feat(canvas): 云端 Agent - 缩减画布目录上下文并展示实时用量
This commit is contained in:
ddcat666
188
web/src/lib/canvas/agent-context-usage.ts
Normal file
188
web/src/lib/canvas/agent-context-usage.ts
Normal file
@@ -0,0 +1,188 @@
|
||||
export type AgentContextUsage = {
|
||||
runId: string;
|
||||
reading: Record<string, unknown> | null;
|
||||
readingSeq: number;
|
||||
readingStale: boolean;
|
||||
compactionPending: Record<string, unknown> | null;
|
||||
lastCompaction: Record<string, unknown> | null;
|
||||
};
|
||||
|
||||
export type AgentContextUsageEvent = {
|
||||
runId: string;
|
||||
seq?: number;
|
||||
type: string;
|
||||
payload?: Record<string, unknown>;
|
||||
};
|
||||
|
||||
/** What the ring is telling the user. Compaction, not the raw window, is the decision. */
|
||||
export type AgentContextPhase = "idle" | "unknown" | "ok" | "watch" | "compress" | "compacting" | "stale";
|
||||
|
||||
export type AgentContextBreakdownItem = {
|
||||
key: string;
|
||||
label: string;
|
||||
tokens: number;
|
||||
bytes?: number;
|
||||
unit: "tokens" | "bytes";
|
||||
};
|
||||
|
||||
export type AgentContextUsageView = {
|
||||
phase: AgentContextPhase;
|
||||
/** Share of the input budget, 0–1 when the model window is known. */
|
||||
ratio?: number;
|
||||
/** Where semantic compaction starts, as a share of the same input budget. */
|
||||
compactRatio?: number;
|
||||
/** Ring fill. 1 means "at or past the compaction line", not "window is full". */
|
||||
ring: number;
|
||||
label: string;
|
||||
detail: string;
|
||||
inputTokens?: number;
|
||||
remainingTokens?: number;
|
||||
usableTokens?: number;
|
||||
compactAtTokens?: number;
|
||||
contextWindowTokens?: number;
|
||||
protocolBytes?: number;
|
||||
tokenSource?: "provider" | "estimate";
|
||||
estimate: boolean;
|
||||
breakdown: AgentContextBreakdownItem[];
|
||||
lastCompaction: Record<string, unknown> | null;
|
||||
};
|
||||
|
||||
export function emptyAgentContextUsage(runId: string): AgentContextUsage {
|
||||
return { runId, reading: null, readingSeq: 0, readingStale: false, compactionPending: null, lastCompaction: null };
|
||||
}
|
||||
|
||||
/** Reduces durable Agent events without mixing readings from different runs. */
|
||||
export function reduceAgentContextUsage(current: AgentContextUsage, event: AgentContextUsageEvent): AgentContextUsage {
|
||||
const scoped = current.runId === event.runId ? current : emptyAgentContextUsage(event.runId);
|
||||
const payload = event.payload && typeof event.payload === "object" ? event.payload : {};
|
||||
if (event.type === "context_pressure") {
|
||||
return { ...scoped, reading: payload, readingSeq: event.seq || 0, readingStale: false, compactionPending: null };
|
||||
}
|
||||
if (event.type === "context_compaction_requested") {
|
||||
return { ...scoped, compactionPending: payload };
|
||||
}
|
||||
if (event.type === "context_compacted") {
|
||||
return { ...scoped, readingStale: Boolean(scoped.reading), compactionPending: null, lastCompaction: payload };
|
||||
}
|
||||
return scoped;
|
||||
}
|
||||
|
||||
export function finiteContextNumber(value: unknown): number | undefined {
|
||||
