@project-ava-supernova/core 0.2.71

package/dist/auto/intent-gate.js

@@ -0,0 +1,247 @@
+import { logger } from '../core/logger.js';
+import { avaEvents } from '../dataset/emitter.js';
+import { chargeCredits, extractUsage } from '../billing/meter.js';
+/**
+ * Resolve a cheap fast model for intent gating. Prefers Qwen 3.5 Flash;
+ * falls back to anything cheap + fast if that's unavailable. Returns
+ * null if nothing viable exists — callers then skip the gate entirely.
+ */
+export function resolveIntentGateModel(providerRegistry, availableProviders, hasPlatform) {
+    // Platform priority — Qwen Flash family first
+    if (hasPlatform) {
+        for (const id of ['qwen3.5-flash', 'qwen-flash']) {
+            const resolved = providerRegistry.resolveModel(`platform:${id}`);
+            if (resolved)
+                return { provider: resolved.provider, model: resolved.model };
+        }
+    }
+    // BYOK — any small/fast model will do the classification.
+    // V4 Flash replaces `deepseek-chat` (retired upstream 2026-07-24) and is the
+    // cheaper of the two V4 tiers, which is what this gate wants.
+    // Nemotron 3.5 Lightning is last, and measured rather than assumed. Over the
+    // 20-prompt gate set on 2026-08-16 it scored 20/20, same as Qwen 3.5 Flash,
+    // at 36.3 output tokens against 29.9 and a 909ms median against 752ms —
+    // slightly chattier and slower, marginally cheaper because its output rate
+    // is half. Ordered last on that basis, not excluded.
+    //
+    // It is only viable because the request shaper now sends
+    // chat_template_kwargs {thinking:false} for it. WITHOUT that it scores
+    // 0/20 — every reply is 120 tokens of "Here's a thinking process:", the
+    // JSON never arrives, and the gate silently returns null and falls back to
+    // the regex Conductor. That failure is invisible: it looks like the gate
+    // simply is not helping. Do not move this entry without checking that the
+    // parameter still reaches the provider.
+    //
+    // What it buys: a BYOK user holding only an NVIDIA key had NO gate at all
+    // before this — no entry here resolved, so every prompt went to the regex.
+    const fallbacks = [
+        'qwen3.5-flash',
+        'deepseek-v4-flash',
+        'nvidia/nemotron-3.5-lightning-30b-a3b',
+    ];
+    for (const id of fallbacks) {
+        const direct = providerRegistry.resolveModel(id);
+        if (direct)
+            return { provider: direct.provider, model: direct.model };
+        for (const provName of availableProviders) {
+            if (provName === 'platform')
+                continue;
+            const qualified = providerRegistry.resolveModel(`${provName}:${id}`);
+            if (qualified)
+                return { provider: qualified.provider, model: qualified.model };
+        }
+    }
+    return null;
+}
+const SYSTEM_PROMPT = `You are the routing classifier for Ava, an agentic coding assistant.
+
+For every user prompt, decide whether Ava (the coordinator) can handle it directly with her own tools (file read/write, shell, search, git, etc.), or whether the task genuinely needs her specialist team (Architect, Scout, Challenger, Verifier, Researcher, etc.) to plan and verify before acting.
+
+**DIRECT** (team NOT needed) — the vast majority of prompts:
+- Questions: "what does this code do?", "explain this function", "why is this broken?"
+- Single-file edits: "rename this variable", "fix this bug", "add a loading state"
+- Quick tasks: "run the tests", "commit this", "check git status"
+- Look-ups: "find all uses of X", "show me the schema", "what's in this folder?"
+- Small features where the approach is obvious
+
+**TEAM** (needed) — only when genuinely warranted:
+- Multi-file architectural changes ("refactor auth across the whole app")
+- Explicit review / audit requests ("full security audit", "architectural review")
+- Non-trivial system design ("design the event bus", "plan the migration strategy")
+- Complex debugging across unknown code ("trace why this happens in prod")
+- Teach-mode curriculum requests (always team)
+
+Bias HEAVILY toward DIRECT. The team is expensive and slow. Only engage it when the user has signalled depth or the task is genuinely large.
+
+Respond with a single JSON object, no prose outside it:
+{"needs_team": true|false, "reasoning": "<one short sentence>"}`;
+/**
+ * Classify a user prompt. Returns null on any failure — callers then
+ * fall back to the regex-based Conductor gate.
+ */
+export async function classifyIntent(opts) {
+    const { userMessage, mode, provider, model, signal } = opts;
+    // Hard rules that don't need Flash
+    if (mode === 'chat')
+        return { needsTeam: false, reasoning: 'Chat mode — no orchestration by design' };
+    if (mode === 'teach')
+        return { needsTeam: true, reasoning: 'Teach mode — curriculum always uses the team' };
+    const promptForFlash = `Mode: ${mode}\nUser message: ${userMessage.slice(0, 2000)}`;
+    try {
+        const response = await provider.createCompletion({
+            model: model.id,
+            messages: [
+                { role: 'system', content: SYSTEM_PROMPT },
+                { role: 'user', content: promptForFlash },
+            ],
+            temperature: 0,
+            max_tokens: 120,
+            // No reasoning pass. This is a classifier: it returns a label, and the
+            // thinking that precedes it is pure latency in front of the real work.
+            // Measured on these exact parameters — 601 output tokens with it on,
+            // 9 with it off, identical answer. max_tokens does NOT contain it;
+            // the cap is 120 and the reasoning pass ignores it.
+            enable_thinking: false,
+        }, signal);
+        // Meter the call — we burned Flash tokens regardless of whether the
+        // response is parseable downstream. Fire-and-forget event.
