dsh-agy-link 0.4.16 → 0.4.19

package/CHANGELOG.md

@@ -1,5 +1,29 @@
 # Changelog
 
+## 0.4.19 (2026-08-26)
+
+- **Fixed: Antigravity Models Missing From the Model Picker (登录成功、额度正常但模型列表为空 — issue #1).**
+  - **Root cause**: when `agy models` lists a bare Gemini base alongside its effort variants (the agy 1.1.13 output shape, e.g. `gemini-3.7-flash` + `gemini-3.7-flash-medium`), `foldEfforts` emitted the base id TWICE — once as the folded base entry and once as a verbatim row. DSH's `llm.listModels` contract throws `INVALID_CATALOG` on any duplicate model id, and the host's model-catalog builder then drops the ENTIRE Antigravity provider group from the picker — so login and quota panels looked perfectly healthy while no Antigravity model could be selected.
+  - **Fix (three layers)**: (1) `foldEfforts` now absorbs the bare base into its folded entry instead of duplicating it; (2) `parseModelsOutput` dedupes repeated raw slugs (first occurrence wins); (3) `AgyAdapter.listModels` dedupes ids as a final guard, so even a user-configured `fallbackModels` list containing repeats can never nuke the whole group.
+  - Regression tests pin both bare-base-plus-variants shapes and the duplicate-slug parse, plus an adapter-level uniqueness guard test.
+  - **Observability**: when the adapter-level guard drops duplicate ids it now logs which ids were removed (`model catalog contained duplicate ids [...]`) so field instances surface in DSH server logs instead of being silently masked. When reporting picker issues, attach the `/agy doctor` report (`~/.dsh/agy-link/diagnostics/doctor-*.md`) and the raw `agy models` output.
+
+## 0.4.18 (2026-08-25)
+
+- **Quota Fallback Never Clobbers Good Data (5h=100%/weekly-missing 根因).**
+  - **What happened**: `retrieveUserQuotaSummary` transiently failed (proxy blip) while `fetchAvailableModels` still answered; the per-model fallback then OVERWROTE the stored family entry with a single-window partial shape — weeklyFraction vanished and the 5h row received wrong-window numbers (observed 5h=100%, reset a week out, weekly `—`, while the live API actually reported 5h 84% / weekly 47%).
+  - **`mergeFallbackFamilyQuota`**: last-known-good complete family data now always wins over partial fallback; the fallback only fills families with no usable previous entry (first-ever refresh). Verified live: both endpoints answer correctly and a successful summary refresh fully restores the display.
+  - Added `scripts/diag-quota.mts` one-shot endpoint probe for future incidents.
+
+## 0.4.17 (2026-08-25)
+
+- **Ghost-Cooldown & Quota-Display Fix (额度显示 0% 根因).**
+  - **What happened**: the UI showed 5h quota as 0% while `agy` reported 98% — the parsed quota data was CORRECT all along, but (a) any active local cooldown forced the 5h bar to render 0%, and (b) the loose rate-limit classifier kept creating ghost cooldowns: it scanned the ENTIRE stdout (model prose mentioning "rate limit"/"quota", hash fragments containing "429") and matched bare keywords, so an unrelated tool/permission error froze a healthy account out of rotation with a 15-minute+ cooldown (captured real reason: `rate limit reached: declaring permissions: cortex tool write_to_file … invalid tool call error`).
+  - **Hard vs soft classification**: new `looksLikeHardRateLimit` (RESOURCE_EXHAUSTED / code·status·HTTP 429 / too many requests / individual quota reached / quota exceeded·reached·exhausted / rate limit exceeded·reached·hit) is the ONLY pattern allowed to put an account into cooldown; soft signals (model overloaded / high traffic) still shape the error message but never cool accounts.
+  - **Scan scope narrowed**: error classification reads stderr + the result envelope's error field only — stdout (event JSON + model prose) no longer participates.
+  - **Honest quota bars**: a local cooldown no longer overwrites the server-reported fraction with 0%; it now appends a `· 本地冷却中` note next to the reset time instead.
+  - Regression tests pin the exact incident text (cortex tool permission error) as a non-rate-limit fixture.
+
 ## 0.4.16 (2026-08-24)
 
