muonroi-cli 1.8.3 → 1.8.5

package/dist/src/cli/cost-forensics.js

@@ -46,21 +46,21 @@ export function collectCostForensics(sessionId) {
     const chain = getSessionChain(sessionId);
     const placeholders = chain.map(() => "?").join(",");
     const rows = db
-        .prepare(`
-    SELECT id, source, model, message_seq, input_tokens, output_tokens,
-           cache_read_tokens, cache_creation_tokens, cost_micros, created_at,
-           provider_options_shape
-    FROM usage_events
-    WHERE session_id IN (${placeholders})
-    ORDER BY id ASC
+        .prepare(`
+    SELECT id, source, model, message_seq, input_tokens, output_tokens,
+           cache_read_tokens, cache_creation_tokens, cost_micros, created_at,
+           provider_options_shape
+    FROM usage_events
+    WHERE session_id IN (${placeholders})
+    ORDER BY id ASC
   `)
         .all(...chain);
     const counts = db
-        .prepare(`
-    SELECT event_type, COUNT(*) AS c
-    FROM interaction_logs
-    WHERE session_id IN (${placeholders})
-    GROUP BY event_type
+        .prepare(`
+    SELECT event_type, COUNT(*) AS c
+    FROM interaction_logs
+    WHERE session_id IN (${placeholders})
+    GROUP BY event_type
   `)
         .all(...chain);
     const countMap = new Map(counts.map((r) => [r.event_type, r.c]));
@@ -83,6 +83,7 @@ export function collectCostForensics(sessionId) {
     let totalCacheCreation = 0;
     let totalCostMicros = 0;
     let peakSingleCallInput = 0;
+    let peakSingleCallFresh = 0;
     for (const e of events) {
         totalInput += e.inputTokens;
         totalOutput += e.outputTokens;
@@ -91,6 +92,9 @@ export function collectCostForensics(sessionId) {
         totalCostMicros += e.costMicros;
         if (e.inputTokens > peakSingleCallInput)
             peakSingleCallInput = e.inputTokens;
+        const fresh = Math.max(0, e.inputTokens - e.cacheReadTokens);
+        if (fresh > peakSingleCallFresh)
+            peakSingleCallFresh = fresh;
     }
     const cacheable = totalInput;
     const cacheHitRatio = cacheable > 0 ? totalCacheRead / cacheable : 0;
@@ -106,6 +110,7 @@ export function collectCostForensics(sessionId) {
         totalCostUsd: totalCostMicros / 1_000_000,
         cacheHitRatio,
         peakSingleCallInput,
+        peakSingleCallFresh,
         events,
         experience: selectSessionExperience(sessionId),
     };
@@ -159,7 +164,8 @@ export function printCostForensics(summary, opts = {}) {
     w(`Total output tokens: ${formatNum(summary.totalOutput)}`);
     w(`Cache read tokens:   ${formatNum(summary.totalCacheRead)} (${formatPct(summary.cacheHitRatio)} of input)`);
     w(`Cache create tokens: ${formatNum(summary.totalCacheCreation)}`);
-    w(`Peak single call:    ${formatNum(summary.peakSingleCallInput)} input`);
+    w(`Peak single call:    ${formatNum(summary.peakSingleCallInput)} input ` +
+        `(${formatNum(summary.peakSingleCallFresh)} fresh)`);
     w(`Estimated cost:      $${summary.totalCostUsd.toFixed(4)}`);
     // Cache-cadence diagnostic — surface fast-loop prompt-cache latency loss.
     const cadence = computeCacheCadence(summary.events);
@@ -199,8 +205,17 @@ export function printCostForensics(summary, opts = {}) {
     }
     // Acceptance hints — surface anomalies relative to Phase A/B/C targets.
     const anomalies = [];
-    if (summary.peakSingleCallInput > 80_000) {
-        anomalies.push(`peak single-call input ${formatNum(summary.peakSingleCallInput)} > 80,000 — sub-agent cap may not have kicked in (Phase B target breach)`);
+    // Phase B breach = a single call that sent >80K FRESH (uncached) tokens.
+    // Use FRESH, not raw input: `task`/`council` usage rows store `totalUsage`
+    // aggregated across a streamText call's internal steps, so a high raw input
+    // there is normally cached-prefix reuse across steps (high cache hit), NOT a
+    // cold cap breach. Evidence: session 6f17a13c1591 peaked at 378,613 raw input
+    // but only 54,389 fresh (324,224 cache-read) — aggregate reuse, no breach.
+    if (summary.peakSingleCallFresh > 80_000) {
+        anomalies.push(`peak single-call FRESH input ${formatNum(summary.peakSingleCallFresh)} > 80,000 — sub-agent cap may not have kicked in (Phase B target breach)`);
+    }
+    else if (summary.peakSingleCallInput > 80_000) {
+        anomalies.push(`peak single-call input ${formatNum(summary.peakSingleCallInput)} > 80,000 but only ${formatNum(summary.peakSingleCallFresh)} FRESH — aggregate multi-step reuse (task/council totalUsage), NOT a cap breach`);
     }
     if (summary.events.some((e) => e.messageSeq === null && e.source === "message")) {
         anomalies.push(`some 'message' events have NULL message_seq — Phase A5 message write-ahead bypassed (check persistMessageWriteAhead wiring)`);