+        chargeCredits('light_call', {
+            model: model.id,
+            rawTokens: extractUsage(response.usage),
+        });
+        const assistantMsg = response.choices?.[0]?.message;
+        if (!assistantMsg) {
+            logger.warn('[intent-gate] Flash returned no choices');
+            return null;
+        }
+        const content = assistantMsg.content;
+        const text = typeof content === 'string'
+            ? content
+            : Array.isArray(content)
+                ? content.map((p) => (p && typeof p === 'object' && 'text' in p ? p.text : '')).join('')
+                : '';
+        // Extract the JSON object — Flash sometimes wraps in ```json fences
+        const jsonMatch = text.match(/\{[\s\S]*?"needs_team"[\s\S]*?\}/);
+        if (!jsonMatch) {
+            logger.warn('[intent-gate] Flash response did not contain valid JSON: ' + text.slice(0, 200));
+            return null;
+        }
+        const parsed = JSON.parse(jsonMatch[0]);
+        if (typeof parsed.needs_team !== 'boolean') {
+            logger.warn('[intent-gate] Flash response missing needs_team boolean');
+            return null;
+        }
+        const decision = {
+            needsTeam: parsed.needs_team,
+            reasoning: parsed.reasoning || '',
+        };
+        avaEvents.emit('routing_decision', {
+            classification: 'intent_gate',
+            chosen_model: decision.needsTeam ? 'team' : 'coordinator_direct',
+            reason: decision.reasoning,
+        });
+        return decision;
+    }
+    catch (err) {
+        logger.warn('[intent-gate] Flash classification failed: ' + (err instanceof Error ? err.message : String(err)));
+        return null;
+    }
+}
+const TEACH_DEPTH_SYSTEM_PROMPT = `You are classifying a user message in Ava's Teach mode.
+
+Decide: does this message ask to CREATE a new learning curriculum, or is it CONTINUATION inside an existing one?
+
+CREATION → "creation":
+- "teach me X" / "show me how to learn Y" / "tutor me on Z"
+- "I want / need / would like to learn / study / understand X"
+- "I'd like to start learning X" / "I'm trying to pick up X" / "I want to get into X"
+- "create / build / make / design / generate a curriculum / course / learning path / lesson plan / syllabus"
+- "where do I start with X" / "help me understand X from scratch"
+- Anything that asks for a structured learning path on a topic Ava hasn't already started
+
+CONTINUATION → "continuation":
+- "continue" / "next lesson" / "next" / "go on"
+- "another example" / "show me more" / "explain again"
+- "I'm stuck" / "I don't get it" / "what does this mean"
+- "let's do the quiz" / "what was wrong with my answer" / "try a harder one"
+- Follow-up questions inside an active lesson
+- Short / ambiguous messages that look like in-flow chatter
+
+Bias toward CONTINUATION for short or ambiguous messages — the Tutor can always ask the user to clarify, but spawning the full team for "another example" wastes ~5× the tokens.
+
+Respond with a single JSON object, no prose outside it:
+{"depth": "creation" | "continuation", "reasoning": "<one short sentence>"}`;
+/**
+ * Classify a Teach-mode user message as curriculum-creation vs delivery
+ * continuation. Replaces the regex-based heuristic in
+ * `detectConductorDepth` for Teach mode — natural phrasings like "I'd
+ * like to learn Rust" or "help me understand React from scratch" don't
+ * fit a fixed regex but a Flash classifier handles them cleanly.
+ *
+ * Returns null on failure — caller falls back to the regex-based
+ * `detectConductorDepth` so a Flash outage never blocks Teach mode.
+ *
+ * Cost: same Flash call shape as `classifyIntent` — ~200 input + ~30
+ * output tokens. For Teach mode this REPLACES the intent-gate's hard
+ * early-return, so there's no net extra Flash call per turn.
+ */
+export async function classifyTeachDepth(opts) {
+    const { userMessage, provider, model, signal } = opts;
+    try {
+        const response = await provider.createCompletion({
+            model: model.id,
+            messages: [
+                { role: 'system', content: TEACH_DEPTH_SYSTEM_PROMPT },
+                { role: 'user', content: userMessage.slice(0, 2000) },
+            ],
+            temperature: 0,
+            max_tokens: 120,
+            // No reasoning pass. This is a classifier: it returns a label, and the
+            // thinking that precedes it is pure latency in front of the real work.
+            // Measured on these exact parameters — 601 output tokens with it on,
+            // 9 with it off, identical answer. max_tokens does NOT contain it;
+            // the cap is 120 and the reasoning pass ignores it.
+            enable_thinking: false,
+        }, signal);
+        chargeCredits('light_call', {
+            model: model.id,
+            rawTokens: extractUsage(response.usage),
+        });
+        const assistantMsg = response.choices?.[0]?.message;
+        if (!assistantMsg) {
+            logger.warn('[intent-gate] Teach depth Flash returned no choices');
+            return null;
+        }
+        const content = assistantMsg.content;
+        const text = typeof content === 'string'
+            ? content
+            : Array.isArray(content)
+                ? content.map((p) => (p && typeof p === 'object' && 'text' in p ? p.text : '')).join('')
+                : '';
+        const jsonMatch = text.match(/\{[\s\S]*?"depth"[\s\S]*?\}/);
+        if (!jsonMatch) {
+            logger.warn('[intent-gate] Teach depth Flash response did not contain valid JSON: ' + text.slice(0, 200));
+            return null;
+        }
+        const parsed = JSON.parse(jsonMatch[0]);
+        if (parsed.depth !== 'creation' && parsed.depth !== 'continuation') {
+            logger.warn('[intent-gate] Teach depth Flash returned invalid depth: ' + String(parsed.depth));
+            return null;
+        }
+        const decision = {