return typeof value === "number" && Number.isFinite(value) && value >= 0 ? value : undefined;
|
||||
}
|
||||
|
||||
/** Returns a ratio only when the backend has a trustworthy configured model budget. */
|
||||
export function contextPressureRatio(reading: Record<string, unknown> | null): number | undefined {
|
||||
if (!reading || reading.modelLimitConfigured !== true) return undefined;
|
||||
const usable = finiteContextNumber(reading.usableInputTokens);
|
||||
if (!usable || usable <= 0) return undefined;
|
||||
const ratio = reading.tokenSource === "provider"
|
||||
? finiteContextNumber(reading.projectedPressureRatio)
|
||||
: finiteContextNumber(reading.pressureRatio);
|
||||
return ratio;
|
||||
}
|
||||
|
||||
export function contextInputTokens(reading: Record<string, unknown> | null): number | undefined {
|
||||
if (!reading) return undefined;
|
||||
if (reading.tokenSource === "provider") {
|
||||
return finiteContextNumber(reading.projectedNextInputTokens) ?? finiteContextNumber(reading.estimatedInputTokens);
|
||||
}
|
||||
return finiteContextNumber(reading.estimatedInputTokens);
|
||||
}
|
||||
|
||||
const BREAKDOWN_LABELS: Record<string, string> = {
|
||||
system: "系统提示（含画布摘要）",
|
||||
tools: "工具 schema",
|
||||
messages: "会话消息（含工具结果）",
|
||||
};
|
||||
|
||||
function contextBreakdown(reading: Record<string, unknown> | null): AgentContextBreakdownItem[] {
|
||||
const breakdown = reading?.breakdown;
|
||||
if (!breakdown || typeof breakdown !== "object") return [];
|
||||
const buckets = (breakdown as { buckets?: unknown }).buckets;
|
||||
if (!Array.isArray(buckets)) return [];
|
||||
return buckets.flatMap((bucket) => {
|
||||
if (!bucket || typeof bucket !== "object") return [];
|
||||
const item = bucket as { key?: unknown; label?: unknown; scaledTokens?: unknown; tokens?: unknown; bytes?: unknown };
|
||||
const key = typeof item.key === "string" ? item.key : "";
|
||||
const tokens = finiteContextNumber(item.scaledTokens) ?? finiteContextNumber(item.tokens);
|
||||
if (!key || tokens === undefined) return [];
|
||||
const label = BREAKDOWN_LABELS[key] || (typeof item.label === "string" ? item.label : key);
|
||||
const bytes = finiteContextNumber(item.bytes);
|
||||
return [{ key, label, tokens, bytes, unit: "tokens" as const }];
|
||||
});
|
||||
}
|
||||
|
||||
|
||||
function contextProtocolBytes(reading: Record<string, unknown> | null): number | undefined {
|
||||
if (!reading?.breakdown || typeof reading.breakdown !== "object") return undefined;
|
||||
return finiteContextNumber((reading.breakdown as { envelopeBytes?: unknown }).envelopeBytes);
|
||||
}
|
||||
|
||||
function formatContextTokens(tokens: number | undefined): string {
|
||||
if (tokens === undefined) return "—";
|
||||
if (tokens >= 1_000_000) return `${Math.round(tokens / 100_000) / 10}M`;
|
||||
if (tokens >= 1_000) return `${Math.round(tokens / 100) / 10}K`;
|
||||
return Math.round(tokens).toLocaleString("zh-CN");
|
||||
}
|
||||
|
||||
/**
|
||||
* One view of the compaction mechanism.
|
||||
* The ring fills against compactAtTokens (85% of the input budget): full means
|
||||
* the next model call is about to pause and compact, not that the window is 100% used.