package/dist/src/cli/keys-bundle.d.ts

@@ -7,7 +7,7 @@
  * Format (JSON):
  *   { v:1, kdf:"scrypt", salt:hex, iv:hex, tag:hex, ct:base64 }
  *
- * Plaintext payload (JSON): { providers: { deepseek: "sk-...", siliconflow: "sf-..." } }
+ * Plaintext payload (JSON): { providers: { deepseek: "sk-...", xai: "xai-..." } }
  *
  * Crypto:
  *   - scrypt(passphrase, salt, 32) → 256-bit key

package/dist/src/cli/keys-bundle.js

@@ -7,7 +7,7 @@
  * Format (JSON):
  *   { v:1, kdf:"scrypt", salt:hex, iv:hex, tag:hex, ct:base64 }
  *
- * Plaintext payload (JSON): { providers: { deepseek: "sk-...", siliconflow: "sf-..." } }
+ * Plaintext payload (JSON): { providers: { deepseek: "sk-...", xai: "xai-..." } }
  *
  * Crypto:
  *   - scrypt(passphrase, salt, 32) → 256-bit key

package/dist/src/cli/keys.d.ts

@@ -15,8 +15,8 @@
  *   keys delete <provider>        — remove a stored key
  *   keys import-bw [providers]    — pull from Bitwarden vault, store in keychain
  *   keys cleanup-settings         — strip plaintext keys from user-settings.json
- *   keys login <provider>         — OAuth login (openai, google/agy)
- *   keys logout <provider>        — OAuth logout (openai, google/agy)
+ *   keys login <provider>         — OAuth login (openai, xai)
+ *   keys logout <provider>        — OAuth logout (openai, xai)
  */
 import { type ChatSecretId } from "../chat/chat-keychain.js";
 export interface KeysSetOptions {

