gsd-pi 2.78.0 → 2.78.1-dev.2d3fd6e71

package/dist/resources/extensions/gsd/metrics.js

@@ -1,3 +1,4 @@
+// GSD-2 + metrics.ts: token & cost tracking for auto-mode units
 /**
  * GSD Metrics — Token & Cost Tracking
  *
@@ -13,12 +14,14 @@
  *   4. On crash recovery or fresh start, the ledger is loaded from disk
  */
 import { join } from "node:path";
+import { openSync, closeSync, unlinkSync, statSync, writeFileSync } from "node:fs";
 import { gsdRoot } from "./paths.js";
 import { getAndClearSkills } from "./skill-telemetry.js";
 import { loadJsonFile, loadJsonFileOrNull, saveJsonFile } from "./json-persistence.js";
 import { parseUnitId } from "./unit-id.js";
 import { buildAuditEnvelope, emitUokAuditEvent } from "./uok/audit.js";
 import { isUnifiedAuditEnabled } from "./uok/audit-toggle.js";
+import { logWarning } from "./workflow-logger.js";
 // Re-export from shared — import directly from format-utils to avoid pulling
 // in the full barrel (mod.js → ui.js → @gsd/pi-tui) which breaks when loaded
 // outside jiti's alias resolution (e.g. dynamic import in auto-loop reports).
@@ -48,10 +51,15 @@ export function classifyUnitPhase(unitType) {
 // ─── In-memory state ──────────────────────────────────────────────────────────
 let ledger = null;
 let basePath = "";
+// Per-workspace ledger map, keyed by workspace.identityKey.
+// Populated by initMetricsByScope; independent of the module singleton.
+const scopedLedgers = new Map();
 // ─── Public API ───────────────────────────────────────────────────────────────
 /**
  * Initialize the metrics system for a given project.
  * Loads existing ledger from disk if present.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use initMetricsByScope instead.
  */
 export function initMetrics(base) {
     basePath = base;
@@ -59,6 +67,8 @@ export function initMetrics(base) {
 }
 /**
  * Reset in-memory state. Called when auto-mode stops.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use resetMetricsByScope instead.
  */
 export function resetMetrics() {
     ledger = null;
@@ -67,6 +77,8 @@ export function resetMetrics() {
 /**
  * Snapshot usage metrics from the current session before it's wiped.
  * Scans session entries for AssistantMessage usage data.
+ *
+ * @deprecated TODO(C-future): remove module singleton. Use snapshotUnitMetricsByScope instead.
  */
 export function snapshotUnitMetrics(ctx, unitType, unitId, startedAt, model, opts) {
     if (!ledger)
@@ -180,6 +192,147 @@ export function snapshotUnitMetrics(ctx, unitType, unitId, startedAt, model, opt
 export function getLedger() {
     return ledger;
 }
+// ─── Scope-aware API (canonical) ─────────────────────────────────────────────
+/**
+ * Initialize the metrics system for a given workspace scope.
+ * Loads existing ledger from disk into the per-scope ledger map.
+ * Does NOT touch the module-level singleton.
+ */
+export function initMetricsByScope(scope) {
+    const base = scope.workspace.projectRoot;
+    const loaded = loadLedger(base);
+    scopedLedgers.set(scope.workspace.identityKey, loaded);
+}
+/**
+ * Get the in-memory ledger for the given scope, or null if not initialized.
+ */
+export function getLedgerByScope(scope) {
+    return scopedLedgers.get(scope.workspace.identityKey) ?? null;
+}
+/**
+ * Reset scoped in-memory state for a workspace. Called when auto-mode stops.
+ */
+export function resetMetricsByScope(scope) {
+    scopedLedgers.delete(scope.workspace.identityKey);
+}
+/**
+ * Snapshot usage metrics using an explicit workspace scope.
+ *
+ * This is the canonical variant. It derives the metrics path from
+ * scope.workspace.projectRoot rather than the module singleton, so it
+ * remains correct across session resume and in multi-workspace processes.
