perf: trim hot path allocations

This commit is contained in:
Peter Steinberger
2026-04-20 23:11:44 +01:00
parent 9f9235b692
commit e0c01bf956
5 changed files with 41 additions and 27 deletions

View File

@@ -409,7 +409,7 @@ export function chunkMarkdown(
}
}
current = kept;
currentChars = kept.reduce((sum, entry) => sum + estimateStringChars(entry.line) + 1, 0);
currentChars = acc;
};
for (let i = 0; i < lines.length; i += 1) {

View File

@@ -83,6 +83,13 @@ describe("detectImageReferences", () => {
expect(refs.some((r) => r.type === "path")).toBe(true);
});
it("does not leak parser state between calls", () => {
expectSingleImageReference("[media attached: /tmp/first.png (image/png)]");
expectSingleImageReference("[Image: source: /tmp/second.jpg]");
expectSingleImageReference("See file:///tmp/third.webp");
expectSingleImageReference("See ./fourth.jpeg");
});
it("handles various image extensions", () => {
const extensions = ["png", "jpg", "jpeg", "gif", "webp", "bmp", "tiff", "heic"];
for (const ext of extensions) {

View File

@@ -41,6 +41,11 @@ const MESSAGE_IMAGE_REGEX_SOURCE =
const FILE_URL_REGEX_SOURCE = "file://[^\\s<>\"'`\\]]+\\.(?:" + IMAGE_EXTENSION_PATTERN + ")";
const PATH_REGEX_SOURCE =
"(?:^|\\s|[\"'`(])((\\.\\.?/|[~/])[^\\s\"'`()\\[\\]]*\\.(?:" + IMAGE_EXTENSION_PATTERN + "))";
const MEDIA_ATTACHED_PATTERN = /\[media attached(?:\s+\d+\/\d+)?:\s*([^\]]+)\]/gi;
const MEDIA_ATTACHED_PATH_PATTERN = new RegExp(MEDIA_ATTACHED_PATH_REGEX_SOURCE, "i");
const MESSAGE_IMAGE_PATTERN = new RegExp(MESSAGE_IMAGE_REGEX_SOURCE, "gi");
const FILE_URL_PATTERN = new RegExp(FILE_URL_REGEX_SOURCE, "gi");
const PATH_PATTERN = new RegExp(PATH_REGEX_SOURCE, "gi");
/**
* Matches the opaque media URI written by the Gateway's claim-check offload:
@@ -251,13 +256,12 @@ export function detectImageReferences(prompt: string): DetectedImageRef[] {
// Pattern for [media attached: path (type) | url] or [media attached N/M: path (type) | url] format
// Each bracket = ONE file. The | separates path from URL, not multiple files.
// Multi-file format uses separate brackets on separate lines.
const mediaAttachedPattern = /\[media attached(?:\s+\d+\/\d+)?:\s*([^\]]+)\]/gi;
const mediaAttachedPathPattern = new RegExp(MEDIA_ATTACHED_PATH_REGEX_SOURCE, "i");
const messageImagePattern = new RegExp(MESSAGE_IMAGE_REGEX_SOURCE, "gi");
const fileUrlPattern = new RegExp(FILE_URL_REGEX_SOURCE, "gi");
const pathPattern = new RegExp(PATH_REGEX_SOURCE, "gi");
MEDIA_ATTACHED_PATTERN.lastIndex = 0;
MESSAGE_IMAGE_PATTERN.lastIndex = 0;
FILE_URL_PATTERN.lastIndex = 0;
PATH_PATTERN.lastIndex = 0;
let match: RegExpExecArray | null;
while ((match = mediaAttachedPattern.exec(prompt)) !== null) {
while ((match = MEDIA_ATTACHED_PATTERN.exec(prompt)) !== null) {
const content = match[1];
// Skip "[media attached: N files]" header lines
@@ -283,14 +287,14 @@ export function detectImageReferences(prompt: string): DetectedImageRef[] {
// Format is: path (type) | url OR just: path (type)
// Path may contain spaces (e.g., "ChatGPT Image Apr 21.png")
// Use non-greedy .+? to stop at first image extension
const pathMatch = content.match(mediaAttachedPathPattern);
const pathMatch = content.match(MEDIA_ATTACHED_PATH_PATTERN);
if (pathMatch?.[1]) {
addPathRef(pathMatch[1].trim());
}
}
// Pattern for [Image: source: /path/...] format from messaging systems
while ((match = messageImagePattern.exec(prompt)) !== null) {
while ((match = MESSAGE_IMAGE_PATTERN.exec(prompt)) !== null) {
const raw = match[1]?.trim();
if (raw) {
addPathRef(raw);
@@ -300,7 +304,7 @@ export function detectImageReferences(prompt: string): DetectedImageRef[] {
// Remote HTTP(S) URLs are intentionally ignored. Native image injection is local-only.
// Pattern for file:// URLs - treat as paths since loadWebMedia handles them
while ((match = fileUrlPattern.exec(prompt)) !== null) {
while ((match = FILE_URL_PATTERN.exec(prompt)) !== null) {
const raw = match[0];
const dedupeKey = normalizeRefForDedupe(raw);
if (seen.has(dedupeKey)) {
@@ -322,7 +326,7 @@ export function detectImageReferences(prompt: string): DetectedImageRef[] {
// - ./relative/path.ext
// - ../parent/path.ext
// - ~/home/path.ext
while ((match = pathPattern.exec(prompt)) !== null) {
while ((match = PATH_PATTERN.exec(prompt)) !== null) {
// Use capture group 1 (the path without delimiter prefix); skip if undefined
if (match[1]) {
addPathRef(match[1]);

View File

@@ -438,19 +438,18 @@ export function resolveFreshestSessionStoreMatchFromStoreKeys(
store: Record<string, SessionEntry>,
storeKeys: string[],
): { key: string; entry: SessionEntry } | undefined {
const matches = storeKeys
.map((key) => {
const entry = store[key];
return entry ? { key, entry } : undefined;
})
.filter((match): match is { key: string; entry: SessionEntry } => match !== undefined);
if (matches.length === 0) {
return undefined;
let freshest: { key: string; entry: SessionEntry } | undefined;
for (const key of storeKeys) {
const entry = store[key];
if (!entry) {
continue;
}
const match = { key, entry };
if (!freshest || (match.entry.updatedAt ?? 0) > (freshest.entry.updatedAt ?? 0)) {
freshest = match;
}
}
if (matches.length === 1) {
return matches[0];
}
return [...matches].toSorted((a, b) => (b.entry.updatedAt ?? 0) - (a.entry.updatedAt ?? 0))[0];
return freshest;
}
export function resolveFreshestSessionEntryFromStoreKeys(
@@ -484,9 +483,13 @@ function findFreshestStoreMatch(
if (matches.size === 0) {
return undefined;
}
return [...matches.values()].toSorted(
(a, b) => (b.entry.updatedAt ?? 0) - (a.entry.updatedAt ?? 0),
)[0];
let freshest: { entry: SessionEntry; key: string } | undefined;
for (const match of matches.values()) {
if (!freshest || (match.entry.updatedAt ?? 0) > (freshest.entry.updatedAt ?? 0)) {
freshest = match;
}
}
return freshest;
}
/**

View File

@@ -388,7 +388,7 @@ export function chunkMarkdown(
}
}
current = kept;
currentChars = kept.reduce((sum, entry) => sum + estimateStringChars(entry.line) + 1, 0);
currentChars = acc;
};
for (let i = 0; i < lines.length; i += 1) {