package/dist/src/cli/keys.js

@@ -15,8 +15,8 @@
  *   keys delete <provider>        — remove a stored key
  *   keys import-bw [providers]    — pull from Bitwarden vault, store in keychain
  *   keys cleanup-settings         — strip plaintext keys from user-settings.json
- *   keys login <provider>         — OAuth login (openai, google/agy)
- *   keys logout <provider>        — OAuth logout (openai, google/agy)
+ *   keys login <provider>         — OAuth login (openai, xai)
+ *   keys logout <provider>        — OAuth logout (openai, xai)
  */
 import { spawnSync } from "node:child_process";
 import { promises as fs } from "node:fs";
@@ -43,7 +43,7 @@ const CHAT_SECRET_IDS = ["discord-token", "discord-guild-id", "slack-token", "sl
 function isChatSecretId(value) {
     return CHAT_SECRET_IDS.includes(value);
 }
-const SETTINGS_PATH = path.join(os.homedir(), ".muonroi-cli", "user-settings.json");
+const getSettingsPath = () => path.join(os.homedir(), ".muonroi-cli", "user-settings.json");
 function maskKey(key) {
     if (key.length <= 10)
         return "***";
@@ -51,8 +51,6 @@ function maskKey(key) {
 }
 function normalizeProvider(p) {
     const lower = p.toLowerCase();
-    if (lower === "agy")
-        return "google"; // agy is alias for google (OAuth + key)
     if (KEYCHAIN_PROVIDER_IDS.includes(lower)) {
         return lower;
     }
@@ -115,7 +113,7 @@ async function promptHidden(question) {
 export async function runKeysSet(provider, options = {}) {
     const norm = normalizeProvider(provider);
     if (!norm) {
-        console.error(`Unknown provider '${provider}'. Valid: ${KEYCHAIN_PROVIDER_IDS.join(", ")} (agy is alias for google)`);
+        console.error(`Unknown provider '${provider}'. Valid: ${KEYCHAIN_PROVIDER_IDS.join(", ")}`);
         process.exit(1);
     }
     const key = (await promptHidden(`Paste ${provider} API key (hidden): `)).trim();
@@ -199,30 +197,11 @@ export async function runKeysList() {
     const oauthIds = await getOAuthProviderIds();
     for (const p of oauthIds) {
         try {
-            let tokens = await loadTokens(p);
-            if (!tokens && p === "google") {
-                // Import from agy storage only for fresh (non-expired) tokens.
-                // Stale agy tokens should be cleaned by running `keys logout google` + removing old .gemini file.
-                const { loadAgyTokensWithRefresh } = await import("../providers/auth/gemini-oauth.js");
-                const agyTokens = await loadAgyTokensWithRefresh();
-                if (agyTokens && Date.now() < agyTokens.expiresAt) {
-                    tokens = agyTokens;
-                }
-            }
-            if (tokens) {
-                if (Date.now() < tokens.expiresAt) {
-                    const cfg = await getOAuthProviderConfig(p);
-                    const display = cfg?.displayName ?? p;
-                    oauthRows.push({ provider: p, display, email: tokens.email, expiresAt: tokens.expiresAt });
-                }
-                else if (p === "google") {
-                    // auto-clean stale Agy token from our store (imported from old ~/.gemini)
-                    try {
-                        const { deleteTokens } = await import("../providers/auth/token-store.js");
-                        await deleteTokens("google");
-                    }
-                    catch { }
-                }
+            const tokens = await loadTokens(p);
+            if (tokens && Date.now() < tokens.expiresAt) {
+                const cfg = await getOAuthProviderConfig(p);
+                const display = cfg?.displayName ?? p;
+                oauthRows.push({ provider: p, display, email: tokens.email, expiresAt: tokens.expiresAt });
             }
         }
         catch {
@@ -234,7 +213,6 @@ export async function runKeysList() {
         console.log("No keys stored in OS keychain.");
         console.log("Run 'muonroi-cli keys set <provider>' or 'muonroi-cli keys import-bw' to add some.");
         console.log("Run 'muonroi-cli keys login openai' to log in with your OpenAI subscription.");
-        console.log("Run 'muonroi-cli keys login google' or 'keys login agy' to log in with Agy (Google OAuth).");
         console.log("Run 'muonroi-cli keys login xai' to log in with your SuperGrok / X Premium+ subscription.");
         return;
     }
@@ -288,35 +266,11 @@ export async function runKeysList() {
 export async function runKeysDelete(provider) {
     const norm = normalizeProvider(provider);
     if (!norm) {
-        console.error(`Unknown provider '${provider}'. Valid: ${KEYCHAIN_PROVIDER_IDS.join(", ")} (agy is alias for google)`);
+        console.error(`Unknown provider '${provider}'. Valid: ${KEYCHAIN_PROVIDER_IDS.join(", ")}`);
         process.exit(1);
     }
     const ok = await deleteKeyForProvider(norm);
     console.log(ok ? `Deleted ${provider} key from keychain.` : `No ${provider} key was stored.`);
-    // If deleting google/agy, also clean the OAuth token (Agy uses the google slot for its OAuth)