+            depth: parsed.depth === 'creation' ? 'full' : 'light',
+            reasoning: parsed.reasoning || '',
+        };
+        avaEvents.emit('routing_decision', {
+            classification: 'teach_depth',
+            chosen_model: decision.depth === 'full' ? 'teach_full_team' : 'teach_tutor_only',
+            reason: decision.reasoning,
+        });
+        return decision;
+    }
+    catch (err) {
+        logger.warn('[intent-gate] Teach depth Flash classification failed: ' + (err instanceof Error ? err.message : String(err)));
+        return null;
+    }
+}
+//# sourceMappingURL=intent-gate.js.map

package/dist/auto/intent-gate.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"intent-gate.js","sourceRoot":"","sources":["../../src/auto/intent-gate.ts"],"names":[],"mappings":"AAGA,OAAO,EAAE,MAAM,EAAE,MAAM,mBAAmB,CAAC;AAC3C,OAAO,EAAE,SAAS,EAAE,MAAM,uBAAuB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AA+ClE;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CACpC,gBAAkC,EAClC,kBAA+B,EAC/B,WAAoB;IAEpB,8CAA8C;IAC9C,IAAI,WAAW,EAAE,CAAC;QAChB,KAAK,MAAM,EAAE,IAAI,CAAC,eAAe,EAAE,YAAY,CAAC,EAAE,CAAC;YACjD,MAAM,QAAQ,GAAG,gBAAgB,CAAC,YAAY,CAAC,YAAY,EAAE,EAAE,CAAC,CAAC;YACjE,IAAI,QAAQ;gBAAE,OAAO,EAAE,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,CAAC;QAC9E,CAAC;IACH,CAAC;IAED,0DAA0D;IAC1D,6EAA6E;IAC7E,8DAA8D;IAC9D,6EAA6E;IAC7E,4EAA4E;IAC5E,wEAAwE;IACxE,2EAA2E;IAC3E,qDAAqD;IACrD,EAAE;IACF,yDAAyD;IACzD,uEAAuE;IACvE,wEAAwE;IACxE,2EAA2E;IAC3E,yEAAyE;IACzE,0EAA0E;IAC1E,wCAAwC;IACxC,EAAE;IACF,0EAA0E;IAC1E,2EAA2E;IAC3E,MAAM,SAAS,GAAG;QAChB,eAAe;QACf,mBAAmB;QACnB,uCAAuC;KACxC,CAAC;IACF,KAAK,MAAM,EAAE,IAAI,SAAS,EAAE,CAAC;QAC3B,MAAM,MAAM,GAAG,gBAAgB,CAAC,YAAY,CAAC,EAAE,CAAC,CAAC;QACjD,IAAI,MAAM;YAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,CAAC,QAAQ,EAAE,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,CAAC;QACtE,KAAK,MAAM,QAAQ,IAAI,kBAAkB,EAAE,CAAC;YAC1C,IAAI,QAAQ,KAAK,UAAU;gBAAE,SAAS;YACtC,MAAM,SAAS,GAAG,gBAAgB,CAAC,YAAY,CAAC,GAAG,QAAQ,IAAI,EAAE,EAAE,CAAC,CAAC;YACrE,IAAI,SAAS;gBAAE,OAAO,EAAE,QAAQ,EAAE,SAAS,CAAC,QAAQ,EAAE,KAAK,EAAE,SAAS,CAAC,KAAK,EAAE,CAAC;QACjF,CAAC;IACH,CAAC;IAED,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,aAAa,GAAG;;;;;;;;;;;;;;;;;;;;;gEAqB0C,CAAC;AAEjE;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,IAMpC;IACC,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC;IAE5D,mCAAmC;IACnC,IAAI,IAAI,KAAK,MAAM;QAAE,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,SAAS,EAAE,wCAAwC,EAAE,CAAC;IACtG,IAAI,IAAI,KAAK,OAAO;QAAE,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,SAAS,EAAE,8CAA8C,EAAE,CAAC;IAE5G,MAAM,cAAc,GAAG,SAAS,IAAI,mBAAmB,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,EAAE,CAAC;IAEpF,IAAI,CAAC;QACH,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,gBAAgB,CAC9C;YACE,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,QAAQ,EAAE;gBACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,aAAa,EAAE;gBAC1C,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,cAAc,EAAE;aAC1C;YACD,WAAW,EAAE,CAAC;YACd,UAAU,EAAE,GAAG;YACf,uEAAuE;YACvE,uEAAuE;YACvE,qEAAqE;YACrE,mEAAmE;YACnE,oDAAoD;YACpD,eAAe,EAAE,KAAK;SACvB,EACD,MAAM,CACP,CAAC;QAEF,oEAAoE;QACpE,2DAA2D;QAC3D,aAAa,CAAC,YAAY,EAAE;YAC1B,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,SAAS,EAAE,YAAY,CAAE,QAAgC,CAAC,KAA2C,CAAC;SACvG,CAAC,CAAC;QAEH,MAAM,YAAY,GAAG,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,CAAC;QACpD,IAAI,CAAC,YAAY,EAAE,CAAC;YAClB,MAAM,CAAC,IAAI,CAAC,yCAAyC,CAAC,CAAC;YACvD,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,OAAO,GAAG,YAAY,CAAC,OAAkB,CAAC;QAChD,MAAM,IAAI,GAAG,OAAO,OAAO,KAAK,QAAQ;YACtC,CAAC,CAAC,OAAO;YACT,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC;gBACtB,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,MAAM,IAAI,CAAC,CAAC,CAAC,CAAE,CAAsB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC9G,CAAC,CAAC,EAAE,CAAC;QAET,oEAAoE;QACpE,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,kCAAkC,CAAC,CAAC;QACjE,IAAI,CAAC,SAAS,EAAE,CAAC;YACf,MAAM,CAAC,IAAI,CAAC,2DAA2D,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;YAC9F,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAiD,CAAC;QACxF,IAAI,OAAO,MAAM,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;YAC3C,MAAM,CAAC,IAAI,CAAC,yDAAyD,CAAC,CAAC;YACvE,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,QAAQ,GAAuB;YACnC,SAAS,EAAE,MAAM,CAAC,UAAU;YAC5B,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,EAAE;SAClC,CAAC;QAEF,SAAS,CAAC,IAAI,CAAC,kBAAkB,EAAE;YACjC,cAAc,EAAE,aAAa;YAC7B,YAAY,EAAE,QAAQ,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,oBAAoB;YAChE,MAAM,EAAE,QAAQ,CAAC,SAAS;SAC3B,CAAC,CAAC;QAEH,OAAO,QAAQ,CAAC;IAClB,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,MAAM,CAAC,IAAI,CAAC,6CAA6C,GAAG,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAChH,OAAO,IAAI,CAAC;IACd,CAAC;AACH,CAAC;AAED,MAAM,yBAAyB,GAAG;;;;;;;;;;;;;;;;;;;;;;;4EAuB0C,CAAC;AAE7E;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,KAAK,UAAU,kBAAkB,CAAC,IAKxC;IACC,MAAM,EAAE,WAAW,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC;IAEtD,IAAI,CAAC;QACH,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,gBAAgB,CAC9C;YACE,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,QAAQ,EAAE;gBACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,yBAAyB,EAAE;gBACtD,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,WAAW,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,EAAE;aACtD;YACD,WAAW,EAAE,CAAC;YACd,UAAU,EAAE,GAAG;YACf,uEAAuE;YACvE,uEAAuE;YACvE,qEAAqE;YACrE,mEAAmE;YACnE,oDAAoD;YACpD,eAAe,EAAE,KAAK;SACvB,EACD,MAAM,CACP,CAAC;QAEF,aAAa,CAAC,YAAY,EAAE;YAC1B,KAAK,EAAE,KAAK,CAAC,EAAE;YACf,SAAS,EAAE,YAAY,CAAE,QAAgC,CAAC,KAA2C,CAAC;SACvG,CAAC,CAAC;QAEH,MAAM,YAAY,GAAG,QAAQ,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,CAAC;QACpD,IAAI,CAAC,YAAY,EAAE,CAAC;YAClB,MAAM,CAAC,IAAI,CAAC,qDAAqD,CAAC,CAAC;YACnE,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,OAAO,GAAG,YAAY,CAAC,OAAkB,CAAC;QAChD,MAAM,IAAI,GAAG,OAAO,OAAO,KAAK,QAAQ;YACtC,CAAC,CAAC,OAAO;YACT,