clone/gemini-cli
1
0
Fork 0
mirror of https://github.com/google-gemini/gemini-cli.git synced 2026-06-01 19:03:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(core): wire up the new ContextManager and AgentChatHistory (#25409)

Browse Source
This commit is contained in:
joshualitt
2026-04-27 11:50:00 -07:00
committed by GitHub
parent 98aca28985
commit 71f313b51a
35 changed files with 707 additions and 377 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
packages/core/src/utils/tokenCalculation.ts
View File

@@ -29,12 +29,14 @@ const MAX_CHARS_FOR_FULL_HEURISTIC = 100_000;
// standard multimodal responses are typically depth 1.
const MAX_RECURSION_DEPTH = 3;
const DEFAULT_CHARS_PER_TOKEN = 4;
/**
* Heuristic estimation of tokens for a text string.
*/
function estimateTextTokens(text: string): number {
function estimateTextTokens(text: string, charsPerToken: number): number {
if (text.length > MAX_CHARS_FOR_FULL_HEURISTIC) {
return text.length / 4;
return text.length / charsPerToken;
}
let tokens = 0;
@@ -73,25 +75,33 @@ function estimateMediaTokens(part: Part): number | undefined {
* Heuristic estimation for tool responses, avoiding massive string copies
* and accounting for nested Gemini 3 multimodal parts.
*/
function estimateFunctionResponseTokens(part: Part, depth: number): number {
function estimateFunctionResponseTokens(
part: Part,
depth: number,
charsPerToken: number,
): number {
const fr = part.functionResponse;
if (!fr) return 0;
let totalTokens = (fr.name?.length ?? 0) / 4;
let totalTokens = (fr.name?.length ?? 0) / charsPerToken;
const response = fr.response as unknown;
if (typeof response === 'string') {
totalTokens += response.length / 4;
totalTokens += response.length / charsPerToken;
} else if (response !== undefined && response !== null) {
// For objects, stringify only the payload, not the whole Part object.
totalTokens += JSON.stringify(response).length / 4;
totalTokens += JSON.stringify(response).length / charsPerToken;
}
// Gemini 3: Handle nested multimodal parts recursively.
// eslint-disable-next-line @typescript-eslint/no-unsafe-type-assertion
const nestedParts = (fr as unknown as { parts?: Part[] }).parts;
if (nestedParts && nestedParts.length > 0) {
totalTokens += estimateTokenCountSync(nestedParts, depth + 1);
totalTokens += estimateTokenCountSync(
nestedParts,
depth + 1,
charsPerToken,
);
}
return totalTokens;
@@ -100,11 +110,12 @@ function estimateFunctionResponseTokens(part: Part, depth: number): number {
/**
* Estimates token count for parts synchronously using a heuristic.
* - Text: character-based heuristic (ASCII vs CJK) for small strings, length/4 for massive ones.
* - Non-text (Tools, etc): JSON string length / 4.
* - Non-text (Tools, etc): JSON string length / charsPerToken.
*/
export function estimateTokenCountSync(
parts: Part[],
depth: number = 0,
charsPerToken: number = DEFAULT_CHARS_PER_TOKEN,
): number {
if (depth > MAX_RECURSION_DEPTH) {
return 0;
@@ -113,9 +124,9 @@ export function estimateTokenCountSync(
let totalTokens = 0;
for (const part of parts) {
if (typeof part.text === 'string') {
totalTokens += estimateTextTokens(part.text);
totalTokens += estimateTextTokens(part.text, charsPerToken);
} else if (part.functionResponse) {
totalTokens += estimateFunctionResponseTokens(part, depth);
totalTokens += estimateFunctionResponseTokens(part, depth, charsPerToken);
} else {
const mediaEstimate = estimateMediaTokens(part);
if (mediaEstimate !== undefined) {
@@ -123,7 +134,7 @@ export function estimateTokenCountSync(
} else {
// Fallback for other non-text parts (e.g., functionCall).
// Note: JSON.stringify(part) here is safe as these parts are typically small.
totalTokens += JSON.stringify(part).length / 4;
totalTokens += JSON.stringify(part).length / charsPerToken;
}
}
}
@@ -162,9 +173,9 @@ export async function calculateRequestTokenCount(
} catch (error) {
// Fallback to local estimation if the API call fails
debugLogger.debug('countTokens API failed:', error);
return estimateTokenCountSync(parts);
return estimateTokenCountSync(parts, 0, DEFAULT_CHARS_PER_TOKEN);
}
}
return estimateTokenCountSync(parts);
return estimateTokenCountSync(parts, 0, DEFAULT_CHARS_PER_TOKEN);
}