bluemsi/connectai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: version up to 2.80.39 and package with UI/config polish

Browse Source
This commit is contained in:
g1nation
2026-05-13 00:39:28 +09:00
parent eb36cec050
commit d2269297ea
13 changed files with 88 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/agent.ts
+17 -8
View File
@@ -551,14 +551,16 @@ export class AgentExecutor {
// (2) 대화 기록을 남은 예산에 맞게 압축하고 (UI 표시용 chatHistory 는 건드리지 않음)
// (3) 동적으로 출력 상한(maxOutputTokens)을 계산한다.
// ──────────────────────────────────────────────────────────────────
// Small models (≤4B) routinely fail on prompts that fit their *nominal* context but
// exceed their *effective* capability (server log shows truncated=0 yet eval time≈0ms —
// Genuinely tiny models (≤3B) sometimes fail on prompts that fit their *nominal* context
// but exceed their *effective* capability (server log shows truncated=0 yet eval time≈0ms —
// the model emitted EOS as the first token). When detected, budget against a smaller
// effective window so the system prompt / RAG / history get shrunk proactively.
// Note: 4B+ models (e.g. gemma-4-e4b with a 100k+ window) are competent — DON'T cap them,
// or the output budget gets squeezed to the minimum and answers come out truncated.
const modelParamB = estimateModelParamsB(actualModel);
const smallModelCap = config.smallModelContextCap; // 0 disables this guard
const cappedForSmallModel = smallModelCap > 0
&& modelParamB !== null && modelParamB <= 4
&& modelParamB !== null && modelParamB <= 3
&& config.contextLength > smallModelCap;
const effectiveContextLength = cappedForSmallModel ? smallModelCap : config.contextLength;
if (cappedForSmallModel) {
@@ -664,7 +666,7 @@ export class AgentExecutor {
brainFiles: brainFiles.length,
imageCount,
tight: outputBudget.tight,
smallModel: cappedForSmallModel || (modelParamB !== null && modelParamB <= 3 && inputTokens > 8000),
smallModel: cappedForSmallModel || (modelParamB !== null && modelParamB <= 3 && inputTokens > 12000),
},
});
// If the user's message reads like a regression complaint ("또 안 돼", "비슷한 실수", "왜 반복돼"…),
@@ -913,8 +915,9 @@ export class AgentExecutor {
if (config.autoContinueOnOutputLimit && config.maxAutoContinuations > 0 && loopDepth === 0) {
const originalUserPrompt = prompt || (this.chatHistory.find(m => m.role === 'user' && typeof m.content === 'string')?.content as string) || '';
let lastOutputTokens = estimateTokens(cleaned.visible);
let lastMaxOutputTokens = maxOutputTokens; // budget the last round actually had (≠ first gen's after round 1)
while (
shouldAutoContinue(classifyStopReason(finishStopReason), cleaned.visible, lastOutputTokens, maxOutputTokens)
shouldAutoContinue(classifyStopReason(finishStopReason), cleaned.visible, lastOutputTokens, lastMaxOutputTokens)
&& continuationCount < config.maxAutoContinuations
&& !this.abortController?.signal.aborted
&& !this.isStaleRun(runId)
@@ -926,10 +929,10 @@ export class AgentExecutor {
{ role: 'system', content: CONTINUATION_SYSTEM_PROMPT, internal: true },
{ role: 'user', content: buildContinuationUserPrompt(originalUserPrompt, cleaned.visible) },
];
const contMax = computeOutputBudget(estimateMessagesTokens(contMsgs), ctxLimits).maxOutputTokens;
lastMaxOutputTokens = computeOutputBudget(estimateMessagesTokens(contMsgs), ctxLimits).maxOutputTokens;
const cr = await this.callNonStreaming({
baseUrl: ollamaUrl, modelName: actualModel, engine, messages: contMsgs,
temperature, maxTokens: contMax, contextLength: ctxLimits.contextLength,
temperature, maxTokens: lastMaxOutputTokens, contextLength: ctxLimits.contextLength,
signal: this.abortController?.signal,
});
finishStopReason = cr.stopReason;
@@ -938,9 +941,15 @@ export class AgentExecutor {
logInfo('Continuation produced no visible text — stopping.', { model: actualModel, round: continuationCount });
break;
}
const before = cleaned.visible;
cleaned = { ...cleaned, visible: mergeContinuationParts(cleaned.visible, ccl.visible), wasThoughtOnly: false };
lastOutputTokens = estimateTokens(ccl.visible);
logInfo('Auto-continued the answer.', { model: actualModel, round: continuationCount, addedChars: ccl.visible.length, totalChars: cleaned.visible.length, contStopReason: cr.stopReason });
logInfo('Auto-continued the answer.', { model: actualModel, round: continuationCount, addedChars: ccl.visible.length, totalChars: cleaned.visible.length, contStopReason: cr.stopReason, contMaxTokens: lastMaxOutputTokens });
// Guard against a continuation that adds (almost) nothing new after dedup — stop instead of spinning.
if (cleaned.visible.length - before.length < 20) {
logInfo('Continuation added negligible new text — stopping.', { model: actualModel, round: continuationCount });
break;
}
} catch (e: any) {
logError('Auto-continuation failed.', { model: actualModel, round: continuationCount, error: e?.message ?? String(e) });
break;
src/config.ts
+2 -2
View File
@@ -122,9 +122,9 @@ export function getConfig(): IAgentConfig {
return v === 'truncateMiddle' || v === 'rollingWindow' ? v : 'stopAtLimit';
})(),
autoCompactHistory: cfg.get<boolean>('autoCompactHistory', true),
smallModelContextCap: Math.max(0, cfg.get<number>('smallModelContextCap', 8192)),
smallModelContextCap: Math.max(0, cfg.get<number>('smallModelContextCap', 16384)),
autoContinueOnOutputLimit: cfg.get<boolean>('autoContinueOnOutputLimit', true),
maxAutoContinuations: Math.max(0, Math.min(10, cfg.get<number>('maxAutoContinuations', 3))),
maxAutoContinuations: Math.max(0, Math.min(10, cfg.get<number>('maxAutoContinuations', 4))),
finalOnlyRetryOnThoughtLeak: cfg.get<boolean>('finalOnlyRetryOnThoughtLeak', true)
};
}