|
||||
*/
|
||||
export function presentAgentContextUsage(usage: AgentContextUsage): AgentContextUsageView {
|
||||
const reading = usage.reading;
|
||||
const inputTokens = contextInputTokens(reading);
|
||||
const usableTokens = finiteContextNumber(reading?.usableInputTokens);
|
||||
const compactAtTokens = finiteContextNumber(reading?.compactAtTokens);
|
||||
const contextWindowTokens = finiteContextNumber(reading?.contextWindowTokens);
|
||||
const ratio = contextPressureRatio(reading);
|
||||
const compactRatio = usableTokens && compactAtTokens ? Math.min(1, compactAtTokens / usableTokens) : undefined;
|
||||
const tokenSource = reading?.tokenSource === "provider" ? "provider" : reading ? "estimate" : undefined;
|
||||
const estimate = tokenSource !== "provider";
|
||||
const remainingTokens = usableTokens !== undefined && inputTokens !== undefined
|
||||
? Math.max(0, usableTokens - inputTokens)
|
||||
: undefined;
|
||||
const base: AgentContextUsageView = {
|
||||
phase: "idle",
|
||||
ratio,
|
||||
compactRatio,
|
||||
ring: 0,
|
||||
label: "未测量",
|
||||
detail: "发出下一条消息后，这里显示下一次请求离压缩还有多远。",
|
||||
inputTokens,
|
||||
remainingTokens,
|
||||
usableTokens,
|
||||
compactAtTokens,
|
||||
contextWindowTokens,
|
||||
protocolBytes: contextProtocolBytes(reading),
|
||||
tokenSource,
|
||||
estimate,
|
||||
breakdown: contextBreakdown(reading),
|
||||
lastCompaction: usage.lastCompaction,
|
||||
};
|
||||
if (usage.compactionPending) {
|
||||
const basis = usage.compactionPending.basis === "bytes" ? "消息体积已到兜底线" : "已到上下文压缩线";
|
||||
return { ...base, phase: "compacting", ring: 1, label: "压缩中", detail: `${basis}，正在把较早对话收成检查点，最近两轮原样保留。` };
|
||||
}
|
||||
if (!reading) return base;
|
||||
if (usage.readingStale) {
|
||||
return { ...base, phase: "stale", ring: ratio === undefined ? 0 : Math.min(1, ratio / (compactRatio || 0.85)), label: "刚压缩", detail: "上一份读数是压缩前的；下一次模型调用会给出压缩后的占用。" };
|
||||
}
|
||||
if (ratio === undefined || usableTokens === undefined || usableTokens <= 0) {
|
||||
const measured = inputTokens === undefined ? "窗口未知" : `约 ${formatContextTokens(inputTokens)} Token`;
|
||||
return { ...base, phase: "unknown", label: measured, detail: "这个模型没有配置可确认的上下文窗口，不能给出占用百分比；对话过长时仍会按条数和体积压缩。" };
|
||||
}
|
||||
const line = compactRatio && compactRatio > 0 ? compactRatio : 0.85;
|
||||
const ring = Math.max(0, Math.min(1, ratio / line));
|
||||
const percent = Math.round(ratio * 100);
|
||||
const source = estimate ? "本地估算" : "模型实测校准";
|
||||
if (ratio >= line) {
|
||||
return { ...base, phase: "compress", ring: 1, label: `${percent}%`, detail: `已到压缩线（输入预算的 ${Math.round(line * 100)}%）。下一次调用前会暂停，把历史收成检查点后再继续。当前 ${formatContextTokens(inputTokens)} / ${formatContextTokens(usableTokens)}（${source}）。` };
|
||||
}
|
||||
if (ring >= 0.72) {
|
||||
return { ...base, phase: "watch", ring, label: `${percent}%`, detail: `接近压缩。当前 ${formatContextTokens(inputTokens)} / ${formatContextTokens(usableTokens)}（${source}），到 ${formatContextTokens(compactAtTokens)} 时开始压缩。` };
|
||||
}
|
||||
return { ...base, phase: "ok", ring, label: `${percent}%`, detail: `当前 ${formatContextTokens(inputTokens)} / ${formatContextTokens(usableTokens)}（${source}）。满格表示到达压缩线，不是窗口已经 100% 用完。` };
|
||||
}
|
||||
Reference in New Issue
Block a user