated token accounting. Don't trust a single chars/4 estimate. Prefer model-reported usage when present; fall back to a per-provider calibrated estimator. Small models and non-OpenAI endpoints are exactly where one estimate drifts and overflows.

3. Endpoint resilience. Retry / backoff + circuit-breaker on the summarization call (Error: Compaction failed: Turn prefix summarization failed: Connection error. #1 was a connection error). Add a last-resort lossless local truncation that keeps tool-call / tool-result pairs intact when the summarizer is unreachable, so the run survives.

4. Keep aggressive multi-point triggering (pre-turn / mid-turn / emergency) — it's the right instinct for tiny windows that overflow fast — but make the emergency path non-destructive-by-default and pair-safe.

5. Adopt deliverAs: "followUp" for the post-compaction resume and retire the setImmediate + isIdle() race heuristic.

Positioning

"aashishd for frontier / subscription models with large windows; capyup for local / self-hosted / small-window / unreliable setups." That's a defensible niche with an underserved user base (LM Studio / Ollama / vLLM) rather than a feature-parity race.