-    if (norm === "google") {
-        try {
-            const { loadTokens, deleteTokens } = await import("../providers/auth/token-store.js");
-            const { getOAuthProviderConfig } = await import("../providers/auth/registry.js");
-            const tokens = await loadTokens("google");
-            if (tokens) {
-                const cfg = await getOAuthProviderConfig("google");
-                if (cfg?.provider) {
-                    try {
-                        await cfg.provider.revoke(tokens);
-                    }
-                    catch {
-                        // best effort
-                    }
-                }
-                await deleteTokens("google");
-                console.log(`Also cleared Agy/Google OAuth token.`);
-            }
-        }
-        catch {
-            // ignore secondary errors
-        }
-    }
 }
 export async function runKeysExport(filePath) {
     const providers = {};
@@ -679,9 +633,9 @@ export async function runChatImportBw(opts = {}) {
  */
 export async function runKeysLogin(provider) {
     const oauthIds = await getOAuthProviderIds();
-    const norm = provider.toLowerCase() === "agy" ? "google" : provider;
+    const norm = provider;
     if (!oauthIds.includes(norm)) {
-        const supported = oauthIds.map((id) => (id === "google" ? "google (or agy)" : id)).join(", ");
+        const supported = oauthIds.join(", ");
         console.error(`OAuth login not supported for '${provider}'. Supported: ${supported}`);
         process.exit(1);
     }
@@ -693,22 +647,6 @@ export async function runKeysLogin(provider) {
         process.exit(1);
     }
     const name = cfg.displayName;
-    // Special handling for google: prefer importing/reusing token from agy local config
-    // (~/.gemini/oauth_creds.json) so existing agy login "just works" without browser.
-    if (provider === "google") {
-        const { loadAgyTokensWithRefresh } = await import("../providers/auth/gemini-oauth.js");
-        const existing = await loadAgyTokensWithRefresh();
-        if (existing?.accessToken) {
-            console.log(`Found existing agy token for ${existing.email || "(no email)"}.`);
-            console.log("Reusing it (refreshed if near expiry). No browser needed.");
-            const expiry = new Date(existing.expiresAt).toLocaleString();
-            const isExpired = Date.now() > existing.expiresAt;
-            console.log(`Token expires: ${expiry}${isExpired ? " (was expired, refresh attempted)" : ""}`);
-            console.log("Run 'muonroi-cli keys list' to verify.");
-            return;
-        }
-        // No existing agy token — fall through to normal browser flow (using agy client)
-    }
     console.log(`Logging in to ${name} via OAuth...`);
     console.log("A browser window will open. Complete the sign-in to continue.");
     const tokens = await cfg.provider.login({
@@ -742,9 +680,9 @@ export async function runKeysLogin(provider) {
  */
 export async function runKeysLogout(provider) {
     const oauthIds = await getOAuthProviderIds();
-    const norm = provider.toLowerCase() === "agy" ? "google" : provider;
+    const norm = provider;
     if (!oauthIds.includes(norm)) {
-        const supported = oauthIds.map((id) => (id === "google" ? "google (or agy)" : id)).join(", ");
+        const supported = oauthIds.join(", ");
         console.error(`OAuth logout not supported for '${provider}'. Supported: ${supported}`);
         process.exit(1);
     }
@@ -766,7 +704,7 @@ export async function runKeysLogout(provider) {
 export async function runKeysCleanupSettings() {
     let raw;
     try {
-        raw = await fs.readFile(SETTINGS_PATH, "utf8");
+        raw = await fs.readFile(getSettingsPath(), "utf8");
     }
     catch (e) {
         if (e.code === "ENOENT") {
@@ -780,11 +718,11 @@ export async function runKeysCleanupSettings() {
         json = JSON.parse(raw);
     }
     catch {
-        console.error(`Settings file is not valid JSON: ${SETTINGS_PATH}`);
+        console.error(`Settings file is not valid JSON: ${getSettingsPath()}`);
         process.exit(1);
     }
     const ts = new Date().toISOString().replace(/[:.]/g, "-");
-    const backup = `${SETTINGS_PATH}.bak.${ts}`;
+    const backup = `${getSettingsPath()}.bak.${ts}`;
     await fs.writeFile(backup, raw, "utf8");
     let removed = 0;
     if ("apiKey" in json) {
@@ -806,9 +744,9 @@ export async function runKeysCleanupSettings() {
             delete json.providers;
         }
     }
-    await fs.writeFile(SETTINGS_PATH, `${JSON.stringify(json, null, 2)}\n`, "utf8");
+    await fs.writeFile(getSettingsPath(), `${JSON.stringify(json, null, 2)}\n`, "utf8");
     console.log(`Backed up to: ${backup}`);
-    console.log(`Removed ${removed} plaintext key field(s) from: ${SETTINGS_PATH}`);
+    console.log(`Removed ${removed} plaintext key field(s) from: ${getSettingsPath()}`);
     if (removed === 0) {
         console.log("(File was already clean.)");
     }

