…e label (#3589) (#3854) * fix(chat): budget injected conversation history to the model context window (#3852) Chat injected the last 40 turns into every prompt (the #3636 always-inject contract) with no size budget, and never clamped individual turn text. A long chat, a huge pasted message, or a big tool result could push the assembled prompt past the model's context window and hard-fail with `413 ... prompt is too long: 206134 tokens > 200000 maximum` instead of degrading gracefully. Extract the (previously duplicated) history-block assembly into a single, unit-tested helper that budgets the injected history in characters against the preset's existing `maxContextChars` setting: it drops the oldest turns first and clamps any single oversized turn, keeping the most recent context. Both the queued path and the main send path now route through it. The #3636 contract is preserved (recent history is always injected when present); only its size is now bounded. Verified by 13 unit tests including a repro of the ~800k-char history that produced the 413. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(settings): label the clear-cache button "clear" not "scan" (#3589) The Clear Cache card, its description, and the confirm dialog all say "clear cache", but the button itself said "scan" — the lone inconsistency the issue reports. The button is the sole entry point to the clear flow (it scans to build the deletion preview, then the dialog confirms and deletes), so "clear" matches the standard pattern and the surrounding copy. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(chat): bound context in pi's compaction layer, not the frontend (#3852) Moves the #3852 fix out of the frontend and into pi's own context/compaction layer, where it can be token-accurate and where context management belongs. 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 every send (#3636); on a long chat that single message overflows and hard-fails with `413 prompt is too long`, and the overflow-retry cannot recover. The context-pruning extension (already installed in every chat + pipe session) now clamps any single oversized message to <=50% of the model's real ctx.model.contextWindow, inside its `context` hook (pi's transformContext, run before every LLM call; returning {messages} replaces what is sent). For the injected history block it drops the oldest turns and always preserves the user's actual message; generic huge payloads keep head + tail. This complements built-in compaction instead of duplicating it. Reverts the earlier frontend char-budget (standalone-chat.tsx back to baseline; deletes lib/chat-context-budget.ts and its test) so context management lives in one place. Tests: 22 cases drive the real shipped extension handlers (imported via a new @screenpipe-ext vitest alias, so no ported copy can drift) across edge cases — the ~800k-char #3852 repro bounded under the window with the question intact, small-window models, old-tool-result pruning, and tool_result feedback. The wdio e2e suite mocks pi (no subprocess) so it cannot exercise the extension; these handler tests are the coverage. next build compiles cleanly. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Louis Beaumont <louis@screenpi.pe> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

[bug] [cosmetic] Clear cache button says "Scan" #3589

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions