smashimo/oh-my-pi
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(model): added bracket-affix stripping and string-keyed resolution cache

Browse Source 
- Replaced WeakMap model cache with provider/id string keys for stable reuse.
- Returned official model ids directly when matched, before heuristics.
- Collapsed non-message token path to system prompt and tool schema totals.
This commit is contained in:
can1357
2026-06-06 22:21:58 +02:00
parent 20d19e8002
commit fde55bf927
5 changed files with 171 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/coding-agent/src/config/model-equivalence.ts
+22 -9
View File
@@ -58,7 +58,7 @@ const EMPTY_COMPILED_EQUIVALENCE: CompiledEquivalenceConfig = {
};
const kModelResolutionCache = Symbol("model-equivalence.resolutionCache");
interface CompiledEquivalenceConfigWithCache extends CompiledEquivalenceConfig {
[kModelResolutionCache]?: WeakMap<Model<Api>, ResolvedCanonicalModel>;
[kModelResolutionCache]?: Map<string, ResolvedCanonicalModel>;
}
const FAMILY_EXTRACTION_PATTERNS = [
/(?:^|[/:._-])((?:claude|gemini|gpt|grok|glm|qwen|minimax|kimi|deepseek|llama|gemma|nova|mistral|ministral|pixtral|codestral|devstral|magistral|ernie|doubao|seed|aion|olmo|molmo|nemotron|palmyra|command|codex|coder|o[1345])[-a-z0-9.]+)(?::|$)/i,
@@ -128,10 +128,18 @@ function normalizeCanonicalIdKey(canonicalId: string): string {
return canonicalId.trim().toLowerCase();
}
function getCanonicalSuffixAliasKey(candidate: string): string {
return PENALTY_HAS_UPPERCASE.test(candidate) ? normalizeCanonicalIdKey(candidate) : candidate;
}
export function formatCanonicalVariantSelector(model: Model<Api>): string {
return `${model.provider}/${model.id}`;
}
function getModelResolutionCacheKey(model: Model<Api>): string {
return `${model.provider}\0${model.id}`;
}
function buildOverrideMap(overrides: Record<string, string> | undefined): Map<string, string> {
const result = new Map<string, string>();
if (!overrides) {
@@ -728,10 +736,10 @@ function getPreferredFallbackCanonicalCandidate(modelId: string, candidates: rea
function resolveCanonicalIdForModel(
model: Model<Api>,
selector: string,
equivalence: CompiledEquivalenceConfig,
referenceData: CanonicalReferenceData,
): ResolvedCanonicalModel {
const selector = formatCanonicalVariantSelector(model);
const normalizedSelector = normalizeSelectorKey(selector);
if (equivalence.overrides.has(normalizedSelector)) {
@@ -752,10 +760,14 @@ function resolveCanonicalIdForModel(
return { id: claudeFamilyAlias, source: claudeFamilyAlias === model.id ? "bundled" : "heuristic" };
}
if (referenceData.officialIds.has(model.id) && !model.id.includes("/") && !model.id.includes(":")) {
return { id: model.id, source: "bundled" };
}
const heuristicCandidates = getHeuristicCanonicalCandidates(model.id, referenceData.officialIds);
const officialMatches = new Set(heuristicCandidates.filter(candidate => referenceData.officialIds.has(candidate)));
for (const candidate of heuristicCandidates) {
const aliased = referenceData.suffixAliases.get(normalizeCanonicalIdKey(candidate));
const aliased = referenceData.suffixAliases.get(getCanonicalSuffixAliasKey(candidate));
if (aliased) {
officialMatches.add(aliased);
}
@@ -814,17 +826,18 @@ export function buildCanonicalModelIndex(
const compiledWithCache = compiledEquivalence as CompiledEquivalenceConfigWithCache;
let modelCache = compiledWithCache[kModelResolutionCache];
if (!modelCache) {
modelCache = new WeakMap<Model<Api>, ResolvedCanonicalModel>();
modelCache = new Map<string, ResolvedCanonicalModel>();
compiledWithCache[kModelResolutionCache] = modelCache;
}
for (const model of models) {
let canonical = modelCache.get(model);
if (!canonical) {
canonical = resolveCanonicalIdForModel(model, compiledEquivalence, referenceData);
modelCache.set(model, canonical);
}
const selector = formatCanonicalVariantSelector(model);
const cacheKey = getModelResolutionCacheKey(model);
let canonical = modelCache.get(cacheKey);
if (!canonical) {
canonical = resolveCanonicalIdForModel(model, selector, compiledEquivalence, referenceData);
modelCache.set(cacheKey, canonical);
}
const variant: CanonicalModelVariant = {
canonicalId: canonical.id,
selector,
packages/coding-agent/src/config/model-id-affixes.ts
+39 -22
View File
@@ -4,34 +4,49 @@ const MODEL_ID_SEGMENT_PATTERN = /[a-z0-9.:-]+/g;
const MODEL_FAMILY_PREFIX_PATTERN =
/^(claude|gemini|gpt|grok|glm|qwen|deepseek|kimi|mimo|doubao|ernie|gpt-oss|gemma|minimax|step|command|jamba|llama|o[1345])/i;
function hasDigit(value: string): boolean {
return /\d/.test(value);
function normalizeModelIdWhitespace(value: string): string {
return value.trim().replace(/\s+/g, " ");
}
/** Ordering for model-like segments: longest first, ties broken lexicographically. */
function compareSegmentPreference(left: string, right: string): number {