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC;gBACtB,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,MAAM,IAAI,CAAC,CAAC,CAAC,CAAE,CAAsB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC9G,CAAC,CAAC,EAAE,CAAC;QAET,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,6BAA6B,CAAC,CAAC;QAC5D,IAAI,CAAC,SAAS,EAAE,CAAC;YACf,MAAM,CAAC,IAAI,CAAC,uEAAuE,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;YAC1G,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAA2C,CAAC;QAClF,IAAI,MAAM,CAAC,KAAK,KAAK,UAAU,IAAI,MAAM,CAAC,KAAK,KAAK,cAAc,EAAE,CAAC;YACnE,MAAM,CAAC,IAAI,CAAC,0DAA0D,GAAG,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;YAC/F,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,QAAQ,GAAuB;YACnC,KAAK,EAAE,MAAM,CAAC,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO;YACrD,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,EAAE;SAClC,CAAC;QAEF,SAAS,CAAC,IAAI,CAAC,kBAAkB,EAAE;YACjC,cAAc,EAAE,aAAa;YAC7B,YAAY,EAAE,QAAQ,CAAC,KAAK,KAAK,MAAM,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,kBAAkB;YAChF,MAAM,EAAE,QAAQ,CAAC,SAAS;SAC3B,CAAC,CAAC;QAEH,OAAO,QAAQ,CAAC;IAClB,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,MAAM,CAAC,IAAI,CAAC,yDAAyD,GAAG,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;QAC5H,OAAO,IAAI,CAAC;IACd,CAAC;AACH,CAAC"}

package/dist/auto/longxiang-router.d.ts

@@ -0,0 +1,123 @@
+import type { TaskCategory } from './types.js';
+/**
+ * Longxiang mode (龙翔, "soaring dragon") — the Chinese-model fleet.
+ *
+ * The fourth fleet. Where Supernova is the polyglot performance stack and
+ * Aurora is the European sovereign stack, Longxiang is the Chinese-model
+ * stack, led by the strongest coder we serve.
+ *
+ * The AMBITION is open weights end to end — every seat a model whose weights
+ * are published under a licence permitting commercial use, so a deployment
+ * could be self-hosted with no vendor in the loop. That is not true yet; see
+ * the open-weights section below. The user-facing copy says "Chinese-model
+ * fleet" and will keep saying it until the weights are actually public.
+ *
+ *   - Kimi K3 (Moonshot)     — Coordinator AND Builder. The fleet's lead seat,
+ *                              held by one model deliberately: K3 is the
+ *                              strongest open coder available to us — 88.3
+ *                              Terminal-Bench 2.1 and 81.2 FrontierSWE, ahead
+ *                              of GLM-5.2 (81.0 / 74.4) — and it is also a
+ *                              strong tool-driver, which is what the
+ *                              coordinator seat actually needs. Modified MIT;
+ *                              attribution only triggers at 100M MAU / $20M
+ *                              monthly revenue.
+ *   - Qwen 3.7 Plus (Alibaba)— Mid-tier builds, vision input, long context.
+ *                              Native vision + video and 1M context, and the
+ *                              model our Builder persona is already tuned
+ *                              against on Supernova. Carries the work that
+ *                              doesn't need K3 depth but does need eyes.
+ *   - DeepSeek V4 Flash      — Intent gate + light personas + chat. MIT,
+ *                              $0.14/$0.28, 1M context. The cheapest place in
+ *                              the fleet to put volume.
+ *
+ * ── Access: exactly like the other fleets ─────────────────────────────────
+ *
+ * A signed-in plan runs Longxiang on credits; a BYOK user unlocks it once all
+ * three keys (Moonshot + Qwen + DeepSeek) are present. Same rule as Maestro,
+ * Supernova and Aurora — no special-casing.
+ *
+ * Cost note for the managed path: K3 is $3.00/$15.00, the priciest model we
+ * serve, so it carries the table's largest credit multiplier (7.28 — derived
+ * as exactly 2× Mistral Medium 3.5, which it doubles on both input and
+ * output). That is genuine cost passed through, not margin.
+ *
+ * ── Open-weights claim: NOT yet fully true ────────────────────────────────
+ *
+ * The fleet's headline promise is "open weights end to end". As of
+ * 2026-07-18 that is true of two of three seats:
+ *
+ *   - Kimi K3        — weights due 2026-07-27 (Modified MIT).      pending
+ *   - DeepSeek V4 Flash — MIT, weights public.                     ✅
+ *   - Qwen 3.7 Plus  — CLOSED. API-only, no public weights. Alibaba's
+ *                      account manager has told us it is "gearing up to
+ *                      open-source soon" with no date given.       ❌
+ *
+ * Do NOT publish the end-to-end claim until Qwen 3.7 Plus weights are
+ * actually public — verify at source, not from vendor intent. If Qwen has
+ * not opened by launch, the contingency is to swap the Base seat (and only
+ * the Base seat) to GLM-5.2, which is MIT today. GLM is a capable code
+ * *writer* but scores weakly on Tool-Decathlon, so it is a substitute for
+ * Qwen's mid-tier build work, never for K3's coordinator/Builder seat.
+ *
+ * Fleet named 2026-07-17; trademark search for "Longxiang" (classes 9 + 42)
+ * is still outstanding and must clear before the name goes public.
+ */
+/**
+ * Master switch for whether Longxiang is OFFERED to users.
+ *
+ * Deliberately a hand-flipped boolean and NOT a date check. Longxiang's
+ * launch depends on two things outside our control, either of which can
+ * slip past a calendar date:
+ *
+ *   1. Kimi K3 weights actually being published (expected 2026-07-27).
+ *   2. The "Longxiang" trademark search clearing (classes 9 + 42).
+ *
+ * A date-based auto-flip would launch the fleet — and its open-weights
+ * claim — on a day when neither may have happened. Flipping this by hand
+ * means launch stays a decision someone makes after verifying, not a timer
+ * that fires regardless.
+ *
+ * While false, the fleet is fully built and testable but is not listed in
+ * any picker and cannot be selected via the CLI's /route. The routing
+ * tables below stay live either way so the test suite can exercise them.