package/dist/src/council/clarifier.d.ts

@@ -69,6 +69,26 @@ export interface ClarifyOptionsResult {
  * so the UI knows to suppress the "(Recommended)" tag.
  */
 export declare function buildClarifyOptions(suggestions: string[] | undefined, recommended?: string): ClarifyOptionsResult;
+/**
+ * Model-designed option object: a choice the user picks between, carrying its
+ * own explanation. `recommended:true` marks the pre-selected default. This is
+ * the richer shape the clarifier prompt now asks for so every choice shows a
+ * per-option "why" (not just the question-level `why`).
+ */
+export interface ClarifyOptionSpec {
+    label: string;
+    description?: string;
+    recommended?: boolean;
+}
+/**
+ * Build council options from the model's richer `{label, description,
+ * recommended}` objects, preserving each choice's explanation as `description`
+ * and pointing `defaultIndex` at the option the model flagged `recommended`
+ * (falling back to the first when none is flagged). The same "Type something" /
+ * "Chat about this" escape-hatches are appended as in `buildClarifyOptions`.
+ * Pure + exported for unit testing.
+ */
+export declare function buildClarifyOptionsRich(specs: ClarifyOptionSpec[] | undefined): ClarifyOptionsResult;
 export declare function runClarification(topic: string, leaderModelId: string, conversationContext: string, respondToQuestion: QuestionResponder, llm: CouncilLLM, signal?: AbortSignal, seedQuestions?: GrayAreaQuestion[], maxRounds?: number, 
 /**
  * Optional pre-filled answers keyed by seed-question id. When the loop's
@@ -83,5 +103,11 @@ prefillAnswers?: Map<string, string>,
  * a cheaper tier model on the same provider as the leader. Falls back to
  * the leader model if no cheaper model is cataloged. Default false.
  */
-costAware?: boolean): AsyncGenerator<StreamChunk, ClarifiedSpec, unknown>;
-export declare function buildSpecFromTopic(topic: string, _conversationContext: string): ClarifiedSpec;
+costAware?: boolean, 
+/**
+ * Healthy panel models to fall back to when the leader's cost-tier spec-synth
+ * model is on a failing proxy — keeps the spec's ≥3 observable criteria (and
+ * thus a meaningful leader eval) instead of degrading to one generic criterion.
+ */
+fallbackModels?: string[]): AsyncGenerator<StreamChunk, ClarifiedSpec, unknown>;
+export declare function buildSpecFromTopic(topic: string, conversationContext: string): ClarifiedSpec;