if (left.length !== right.length) {
return right.length - left.length;
}
return left.localeCompare(right);
return left.length !== right.length ? right.length - left.length : left.localeCompare(right);
}
export function getModelLikeIdSegments(modelId: string): string[] {
const normalized = normalizeModelIdWhitespace(modelId).toLowerCase();
if (!normalized) return [];
const segments = (normalized.match(MODEL_ID_SEGMENT_PATTERN) ?? []).filter(
segment => MODEL_FAMILY_PREFIX_PATTERN.test(segment) && hasDigit(segment),
);
const unique = [...new Set(segments)];
unique.sort(compareSegmentPreference);
return unique;
const matches = normalizeModelIdWhitespace(modelId).toLowerCase().match(MODEL_ID_SEGMENT_PATTERN);
if (!matches) return [];
const segments = new Set<string>();
for (const segment of matches) {
if (MODEL_FAMILY_PREFIX_PATTERN.test(segment) && /\d/.test(segment)) segments.add(segment);
}
return [...segments].sort(compareSegmentPreference);
}
export function getLongestModelLikeIdSegment(modelId: string): string | undefined {
return getModelLikeIdSegments(modelId)[0];
const matches = normalizeModelIdWhitespace(modelId).toLowerCase().match(MODEL_ID_SEGMENT_PATTERN);
if (!matches) return undefined;
let best: string | undefined;
for (const segment of matches) {
if (
MODEL_FAMILY_PREFIX_PATTERN.test(segment) &&
/\d/.test(segment) &&
(best === undefined || compareSegmentPreference(segment, best) < 0)
) {
best = segment;
}
}
return best;
}
function normalizeModelIdWhitespace(value: string): string {
return value.trim().replace(/\s+/g, " ");
function hasBracketAffixMarker(value: string): boolean {
for (let index = 0; index < value.length; index++) {
const code = value.charCodeAt(index);
if (code === 91 || code === 93 || code === 0x3010 || code === 0x3011) {
return true;
}
}
return false;
}
/**
@@ -39,18 +54,20 @@ function normalizeModelIdWhitespace(value: string): string {
* upstream model id, e.g.
* "[Kiro] claude-opus-4-8" -> "claude-opus-4-8"
* "[gcli转] gemini-3.1-pro-preview [假流]" -> "gemini-3.1-pro-preview"
*
* Candidates are returned most-stripped first: both ends, then leading-only, then trailing-only.
*/
export function getBracketStrippedModelIdCandidates(modelId: string): string[] {
if (!hasBracketAffixMarker(modelId)) return [];
const normalized = normalizeModelIdWhitespace(modelId);
if (!normalized) return [];
const candidates = new Set<string>();
const withoutLeading = normalizeModelIdWhitespace(normalized.replace(LEADING_BRACKETED_AFFIX_PATTERN, ""));
const strippedLeading = normalized.replace(LEADING_BRACKETED_AFFIX_PATTERN, "");
const withoutLeading = normalizeModelIdWhitespace(strippedLeading);
const withoutTrailing = normalizeModelIdWhitespace(normalized.replace(TRAILING_BRACKETED_AFFIX_PATTERN, ""));
const withoutBoth = normalizeModelIdWhitespace(
normalized.replace(LEADING_BRACKETED_AFFIX_PATTERN, "").replace(TRAILING_BRACKETED_AFFIX_PATTERN, ""),
);
const withoutBoth = normalizeModelIdWhitespace(strippedLeading.replace(TRAILING_BRACKETED_AFFIX_PATTERN, ""));
const candidates = new Set<string>();
for (const candidate of [withoutBoth, withoutLeading, withoutTrailing]) {
if (candidate && candidate !== normalized) {
candidates.add(candidate);
packages/coding-agent/src/modes/utils/context-usage.ts
+10 -6
View File
@@ -37,6 +37,9 @@ export interface ContextBreakdown {
freeTokens: number;
}
const EMPTY_STRING_PARTS: readonly string[] = [];
const EMPTY_TOOLS: ReadonlyArray<Pick<Tool, "name" | "description" | "parameters">> = [];
export function estimateSkillsTokens(skills: readonly Skill[]): number {
const fragments: string[] = [];
for (const skill of skills) {
@@ -75,15 +78,16 @@ export function estimateToolSchemaTokens(
* messages walked incrementally as new entries append.
*/
export function computeNonMessageTokens(session: AgentSession): number {
const parts = computeNonMessageBreakdown(session);
return parts.systemPromptTokens + parts.systemContextTokens + parts.toolsTokens + parts.skillsTokens;
const systemPromptParts = session.systemPrompt ?? EMPTY_STRING_PARTS;
const tools = session.agent?.state?.tools ?? EMPTY_TOOLS;
return countTokens(systemPromptParts) + estimateToolSchemaTokens(tools);
}
/**
* Shared helper for the four non-message token totals. Single source of truth
* for both `computeNonMessageTokens` (status-line incremental cache) and
* `computeContextBreakdown` (/context panel). The split avoids drift between
* the two surfaces — they MUST report the same numbers.
* Shared helper for the four non-message token totals used by
* `computeContextBreakdown` (/context panel). Keep this category split stable:
* the status-line fast path intentionally uses the equivalent collapsed total
* in `computeNonMessageTokens`.
*/
function computeNonMessageBreakdown(session: AgentSession): {
skillsTokens: number;