+ *
+ * Before flipping to true, verify AT SOURCE:
+ *   - K3 weights are on Hugging Face (not just announced)
+ *   - Qwen 3.7 Plus weights are public, OR the Base seat has been swapped
+ *     to GLM-5.2 and the marketing copy no longer claims "open end to end"
+ *   - The trademark search has come back clean
+ */
+export declare const LONGXIANG_ENABLED = true;
+/** The conductor that classifies tasks, picks specialists, runs the loop.
+ *  K3 holds this seat on tool-driving strength as much as raw coding. */
+export declare const LONGXIANG_COORDINATOR_ID = "kimi-k3";
+/** Builder agent — TaskExecutor spawn for any session task. Same model as the
+ *  coordinator: K3 is the best open coder we have (88.3 Terminal-Bench 2.1,
+ *  81.2 FrontierSWE), so there is nothing to gain by handing the build to a
+ *  weaker seat. Front and centre by design. */
+export declare const LONGXIANG_BUILDER_ID = "kimi-k3";
+/** Vision input override — Qwen 3.7 Plus, not K3. This is a COST call, not a
+ *  capability one: K3 sees natively too (text+image+video), but at $3/$15 it
+ *  is ~3x the mid-tier, and Qwen is the model our vision paths are already
+ *  tuned against on Supernova. K3 remains a genuine vision fallback. */
+export declare const LONGXIANG_VISION_ID = "qwen3.7-plus";
+/** Intent gate — the cheapest classifier in the fleet. V4 Flash at
+ *  $0.14/$0.28 with 1M context; short routing calls don't need K3. */
+export declare const LONGXIANG_INTENT_GATE_ID = "deepseek-v4-flash";
+export interface LongxiangRouteEntry {
+    modelId: string;
+    reason: string;
+    fallbackModelId?: string;
+    requiresVision?: boolean;
+    /** Used by `teach` — coordinator-tier model when curriculum is being created (depth='full'). */
+    creationModelId?: string;
+}
+export declare const LONGXIANG_ROUTES: Record<TaskCategory, LongxiangRouteEntry>;
+export declare const LONGXIANG_PERSONA_MODEL: Record<string, string>;
+/**
+ * Resolve the right model id for a persona in Longxiang mode. Falls back to
+ * `defaultModelId` when the persona isn't in the map (a new persona shipped
+ * without a Longxiang route — graceful degrade rather than throw).
+ */
+export declare function resolveLongxiangPersonaModel(persona: string, defaultModelId: string): string;
+//# sourceMappingURL=longxiang-router.d.ts.map

package/dist/auto/longxiang-router.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"longxiang-router.d.ts","sourceRoot":"","sources":["../../src/auto/longxiang-router.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAE/C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8DG;AAIH;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,eAAO,MAAM,iBAAiB,OAAO,CAAC;AAItC;yEACyE;AACzE,eAAO,MAAM,wBAAwB,YAAY,CAAC;AAElD;;;+CAG+C;AAC/C,eAAO,MAAM,oBAAoB,YAAY,CAAC;AAE9C;;;wEAGwE;AACxE,eAAO,MAAM,mBAAmB,iBAAiB,CAAC;AAElD;sEACsE;AACtE,eAAO,MAAM,wBAAwB,sBAAsB,CAAC;AAO5D,MAAM,WAAW,mBAAmB;IAClC,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,gGAAgG;IAChG,eAAe,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,eAAO,MAAM,gBAAgB,EAAE,MAAM,CAAC,YAAY,EAAE,mBAAmB,CA4BtE,CAAC;AAeF,eAAO,MAAM,uBAAuB,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CA8B1D,CAAC;AAEF;;;;GAIG;AACH,wBAAgB,4BAA4B,CAC1C,OAAO,EAAE,MAAM,EACf,cAAc,EAAE,MAAM,GACrB,MAAM,CAER"}

package/dist/auto/longxiang-router.js

@@ -0,0 +1,184 @@
+/**
+ * Longxiang mode (龙翔, "soaring dragon") — the Chinese-model fleet.
+ *
+ * The fourth fleet. Where Supernova is the polyglot performance stack and
+ * Aurora is the European sovereign stack, Longxiang is the Chinese-model
+ * stack, led by the strongest coder we serve.
+ *
+ * The AMBITION is open weights end to end — every seat a model whose weights
+ * are published under a licence permitting commercial use, so a deployment
+ * could be self-hosted with no vendor in the loop. That is not true yet; see
+ * the open-weights section below. The user-facing copy says "Chinese-model
+ * fleet" and will keep saying it until the weights are actually public.
+ *
+ *   - Kimi K3 (Moonshot)     — Coordinator AND Builder. The fleet's lead seat,
+ *                              held by one model deliberately: K3 is the
+ *                              strongest open coder available to us — 88.3
+ *                              Terminal-Bench 2.1 and 81.2 FrontierSWE, ahead
+ *                              of GLM-5.2 (81.0 / 74.4) — and it is also a
+ *                              strong tool-driver, which is what the
+ *                              coordinator seat actually needs. Modified MIT;
+ *                              attribution only triggers at 100M MAU / $20M
+ *                              monthly revenue.
+ *   - Qwen 3.7 Plus (Alibaba)— Mid-tier builds, vision input, long context.
+ *                              Native vision + video and 1M context, and the
+ *                              model our Builder persona is already tuned
+ *                              against on Supernova. Carries the work that
+ *                              doesn't need K3 depth but does need eyes.
+ *   - DeepSeek V4 Flash      — Intent gate + light personas + chat. MIT,
+ *                              $0.14/$0.28, 1M context. The cheapest place in
+ *                              the fleet to put volume.
+ *
+ * ── Access: exactly like the other fleets ─────────────────────────────────
+ *
+ * A signed-in plan runs Longxiang on credits; a BYOK user unlocks it once all
+ * three keys (Moonshot + Qwen + DeepSeek) are present. Same rule as Maestro,
+ * Supernova and Aurora — no special-casing.
+ *
+ * Cost note for the managed path: K3 is $3.00/$15.00, the priciest model we
+ * serve, so it carries the table's largest credit multiplier (7.28 — derived
+ * as exactly 2× Mistral Medium 3.5, which it doubles on both input and
+ * output). That is genuine cost passed through, not margin.