package/dist/src/council/clarifier.js

@@ -1,5 +1,5 @@
 import { pickCouncilTaskModel } from "./leader.js";
-import { tracedGenerate } from "./llm.js";
+import { tracedGenerate, tracedGenerateWithFallback } from "./llm.js";
 import { phaseDone, phaseError, phaseStart } from "./phase-events.js";
 import { buildClarificationPrompt, buildReadinessJudgePrompt, buildSpecSynthesisPrompt } from "./prompts.js";
 /** P5: Hard cap on clarification rounds regardless of judge verdict. */
@@ -138,6 +138,45 @@ export function buildClarifyOptions(suggestions, recommended) {
     }
     return { options, defaultIndex };
 }
+/**
+ * Build council options from the model's richer `{label, description,
+ * recommended}` objects, preserving each choice's explanation as `description`
+ * and pointing `defaultIndex` at the option the model flagged `recommended`
+ * (falling back to the first when none is flagged). The same "Type something" /
+ * "Chat about this" escape-hatches are appended as in `buildClarifyOptions`.
+ * Pure + exported for unit testing.
+ */
+export function buildClarifyOptionsRich(specs) {
+    const cleaned = (specs ?? []).filter((s) => !!s && typeof s.label === "string" && s.label.trim().length > 0);
+    const choices = cleaned.map((s) => ({
+        label: s.label.trim(),
+        value: s.label.trim(),
+        kind: "choice",
+        ...(typeof s.description === "string" && s.description.trim().length > 0
+            ? { description: s.description.trim() }
+            : {}),
+    }));
+    const options = [
+        ...choices,
+        {
+            label: "Type something",
+            description: "Nhập câu trả lời tự do",
+            value: "",
+            kind: "freetext",
+        },
+        {
+            label: "Chat about this",
+            description: "Thảo luận thêm trước khi trả lời",
+            value: "",
+            kind: "chat",
+        },
+    ];
+    let defaultIndex;
+    const recIdx = cleaned.findIndex((s) => s.recommended === true);
+    if (recIdx !== -1)
+        defaultIndex = recIdx;
+    return { options, defaultIndex };
+}
 export async function* runClarification(topic, leaderModelId, conversationContext, respondToQuestion, llm, signal, seedQuestions, maxRounds, 
 /**
  * Optional pre-filled answers keyed by seed-question id. When the loop's
@@ -152,7 +191,13 @@ prefillAnswers,
  * a cheaper tier model on the same provider as the leader. Falls back to
  * the leader model if no cheaper model is cataloged. Default false.
  */
-costAware = false) {
+costAware = false, 
+/**
+ * Healthy panel models to fall back to when the leader's cost-tier spec-synth
+ * model is on a failing proxy — keeps the spec's ≥3 observable criteria (and
+ * thus a meaningful leader eval) instead of degrading to one generic criterion.
+ */
+fallbackModels = []) {
     // P5: use MAX_CLARIFY_ROUNDS (12) as the hard cap; respect explicit override
     // from callers that pass maxRounds (e.g. tests that want old 3-round behavior).
     const max = typeof maxRounds === "number" && maxRounds > 0 ? maxRounds : MAX_CLARIFY_ROUNDS;
@@ -284,7 +329,16 @@ costAware = false) {
                 continue;
             }
             const questionId = crypto.randomUUID();
-            const { options, defaultIndex } = buildClarifyOptions(q.suggestions, q.recommended);
+            // Prefer the model's richer {label, description, recommended} options so
+            // every choice shows its own explanation + the recommended default is
+            // pre-selected. Fall back to the legacy string `suggestions` shape (PIL
+            // gray-area seeds and any model still emitting the old format).
+            const { options, defaultIndex } = q.options && q.options.length > 0
+                ? buildClarifyOptionsRich(q.options)
+                : buildClarifyOptions(q.suggestions, q.recommended);
+            // Keep the deprecated `suggestions` mirror populated for consumers that
+            // still read it (audit/replay), deriving labels from rich options.
+            const suggestionLabels = q.suggestions ?? (q.options ? q.options.map((o) => o.label).filter((l) => typeof l === "string") : undefined);
             yield {
                 type: "council_question",
                 content: `**${q.question}**\n${q.why ? `> ${q.why}` : ""}`,
@@ -293,7 +347,7 @@ costAware = false) {
                     phase: "clarify",
                     question: q.question,
                     context: q.why,
-                    suggestions: q.suggestions,
+                    suggestions: suggestionLabels,
                     options,
                     isRequired: q.isRequired,
                     defaultIndex,
@@ -329,7 +383,7 @@ costAware = false) {
         startedAt: phaseStartedAt,
         detail: allQA.length > 0 ? `${allQA.length} Q&A` : "no clarification needed",
     });
-    const spec = yield* synthesizeSpec(topic, conversationContext, allQA, leaderModelId, llm, costAware);
+    const spec = yield* synthesizeSpec(topic, conversationContext, allQA, leaderModelId, llm, costAware, fallbackModels);