 - **External Re-Login Sync (换号自动/手动同步).**

package/dist/client.js

@@ -963,8 +963,8 @@ body.dark,
 					const info = acc.quotas[familyKey];
 					const cd = acc.cooldowns[familyKey];
 					const inCooldown = cd && cd.cooldownUntil > Date.now();
-					let pct5h = typeof info?.remainingFraction === "number" && Number.isFinite(info.remainingFraction) ? Math.max(0, Math.min(100, Math.round(info.remainingFraction * 100))) : -1;
-					if (inCooldown) pct5h = 0;
+					const pct5h = typeof info?.remainingFraction === "number" && Number.isFinite(info.remainingFraction) ? Math.max(0, Math.min(100, Math.round(info.remainingFraction * 100))) : -1;
+					const cdNote = inCooldown ? " · 本地冷却中" : "";
 					const w5h = formatQuotaWindow(info?.resetTime);
 					const pctWeekly = typeof info?.weeklyFraction === "number" && Number.isFinite(info.weeklyFraction) ? Math.max(0, Math.min(100, Math.round(info.weeklyFraction * 100))) : -1;
 					const wWeekly = formatQuotaWindow(info?.weeklyResetTime);
@@ -1052,7 +1052,7 @@ body.dark,
 						fontSize: "12.5px",
 						marginBottom: "5px",
 						color: "var(--agy-text-primary)"
-					} }, brandIcon(FAMILY_BRAND[familyKey], 14), h("span", null, label)), renderLine("5h 额度", pct5h, c5h, w5h.resetText), renderLine("周额度", pctWeekly, cWeekly, wWeekly.resetText));
+					} }, brandIcon(FAMILY_BRAND[familyKey], 14), h("span", null, label)), renderLine("5h 额度", pct5h, c5h, w5h.resetText ? w5h.resetText + cdNote : cdNote.replace(/^ · /, "")), renderLine("周额度", pctWeekly, cWeekly, wWeekly.resetText));
 				};
 				const renderedAccountCards = accounts.map((acc) => {
 					const isPrimary = acc.id === pool?.primaryAccountId;