+ *
+ * ── Open-weights claim: NOT yet fully true ────────────────────────────────
+ *
+ * The fleet's headline promise is "open weights end to end". As of
+ * 2026-07-18 that is true of two of three seats:
+ *
+ *   - Kimi K3        — weights due 2026-07-27 (Modified MIT).      pending
+ *   - DeepSeek V4 Flash — MIT, weights public.                     ✅
+ *   - Qwen 3.7 Plus  — CLOSED. API-only, no public weights. Alibaba's
+ *                      account manager has told us it is "gearing up to
+ *                      open-source soon" with no date given.       ❌
+ *
+ * Do NOT publish the end-to-end claim until Qwen 3.7 Plus weights are
+ * actually public — verify at source, not from vendor intent. If Qwen has
+ * not opened by launch, the contingency is to swap the Base seat (and only
+ * the Base seat) to GLM-5.2, which is MIT today. GLM is a capable code
+ * *writer* but scores weakly on Tool-Decathlon, so it is a substitute for
+ * Qwen's mid-tier build work, never for K3's coordinator/Builder seat.
+ *
+ * Fleet named 2026-07-17; trademark search for "Longxiang" (classes 9 + 42)
+ * is still outstanding and must clear before the name goes public.
+ */
+// ── Launch flag ───────────────────────────────────────────────────────────
+/**
+ * Master switch for whether Longxiang is OFFERED to users.
+ *
+ * Deliberately a hand-flipped boolean and NOT a date check. Longxiang's
+ * launch depends on two things outside our control, either of which can
+ * slip past a calendar date:
+ *
+ *   1. Kimi K3 weights actually being published (expected 2026-07-27).
+ *   2. The "Longxiang" trademark search clearing (classes 9 + 42).
+ *
+ * A date-based auto-flip would launch the fleet — and its open-weights
+ * claim — on a day when neither may have happened. Flipping this by hand
+ * means launch stays a decision someone makes after verifying, not a timer
+ * that fires regardless.
+ *
+ * While false, the fleet is fully built and testable but is not listed in
+ * any picker and cannot be selected via the CLI's /route. The routing
+ * tables below stay live either way so the test suite can exercise them.
+ *
+ * Before flipping to true, verify AT SOURCE:
+ *   - K3 weights are on Hugging Face (not just announced)
+ *   - Qwen 3.7 Plus weights are public, OR the Base seat has been swapped
+ *     to GLM-5.2 and the marketing copy no longer claims "open end to end"
+ *   - The trademark search has come back clean
+ */
+export const LONGXIANG_ENABLED = true;
+// ── Coordinator + special-case routes (highest priority) ──────────────────
+/** The conductor that classifies tasks, picks specialists, runs the loop.
+ *  K3 holds this seat on tool-driving strength as much as raw coding. */
+export const LONGXIANG_COORDINATOR_ID = 'kimi-k3';
+/** Builder agent — TaskExecutor spawn for any session task. Same model as the
+ *  coordinator: K3 is the best open coder we have (88.3 Terminal-Bench 2.1,
+ *  81.2 FrontierSWE), so there is nothing to gain by handing the build to a
+ *  weaker seat. Front and centre by design. */
+export const LONGXIANG_BUILDER_ID = 'kimi-k3';
+/** Vision input override — Qwen 3.7 Plus, not K3. This is a COST call, not a
+ *  capability one: K3 sees natively too (text+image+video), but at $3/$15 it
+ *  is ~3x the mid-tier, and Qwen is the model our vision paths are already
+ *  tuned against on Supernova. K3 remains a genuine vision fallback. */
+export const LONGXIANG_VISION_ID = 'qwen3.7-plus';
+/** Intent gate — the cheapest classifier in the fleet. V4 Flash at
+ *  $0.14/$0.28 with 1M context; short routing calls don't need K3. */
+export const LONGXIANG_INTENT_GATE_ID = 'deepseek-v4-flash';
+export const LONGXIANG_ROUTES = {
+    // ── Depth routes → K3 (the Peak) ──────────────────────────────────────
+    // Builder territory. This is the seat the whole fleet is built around.
+    coding: { modelId: 'kimi-k3', reason: 'Kimi K3 — 88.3 Terminal-Bench 2.1, 81.2 FrontierSWE; strongest open coder in the fleet', fallbackModelId: 'qwen3.7-plus' },
+    // Planning leans on synthesis depth — K3 leads, Qwen carries the 1M ceiling.
+    planning: { modelId: 'kimi-k3', reason: 'Kimi K3 — reasoning depth for plan synthesis and specialist dispatch', fallbackModelId: 'qwen3.7-plus' },
+    // Security = CVE Researcher leads — deep reasoning over the attack surface.
+    security: { modelId: 'kimi-k3', reason: 'Kimi K3 — deep reasoning over attack surface', fallbackModelId: 'qwen3.7-plus' },
+    // ── Vision + long-context → Qwen 3.7 Plus (the Base) ──────────────────
+    // Both seats see natively; Qwen carries vision on cost, K3 is a real
+    // fallback (not a degraded one) if Qwen is unreachable.
+    vision: { modelId: 'qwen3.7-plus', reason: 'Qwen 3.7 Plus — native vision + video, 1M context, mid-tier cost', fallbackModelId: 'kimi-k3', requiresVision: true },
+    // Both K3 and Qwen are long-context; route on cost — Qwen is the cheaper
+    // place to push a large prompt, K3 is the depth ceiling.
+    long_context: { modelId: 'qwen3.7-plus', reason: 'Qwen 3.7 Plus — 1M context at mid-tier cost; K3 reserved for depth', fallbackModelId: 'kimi-k3' },
+    // Teach = Tutor + Curriculum Architect, medium depth. Curriculum *creation*
+    // upgrades to K3 — running the 5-persona prep team mid-tier goes shallow.
+    teach: { modelId: 'qwen3.7-plus', reason: 'Qwen 3.7 Plus — long-form coherence for tutorials and lesson delivery', fallbackModelId: 'deepseek-v4-flash', creationModelId: 'kimi-k3' },
+    // ── Volume routes → DeepSeek V4 Flash (the River) ─────────────────────
+    // image_gen routes a generate_image tool call out to Qwen-Image — the model
+    // here only orchestrates, so depth is wasted spend.
+    image_gen: { modelId: 'deepseek-v4-flash', reason: 'DeepSeek V4 Flash — orchestrates generate_image tool calls; depth not required at this layer', fallbackModelId: 'qwen3.7-plus' },
+    // Chat is a single-turn response — doesn't exercise the coordinator pattern.
+    chat: { modelId: 'deepseek-v4-flash', reason: 'DeepSeek V4 Flash — fast, cheapest tier in the fleet for typical chat turns', fallbackModelId: 'qwen3.7-plus' },
+    // Brainstorm = breadth over depth — cheap and creatively wide beats careful.