     // P5: attach ready-gate metadata to the returned spec
     spec.confidenceScore = gateConfidence;
     spec.remainingGaps = gateGaps;
@@ -355,13 +409,18 @@ costAware = false) {
     };
     return spec;
 }
-export function buildSpecFromTopic(topic, _conversationContext) {
+export function buildSpecFromTopic(topic, conversationContext) {
     return {
         problemStatement: topic,
         constraints: [],
         successCriteria: [`Address the topic: ${topic.slice(0, 100)}`],
         scope: "Determined by conversation context",
         rawQA: [],
+        // Carry the session/task context onto the spec so every debate stage (not
+        // just the opening) can stay anchored to the parent task. Previously this
+        // argument was dropped (prefixed `_`), which is why auto-council debates saw
+        // only the isolated sub-task and drifted off-topic.
+        parentContext: conversationContext?.trim() || undefined,
     };
 }
 /**
@@ -374,7 +433,7 @@ export function buildSpecFromTopic(topic, _conversationContext) {
  * We LLM-extract criteria/constraints/scope from the topic when QA is empty.
  * Falls back to the original buildSpecFromTopic if the LLM call fails.
  */
-async function* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId, llm, costAware) {
+async function* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId, llm, costAware, fallbackModels = []) {
     const system = `You are extracting an implicit specification from a short topic statement. ` +
         `The user did not answer clarification questions, but the topic itself often ` +
         `implies scope, criteria, and constraints. Tease them out.\n\n` +
@@ -396,16 +455,21 @@ async function* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId
         `- Write all fields in the user's language (detected from the topic).`;
     const prompt = `## Topic\n${topic}\n\n${conversationContext ? `## Context\n${conversationContext}\n` : ""}`;
     try {
+        // Model-fallback: the cost-tier synth model can sit on a flaky proxy
+        // (Console Go glm/kimi → "Upstream request failed"). Falling back to
+        // buildSpecFromTopic yields a single generic criterion, which makes the
+        // leader eval read a vague "1/1 criteria met". Retry on healthy panel models
+        // first so the spec keeps its ≥3 observable criteria.
         const synthModel = pickCouncilTaskModel("spec_synthesis", leaderModelId, costAware);
-        const raw = yield* tracedGenerate(llm, {
+        const raw = yield* tracedGenerateWithFallback(llm, {
             phase: "synthesis",
             label: "Inferring spec from topic (no clarification answers)",
-            modelId: synthModel,
+            models: [synthModel, ...fallbackModels],
             system,
             prompt,
             maxTokens: 1024,
         });
-        const match = raw.match(/\{[\s\S]*\}/);
+        const match = raw?.match(/\{[\s\S]*\}/);
         if (match) {
             const parsed = JSON.parse(match[0]);
             const criteria = Array.isArray(parsed.successCriteria) && parsed.successCriteria.length >= 1
@@ -429,12 +493,12 @@ async function* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId
     }
     return buildSpecFromTopic(topic, conversationContext);
 }
-async function* synthesizeSpec(topic, conversationContext, qa, leaderModelId, llm, costAware = false) {
+async function* synthesizeSpec(topic, conversationContext, qa, leaderModelId, llm, costAware = false, fallbackModels = []) {
     if (qa.length === 0) {
         // Clarifier asked 0 questions OR user skipped all of them. Use the LLM to
         // pull implicit criteria/constraints/scope from the topic alone instead of
         // returning the single-criterion fallback that makes leader-eval trivial.
-        return yield* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId, llm, costAware);
+        return yield* inferSpecFromTopicOnly(topic, conversationContext, leaderModelId, llm, costAware, fallbackModels);
     }
     const { system, prompt } = buildSpecSynthesisPrompt(topic, conversationContext, qa);
     const resolved = {};
@@ -444,16 +508,18 @@ async function* synthesizeSpec(topic, conversationContext, qa, leaderModelId, ll
         }
     }
     try {
+        // Model-fallback (see inferSpecFromTopicOnly): keep the ≥3-criteria spec even
+        // when the leader's cost-tier synth model is on a failing proxy.
         const synthModel = pickCouncilTaskModel("spec_synthesis", leaderModelId, costAware);
-        const raw = yield* tracedGenerate(llm, {
+        const raw = yield* tracedGenerateWithFallback(llm, {
             phase: "synthesis",
             label: "Synthesizing clarified spec",
-            modelId: synthModel,
+            models: [synthModel, ...fallbackModels],
             system,
             prompt,
             maxTokens: 2048,
         });
-        const match = raw.match(/\{[\s\S]*\}/);
+        const match = raw?.match(/\{[\s\S]*\}/);
         if (match) {
             const parsed = JSON.parse(match[0]);
             return {

