What

Two fixes from the user-feedback triage.

1. [bug] Chat fails with 413 "prompt is too long" (>200k tokens) instead of trimming context to the model window #3852 chat hard-fails with 413 prompt is too long on long conversations. Now fixed in pi's own context/compaction layer.
2. [bug] [cosmetic] Clear cache button says "Scan" #3589 the clear-cache button said "scan" instead of "clear".

1. Bound oversized messages in pi's compaction layer (closes #3852)

Root cause. pi's built-in compaction summarizes across messages but cuts at message boundaries, so it can never shrink a single message that is itself larger than the context window. The chat re-injects recent history as one big <conversation_history> user message on every send (the #3636 contract). On a long chat (or with a huge pasted/tool turn) that single message exceeds the window, the request 413s, and pi's overflow-retry cannot recover because no amount of summarizing other turns shrinks that one message.

Fix. The context-pruning.ts extension already runs in every chat and pipe session. Its context hook is pi's transformContext slot: it runs before every LLM call and the array it returns is what actually gets sent. The hook now clamps any single message whose text exceeds 50% of the model's real ctx.model.contextWindow. For the injected history block it drops the oldest turns and always preserves the user's actual message; other oversized payloads keep head and tail. This complements built-in compaction (which then handles cross-message totals) rather than duplicating it.

flowchart LR
  A["chat injects last 40 turns as one user message (#3636)"] --> C["pi context hook, before every LLM call"]
  C --> D{"any single message > 50% of window?"}
  D -->|"before #3852"| E["built-in compaction can't split one message"] --> F["413 prompt too long"]
  D -->|"after"| G["clamp it: keep recent turns + the question"] --> H["compaction handles the rest, no 413"]

This replaces an earlier attempt that bounded the block on the frontend with a character budget. Per review that was the wrong layer (it bypassed pi's compaction and was a blunt char clamp), so it is reverted: standalone-chat.tsx is back to baseline and the chat-context-budget.ts helper is removed.

2. Clear-cache button label (closes #3589)

The card title, description, and confirm dialog all say "clear cache"; the button said "scan". Changed it to "clear".

Verification

• bunx vitest run (22 tests, all green) drive the real shipped extension handlers, imported via a new @screenpipe-ext vitest alias so there is no ported copy to drift: the pure helpers (including a throwing getContextUsage), the context handler end to end (empty/short no-op, the ~800k-char [bug] Chat fails with 413 "prompt is too long" (>200k tokens) instead of trimming context to the model window #3852 repro bounded under the window with the question preserved, small-window models, old-tool-result pruning still intact, combined cases), and the tool_result handler.
• next build compiles and type-checks clean (the one warning is a pre-existing unpdf dependency note, unrelated).
• Honest limitation. The wdio e2e suite mocks pi (no real subprocess), so it cannot exercise this extension. The handler-level tests above are the coverage. e2e still covers the frontend <conversation_history> injection contract (Bug: Chat Suddenly Loses Conversation Context #3636), which the revert leaves unchanged.

Scope

Other live issues from the triage (transcription backlog / proxy resets #3850, model 404s #3786, timezone scheduling #3851) need staging or manual verification and are tracked separately.

🤖 Generated with Claude Code