+    brainstorm: { modelId: 'deepseek-v4-flash', reason: 'DeepSeek V4 Flash — breadth and speed for ideation at the fleet\'s lowest cost', fallbackModelId: 'qwen3.7-plus' },
+};
+// ── Per-persona override map ──────────────────────────────────────────────
+//
+// Used when spawning specific personas. Persona is finer-grained than task
+// category — a "planning" task might invoke Architect AND Researcher within
+// the same orchestration. Persona override wins over the category route.
+//
+// Keys match persona names from packages/core/src/personas/definitions.ts.
+//
+// NOTE: as of writing, neither SUPERNOVA_PERSONA_MODEL nor AURORA_PERSONA_MODEL
+// has a live consumer — the maps are declared ahead of the wiring. This one
+// mirrors them deliberately so Longxiang doesn't become the odd fleet out when
+// that wiring lands.
+export const LONGXIANG_PERSONA_MODEL = {
+    // Heavy specialists — the seats where K3's coding + reasoning lead pays.
+    architect: 'kimi-k3', // plan depth, tool-aware
+    builder: 'kimi-k3', // the coder — fleet's whole point
+    researcher: 'kimi-k3', // synthesis depth
+    cve_researcher: 'kimi-k3', // deep reasoning over attack chain
+    ideator: 'kimi-k3', // reasoning depth 5
+    // Mid-tier — long output and anything that needs eyes.
+    content_writer: 'qwen3.7-plus', // flagship long-form coherence
+    design_reviewer: 'qwen3.7-plus', // sees images + video natively
+    // Light + mid-light specialists — volume tier.
+    scout: 'deepseek-v4-flash',
+    verifier: 'deepseek-v4-flash',
+    sequencer: 'deepseek-v4-flash',
+    challenger: 'deepseek-v4-flash',
+    integrator: 'deepseek-v4-flash',
+    curator: 'deepseek-v4-flash',
+    tutor: 'deepseek-v4-flash', // mid-depth, latency matters
+    code_reviewer: 'deepseek-v4-flash',
+    fact_checker: 'deepseek-v4-flash',
+    quiz_master: 'deepseek-v4-flash',
+    recon: 'deepseek-v4-flash',
+    scanner: 'deepseek-v4-flash',
+    security_verifier: 'deepseek-v4-flash',
+    security_reporter: 'deepseek-v4-flash',
+    curriculum_architect: 'deepseek-v4-flash',
+    explorer: 'deepseek-v4-flash',
+    refiner: 'deepseek-v4-flash',
+};
+/**
+ * Resolve the right model id for a persona in Longxiang mode. Falls back to
+ * `defaultModelId` when the persona isn't in the map (a new persona shipped
+ * without a Longxiang route — graceful degrade rather than throw).
+ */
+export function resolveLongxiangPersonaModel(persona, defaultModelId) {
+    return LONGXIANG_PERSONA_MODEL[persona.toLowerCase()] ?? defaultModelId;
+}
+//# sourceMappingURL=longxiang-router.js.map

package/dist/auto/longxiang-router.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"longxiang-router.js","sourceRoot":"","sources":["../../src/auto/longxiang-router.ts"],"names":[],"mappings":"AAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8DG;AAEH,6EAA6E;AAE7E;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAG,IAAI,CAAC;AAEtC,6EAA6E;AAE7E;yEACyE;AACzE,MAAM,CAAC,MAAM,wBAAwB,GAAG,SAAS,CAAC;AAElD;;;+CAG+C;AAC/C,MAAM,CAAC,MAAM,oBAAoB,GAAG,SAAS,CAAC;AAE9C;;;wEAGwE;AACxE,MAAM,CAAC,MAAM,mBAAmB,GAAG,cAAc,CAAC;AAElD;sEACsE;AACtE,MAAM,CAAC,MAAM,wBAAwB,GAAG,mBAAmB,CAAC;AAgB5D,MAAM,CAAC,MAAM,gBAAgB,GAA8C;IACzE,yEAAyE;IACzE,uEAAuE;IACvE,MAAM,EAAQ,EAAE,OAAO,EAAE,SAAS,EAAY,MAAM,EAAE,wFAAwF,EAAE,eAAe,EAAE,cAAc,EAAE;IACjL,6EAA6E;IAC7E,QAAQ,EAAM,EAAE,OAAO,EAAE,SAAS,EAAY,MAAM,EAAE,sEAAsE,EAAoB,eAAe,EAAE,cAAc,EAAE;IACjL,4EAA4E;IAC5E,QAAQ,EAAM,EAAE,OAAO,EAAE,SAAS,EAAY,MAAM,EAAE,8CAA8C,EAA4C,eAAe,EAAE,cAAc,EAAE;IAEjL,yEAAyE;IACzE,qEAAqE;IACrE,wDAAwD;IACxD,MAAM,EAAQ,EAAE,OAAO,EAAE,cAAc,EAAO,MAAM,EAAE,kEAAkE,EAAwB,eAAe,EAAE,SAAS,EAAE,cAAc,EAAE,IAAI,EAAE;IAClM,yEAAyE;IACzE,yDAAyD;IACzD,YAAY,EAAE,EAAE,OAAO,EAAE,cAAc,EAAO,MAAM,EAAE,oEAAoE,EAAsB,eAAe,EAAE,SAAS,EAAE;IAC5K,4EAA4E;IAC5E,0EAA0E;IAC1E,KAAK,EAAS,EAAE,OAAO,EAAE,cAAc,EAAO,MAAM,EAAE,uEAAuE,EAAmB,eAAe,EAAE,mBAAmB,EAAE,eAAe,EAAE,SAAS,EAAE;IAElN,yEAAyE;IACzE,4EAA4E;IAC5E,oDAAoD;IACpD,SAAS,EAAK,EAAE,OAAO,EAAE,mBAAmB,EAAE,MAAM,EAAE,8FAA8F,EAAE,eAAe,EAAE,cAAc,EAAE;IACvL,6EAA6E;IAC7E,IAAI,EAAU,EAAE,OAAO,EAAE,mBAAmB,EAAE,MAAM,EAAE,6EAA6E,EAAa,eAAe,EAAE,cAAc,EAAE;IACjL,6EAA6E;IAC7E,UAAU,EAAI,EAAE,OAAO,EAAE,mBAAmB,EAAE,MAAM,EAAE,gFAAgF,EAAW,eAAe,EAAE,cAAc,EAAE;CACnL,CAAC;AAEF,6EAA6E;AAC7E,EAAE;AACF,2EAA2E;AAC3E,4EAA4E;AAC