package/dist/src/council/context.js

@@ -98,7 +98,7 @@ function getCompactionText(msg) {
         return null;
     return msg.content;
 }
-function extractUserText(content) {
+function extractText(content) {
     if (typeof content === "string")
         return content;
     if (Array.isArray(content)) {
@@ -109,40 +109,74 @@ function extractUserText(content) {
     }
     return "";
 }
+/** Filter assistant messages to exclude tool-call-only turns (tool_use + tool_result
+ *  pairs that carry no analytical text). An assistant turn with ≥80 chars of text
+ *  content is considered "analytical" (contains decisions, analysis, code, or plans). */
+function isAnalyticalAssistant(content) {
+    const text = extractText(content);
+    return text.trim().length >= 80;
+}
+/** Maximum total characters for the built context string. Raised from 12K to 32K
+ *  (~8K tokens) so council participants get meaningful session history rather than
+ *  only 5 user-message snippets. */
+const MAX_CONTEXT_CHARS = 32_000;
 export function buildCouncilContext(messages) {
     const parts = [];
+    // 1. Compaction summary (first message if it's a system compaction)
     if (messages.length > 0 && isCompactionSummary(messages[0])) {
         const summary = getCompactionText(messages[0]);
         if (summary) {
             parts.push(`## Session Context (from compaction summary)\n${summary}`);
         }
     }
-    const userMessages = [];
-    for (let i = messages.length - 1; i >= 0 && userMessages.length < 5; i--) {
+    // 2. Recent conversation turns — scan BOTH user + analytical assistant messages
+    //    from the tail, capturing at most 8 meaningful entries.
+    //    Previously only captured 5 user messages — lost ALL assistant analysis,
+    //    code changes, and decisions made during the session.
+    const conversationTurns = [];
+    for (let i = messages.length - 1; i >= 0 && conversationTurns.length < 8; i--) {
         const msg = messages[i];
         if (msg.role === "user") {
-            const text = extractUserText(msg.content);
+            const text = extractText(msg.content);
+            if (text.trim()) {
+                conversationTurns.unshift(`[User] ${text.slice(0, 3000).trim()}`);
+            }
+        }
+        else if (msg.role === "assistant" && isAnalyticalAssistant(msg.content)) {
+            const text = extractText(msg.content);
             if (text.trim()) {
-                userMessages.unshift(`- ${text.slice(0, 2000).trim()}`);
+                // Cap assistant text to 1500 chars — enough to convey analysis/decisions
+                // without blowing the context budget on long code outputs.
+                conversationTurns.unshift(`[Assistant] ${text.slice(0, 1500).trim()}`);
             }
         }
     }
-    if (userMessages.length > 0) {
-        parts.push(`## Recent User Messages\n${userMessages.join("\n")}`);
+    if (conversationTurns.length > 0) {
+        parts.push(`## Recent Conversation\n${conversationTurns.join("\n\n")}`);
     }
-    const councilMemories = [];
+    // 3. Key decisions from system messages (task outcomes, decisions, plans)
+    const decisionMsgs = [];
     for (const msg of messages) {
-        if (msg.role === "system" && typeof msg.content === "string" && msg.content.includes("[Council Memory]")) {
-            councilMemories.push(msg.content);
+        if (msg.role !== "system" || typeof msg.content !== "string")
+            continue;
+        const c = msg.content;
+        if (c.includes("[Council Memory]") ||
+            c.includes("[Council Decision]") ||
+            c.includes("[Council Action Items]") ||
+            c.includes("[Council Plan Update]") ||
+            c.includes("[Decision]") ||
+            c.includes("[Task Completed]") ||
+            c.includes("[NEEDS HUMAN REVIEW]")) {
+            decisionMsgs.push(c);
         }
     }
-    if (councilMemories.length > 0) {
-        const digests = councilMemories.slice(-2).map(formatCouncilMemoryDigest);
-        parts.push(`## Previous Council Outcomes (cite by role/round)\n${digests.join("\n\n")}`);
+    if (decisionMsgs.length > 0) {
+        const digests = decisionMsgs.slice(-4).map(formatCouncilMemoryDigest);
+        parts.push(`## Key Decisions\n${digests.join("\n\n")}`);
     }
     const combined = parts.join("\n\n---\n\n");
-    if (combined.length > 12000) {
-        return `${combined.slice(0, 12000)}\n\n[... context truncated to fit token budget]`;
+    if (combined.length > MAX_CONTEXT_CHARS) {
+        return `${combined.slice(0, MAX_CONTEXT_CHARS)}\n\n[... context truncated to fit token budget]`;
     }
     return combined;
 }