package/dist/index.js

@@ -115,9 +115,25 @@ function extractAuthUrl(text) {
 	if (!m) return void 0;
 	return m[0].replace(/[)\]>.,;\x27\x22]+$/, "");
 }
+/**
+* HARD, server-issued rate-limit signatures — the ONLY patterns allowed to
+* put an account into cooldown. Deliberately narrow: bare `429` or
+* `rate limit` matched incidental substrings in the wild (hash/UUID
+* fragments, model prose mentioning quotas, unrelated tool/permission
+* errors quoting such words) and produced ghost cooldowns that froze
+* healthy accounts out of rotation.
+*/
+function looksLikeHardRateLimit(text) {
+	if (!text) return false;
+	return /RESOURCE_EXHAUSTED|code[ :]?429\b|status[ :]?429\b|HTTP[ :]?429\b|too many requests|individual quota reached|quota (?:exceeded|reached|exhausted)|rate[ -]?limit(?:ed)? (?:exceeded|reached|hit)|exceeded (?:your |the )?quota/i.test(text);
+}
+/**
+* Soft heuristic adds capacity signals (model overloaded / high traffic).
+* Shapes the user-facing error message only — NEVER cools an account down.
+*/
 function looksLikeRateLimit(text) {
 	if (!text) return false;
-	return /429|too many requests|resource_exhausted|quota exceeded|quota reached|individual quota reached|rate limit|model overloaded|server.*experiencing high traffic|exceeded.*quota/i.test(text);
+	return looksLikeHardRateLimit(text) || /model overloaded|experiencing high traffic/i.test(text);
 }
 /**
 * Parse reset duration in milliseconds from rate-limit / quota-exhausted error strings.
@@ -1058,7 +1074,7 @@ function parseModelsOutput(stdout) {
 	if (text === "") return [];
 	try {
 		const list = extractModelList(JSON.parse(text));
-		if (list) return list;
+		if (list) return dedupeBySlug(list);
 	} catch {}
 	const out = [];
 	for (const line of text.split(/\n/)) {
@@ -1074,6 +1090,17 @@ function parseModelsOutput(stdout) {
 			label: t
 		});
 	}
+	return dedupeBySlug(out);
+}
+/** First occurrence wins: duplicate raw slugs would become duplicate catalog ids. */
+function dedupeBySlug(raw) {
+	const seen = /* @__PURE__ */ new Set();
+	const out = [];
+	for (const r of raw) {
+		if (seen.has(r.slug)) continue;
+		seen.add(r.slug);
+		out.push(r);
+	}
 	return out;
 }
 function extractModelList(parsed) {
@@ -1173,7 +1200,7 @@ function foldEfforts(raw) {
 	};
 	folded.sort((a, b) => rank(a) - rank(b));
 	verbatim.sort((a, b) => rank(a) - rank(b));
-	return [...folded, ...verbatim];
+	return [...folded, ...verbatim.filter((e) => !bases.has(e.id))];
 }
 function stripEffortLabel(label, eff) {
 	const re = new RegExp("\\s*\\(?" + eff + "\\)?\\s*$", "i");
@@ -2055,12 +2082,25 @@ var AgyAdapter = class extends LlmAdapter {
 	}
 	async listModels(_provider) {
 		this.deps.catalog.refreshIfNeeded();
-		return this.deps.catalog.get().models.map((m) => ({
-			provider: PROVIDER_ID,
-			id: m.id,
-			name: m.name,
-			inputModalities: ["text", "image"]
-		}));
+		const cat = this.deps.catalog.get();
+		const seen = /* @__PURE__ */ new Set();
+		const models = [];
+		const dropped = [];
+		for (const m of cat.models) {
+			if (seen.has(m.id)) {
+				dropped.push(m.id);
+				continue;
+			}
+			seen.add(m.id);
+			models.push({
+				provider: PROVIDER_ID,
+				id: m.id,
+				name: m.name,
+				inputModalities: ["text", "image"]
+			});
+		}
+		if (dropped.length > 0) this.warnOnce("catalog-dupes", "model catalog contained duplicate ids [" + dropped.join(", ") + "] — kept first occurrence so DSH does not drop the whole provider group (INVALID_CATALOG)");
+		return models;
 	}
 	async resolveModel(_provider, model, _signal) {
 		const cfg = this.deps.getConfig();
@@ -2364,11 +2404,8 @@ var AgyAdapter = class extends LlmAdapter {
 			const conversationId = streamCid ?? diffed;
 			const r = rec.getResultEvent();
 			const consumable = r !== null && (r.ok || r.response !== "");
-			const isRateLimit = looksLikeRateLimit([
-				outcome.stderrTail,
-				outcome.stdout,
-				parser.stats.lastResultError
-			].filter(Boolean).join(" "));
+			const rawErrText = [outcome.stderrTail, parser.stats.lastResultError].filter(Boolean).join(" ");
+			const isRateLimit = looksLikeRateLimit(rawErrText);
 			let failure = null;
 			if (outcome.aborted) failure = {
 				kind: "aborted",
@@ -2423,7 +2460,7 @@ var AgyAdapter = class extends LlmAdapter {
 				}
 			} else {
 				const effectiveRateLimit = isRateLimit || looksLikeRateLimit(failure.message);
-				if (account && effectiveRateLimit) this.deps.pool?.recordFailure(account.id, family, failure.message);
+				if (account && looksLikeHardRateLimit(rawErrText)) this.deps.pool?.recordFailure(account.id, family, failure.message);
 				if (account && (failure.code === Err.AUTH || /invalid_grant|not signed in|auth/i.test(failure.message))) this.deps.pool?.markAuthRequired(account.id, failure.message);
 				if (!isAux && sessionAccountKey !== "") {
 					if (failure.code === Err.AUTH || effectiveRateLimit || failure.message && /conversation.*(not found|invalid|not recognized|expired|does not exist)|session.*(expired|invalid)/i.test(failure.message)) this.deps.store.delete(sessionAccountKey);
@@ -27668,6 +27705,18 @@ var PoolAuthFlow = class {
 };
 //#endregion
 //#region src/host/quota.ts
+/**
+* When the quota-summary endpoint transiently fails, per-model fallback
+* data carries a SINGLE window (sometimes the weekly one) and no weekly
+* fields. Overwriting a previously COMPLETE family entry with that partial
+* shape dropped weeklyFraction to none and put wrong-window numbers into
+* the 5h row (observed: 5h=100% / reset a week out / weekly missing).
+* Rule: last-known-good complete data always wins over partial fallback.
+*/
+function mergeFallbackFamilyQuota(prev, fallback) {
+	if (prev && typeof prev.remainingFraction === "number") return prev;
+	return fallback;
+}
 function detectEmailFromAgyLogs(homeDir) {
 	const logDir = join(homeDir, ".gemini", "antigravity-cli", "log");
 	if (!existsSync(logDir)) return void 0;
@@ -27972,11 +28021,11 @@ var QuotaService = class {
 				remainingFraction: remaining,
 				resetTime
 			});
-			if (!familyQuotas[fam]) familyQuotas[fam] = {
+			if (!familyQuotas[fam]) familyQuotas[fam] = mergeFallbackFamilyQuota(account.quotas[fam], {
 				remainingFraction: remaining,
 				resetTime,
 				updatedAt: now
-			};
+			});
 			else if (familyQuotas[fam].remainingFraction === void 0) {
 				const curRemaining = familyQuotas[fam].remainingFraction ?? 1;
 				if (remaining < curRemaining) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-agy-link",
-  "version": "0.4.16",
+  "version": "0.4.19",
   "description": "Google Antigravity (agy CLI) models for DeepSeek Harness — stream Gemini/Claude/GPT-OSS subscriptions into DSH with thinking, tool activity, token usage and in-GUI Google OAuth login.",
   "type": "module",
   "license": "MIT",