5E,yEAAyE;AACzE,EAAE;AACF,2EAA2E;AAC3E,EAAE;AACF,gFAAgF;AAChF,4EAA4E;AAC5E,+EAA+E;AAC/E,qBAAqB;AAErB,MAAM,CAAC,MAAM,uBAAuB,GAA2B;IAC7D,yEAAyE;IACzE,SAAS,EAAY,SAAS,EAAa,yBAAyB;IACpE,OAAO,EAAc,SAAS,EAAa,kCAAkC;IAC7E,UAAU,EAAW,SAAS,EAAa,kBAAkB;IAC7D,cAAc,EAAO,SAAS,EAAa,mCAAmC;IAC9E,OAAO,EAAc,SAAS,EAAa,oBAAoB;IAE/D,uDAAuD;IACvD,cAAc,EAAO,cAAc,EAAQ,+BAA+B;IAC1E,eAAe,EAAM,cAAc,EAAQ,+BAA+B;IAE1E,+CAA+C;IAC/C,KAAK,EAAgB,mBAAmB;IACxC,QAAQ,EAAa,mBAAmB;IACxC,SAAS,EAAY,mBAAmB;IACxC,UAAU,EAAW,mBAAmB;IACxC,UAAU,EAAW,mBAAmB;IACxC,OAAO,EAAc,mBAAmB;IACxC,KAAK,EAAgB,mBAAmB,EAAG,6BAA6B;IACxE,aAAa,EAAQ,mBAAmB;IACxC,YAAY,EAAS,mBAAmB;IACxC,WAAW,EAAU,mBAAmB;IACxC,KAAK,EAAgB,mBAAmB;IACxC,OAAO,EAAc,mBAAmB;IACxC,iBAAiB,EAAI,mBAAmB;IACxC,iBAAiB,EAAI,mBAAmB;IACxC,oBAAoB,EAAC,mBAAmB;IACxC,QAAQ,EAAa,mBAAmB;IACxC,OAAO,EAAc,mBAAmB;CACzC,CAAC;AAEF;;;;GAIG;AACH,MAAM,UAAU,4BAA4B,CAC1C,OAAe,EACf,cAAsB;IAEtB,OAAO,uBAAuB,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC,IAAI,cAAc,CAAC;AAC1E,CAAC"}

package/dist/auto/model-router.d.ts

@@ -0,0 +1,42 @@
+import type { ProviderRegistry } from '../providers/provider-registry.js';
+import type { TaskCategory, RouteResult, UserRoutePreferences } from './types.js';
+export { ROUTING_MODES, isRoutingMode } from './routing-modes.js';
+export type { RoutingMode } from './routing-modes.js';
+import type { RoutingMode } from './routing-modes.js';
+/**
+ * Routes task categories to the best available model.
+ * Respects user overrides, checks provider availability, falls back gracefully.
+ *
+ * `mode` selects the routing table — 'auto' (default) uses the static
+ * Qwen-coordinated table; 'supernova' uses the polyglot table that picks
+ * different models per task based on what each is best at.
+ */
+export declare class ModelRouter {
+    private providerRegistry;
+    private availableProviders;
+    private hasPlatform;
+    private userPreferences;
+    private mode;
+    constructor(providerRegistry: ProviderRegistry, availableProviders: Set<string>, platformKey?: string, userPreferences?: UserRoutePreferences, mode?: RoutingMode);
+    /**
+     * Route a task category to the best available model.
+     * Returns null only if no model at all is available (shouldn't happen in practice).
+     *
+     * `teachDepth` is consulted only for `category === 'teach'`. When the
+     * Flash gate upstream determines this is a curriculum-creation request
+     * ('full'), the router upgrades to `entry.creationModelId` (the
+     * coordinator-tier model) — running the 5-persona prep team on a
+     * mid-tier model produced shallow curriculums. Delivery turns
+     * ('light') and undefined (Flash unavailable / non-Teach) keep the
+     * standard `entry.modelId`.
+     */
+    route(category: TaskCategory, modelOverride?: string, teachDepth?: 'light' | 'full'): RouteResult | null;
+    /** Check if a category can be routed to any model */
+    canRoute(category: TaskCategory): boolean;
+    /** Update user preferences (e.g. from settings change) */
+    setPreferences(prefs: UserRoutePreferences): void;
+    private resolveModel;
+    private resolveAnyModel;
+    private isProviderAvailable;
+}
+//# sourceMappingURL=model-router.d.ts.map

package/dist/auto/model-router.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"model-router.d.ts","sourceRoot":"","sources":["../../src/auto/model-router.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,mCAAmC,CAAC;AAC1E,OAAO,KAAK,EAAE,YAAY,EAAE,WAAW,EAAE,oBAAoB,EAAE,MAAM,YAAY,CAAC;AAQlF,OAAO,EAAE,aAAa,EAAE,aAAa,EAAE,MAAM,oBAAoB,CAAC;AAClE,YAAY,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACtD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AAoDtD;;;;;;;GAOG;AACH,qBAAa,WAAW;IACtB,OAAO,CAAC,gBAAgB,CAAmB;IAC3C,OAAO,CAAC,kBAAkB,CAAc;IACxC,OAAO,CAAC,WAAW,CAAU;IAC7B,OAAO,CAAC,eAAe,CAAuB;IAC9C,OAAO,CAAC,IAAI,CAAc;gBAGxB,gBAAgB,EAAE,gBAAgB,EAClC,kBAAkB,EAAE,GAAG,CAAC,MAAM,CAAC,EAC/B,WAAW,CAAC,EAAE,MAAM,EACpB,eAAe,CAAC,EAAE,oBAAoB,EACtC,IAAI,GAAE,WAAoB;IAS5B;;;;;;;;;;;OAWG;IACH,KAAK,CAAC,QAAQ,EAAE,YAAY,EAAE,aAAa,CAAC,EAAE,MAAM,EAAE,UAAU,CAAC,EAAE,OAAO,GAAG,MAAM,GAAG,WAAW,GAAG,IAAI;IA8DxG,qDAAqD;IACrD,QAAQ,CAAC,QAAQ,EAAE,YAAY,GAAG,OAAO;IAIzC,0DAA0D;IAC1D,cAAc,CAAC,KAAK,EAAE,oBAAoB,GAAG,IAAI;IAIjD,OAAO,CAAC,YAAY;IA4BpB,OAAO,CAAC,eAAe;IA4CvB,OAAO,CAAC,mBAAmB;CAG5B"}