UNPKG

@bitkyc08/opencodex

Version:

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code

72 lines (67 loc) • 3.89 kB
/** * Gemini progress-text history trimming. * * Gemini models, observed on Gemini 4 Argon over Cloud Code Assist, sometimes skip their thinking * channel on a tool-calling turn and plan in the visible text instead. The provider sends that * planning as an ordinary text part without a thought marker, for example a progress line in the * user's language followed by "Let's read google.ts 580-1700 first." Once one turn like that is * replayed as history, the model copies the pattern on the turns that follow, so the leak grows * through the rest of the session. Replaying the same 300K-token request with the leaked tails * removed from the history stopped the leak, while stronger system-prompt wording did not. * * The fix is on the history side only. Text that precedes the tool call of a replayed model turn is * intermediate progress, and the Google brevity instruction already asks for a single short line * there, so it is replayed as its lead sentence. A turn Codex marks `final_answer` and text after a * tool call are replayed whole. Visible output is never rewritten, final * answers (turns without tool calls) are replayed in full, and the split is punctuation-based so * it works the same in every language without matching any wording. */ /** * Sentence end: ASCII terminal punctuation followed by whitespace or the end (so "google.ts" and * "3.5" are not boundaries), a script's own sentence terminator (CJK 。!?, Devanagari * ।॥, Arabic ۔؟, Armenian ։, Ethiopic ።, Myanmar ။), or a line break. Closing quotes, brackets and * Markdown emphasis right after the punctuation belong to the sentence. */ const SENTENCE_END = /[.!?]["'\u201D\u2019)\]*_`]*(?=\s|$)|[\u3002\uFF01\uFF1F\u0964\u0965\u06D4\u061F\u0589\u1362\u104B]["'\u201D\u2019\u300D\u300F)\]*_`]*|\n/g; /** A boundary that only closes a leading list marker ("1.", "a.") or a common abbreviation ("e.g."). */ const NOT_A_SENTENCE = /^(?:\d{1,3}|[a-z])\.$|(?:^|\s)(?:e\.g|i\.e|etc|vs)\.$/i; const WHITESPACE = /\s/; /** Whether replayed progress text for this model is trimmed to its lead sentence. */ export function trimsGeminiProgressHistory(modelId: string): boolean { return /gemini/i.test(modelId); } /** * The lead sentence of a replayed progress text, or undefined when nothing usable remains. * Text without a sentence boundary is already a single sentence and is returned whole. */ export function geminiProgressLead(text: string): string | undefined { const trimmed = text.trim(); if (trimmed.length === 0) return undefined; SENTENCE_END.lastIndex = 0; let end = trimmed.length; let checked = 0; let contentEnd = 0; for (let match = SENTENCE_END.exec(trimmed); match; match = SENTENCE_END.exec(trimmed)) { const boundary = match.index + match[0].length; // Only a newline boundary can end in whitespace. Scan its new interval once; repeated // blank lines retain the previous content end without rescanning the preceding text. if (match[0] === "\n") { for (let index = checked; index < boundary; index++) { if (!WHITESPACE.test(trimmed[index])) contentEnd = index + 1; } } else { contentEnd = boundary; } checked = boundary; // The longest exception is four characters plus its preceding whitespace. The list // marker alternative is anchored to the whole prefix, so allow it only at the start. const start = Math.max(0, contentEnd - 5); const suffix = trimmed.slice(start, contentEnd); const exception = NOT_A_SENTENCE.exec(suffix); if (!exception || (exception.index === 0 && start > 0 && !/^\s/.test(suffix))) { end = boundary; break; } } const lead = trimmed.slice(0, end).trim(); // A code-fence opener is not a sentence; replaying it alone would leave an unclosed fence. if (lead.length === 0 || lead.startsWith("```")) return undefined; return lead; }