+ *
+ * Preserves the atomic write-merge logic from saveLedger so concurrent
+ * workers cannot silently discard each other's entries.
+ *
+ * If initMetricsByScope has not been called, the ledger is loaded from
+ * disk on first call (lazy init).
+ */
+export function snapshotUnitMetricsByScope(scope, ctx, unitType, unitId, startedAt, model, opts) {
+    const base = scope.workspace.projectRoot;
+    const key = scope.workspace.identityKey;
+    // Lazy init: load from disk if not yet in scoped map.
+    if (!scopedLedgers.has(key)) {
+        scopedLedgers.set(key, loadLedger(base));
+    }
+    const scopedLedger = scopedLedgers.get(key);
+    const entries = ctx.sessionManager.getEntries();
+    if (!entries || entries.length === 0)
+        return null;
+    const tokens = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 };
+    let cost = 0;
+    let toolCalls = 0;
+    let assistantMessages = 0;
+    let userMessages = 0;
+    for (const entry of entries) {
+        if (entry.type !== "message")
+            continue;
+        const msg = entry.message;
+        if (!msg)
+            continue;
+        if (msg.role === "assistant") {
+            assistantMessages++;
+            if (msg.usage) {
+                tokens.input += msg.usage.input ?? 0;
+                tokens.output += msg.usage.output ?? 0;
+                tokens.cacheRead += msg.usage.cacheRead ?? 0;
+                tokens.cacheWrite += msg.usage.cacheWrite ?? 0;
+                tokens.total += msg.usage.totalTokens ?? 0;
+                if (msg.usage.cost != null) {
+                    const c = msg.usage.cost;
+                    cost += typeof c === "number" ? c : (c.total ?? 0);
+                }
+            }
+            if (msg.content && Array.isArray(msg.content)) {
+                for (const block of msg.content) {
+                    if (block.type === "toolCall")
+                        toolCalls++;
+                }
+            }
+        }
+        else if (msg.role === "user") {
+            userMessages++;
+        }
+    }
+    const unit = {
+        type: unitType,
+        id: unitId,
+        model,
+        startedAt,
+        finishedAt: Date.now(),
+        ...(opts?.autoSessionKey ? { autoSessionKey: opts.autoSessionKey } : {}),
+        tokens,
+        cost,
+        toolCalls,
+        assistantMessages,
+        userMessages,
+        apiRequests: assistantMessages,
+        ...(opts?.tier ? { tier: opts.tier } : {}),
+        ...(opts?.modelDowngraded !== undefined ? { modelDowngraded: opts.modelDowngraded } : {}),
+        ...(opts?.contextWindowTokens !== undefined ? { contextWindowTokens: opts.contextWindowTokens } : {}),
+        ...(opts?.truncationSections !== undefined ? { truncationSections: opts.truncationSections } : {}),
+        ...(opts?.continueHereFired !== undefined ? { continueHereFired: opts.continueHereFired } : {}),
+        ...(opts?.promptCharCount != null ? { promptCharCount: opts.promptCharCount } : {}),
+        ...(opts?.baselineCharCount != null ? { baselineCharCount: opts.baselineCharCount } : {}),
+    };
+    // Auto-capture skill telemetry (#599)
+    const skills = getAndClearSkills();
+    if (skills.length > 0) {
+        unit.skills = skills;
+    }
+    // Compute cache hit rate
+    if (tokens.cacheRead > 0 || tokens.input > 0) {
+        const totalInput = tokens.cacheRead + tokens.input;
+        unit.cacheHitRate = totalInput > 0 ? Math.round((tokens.cacheRead / totalInput) * 100) : 0;
+    }
+    // Idempotency guard: update in-place on duplicate, append otherwise.
+    const dupeIdx = scopedLedger.units.findIndex((u) => u.type === unit.type && u.id === unit.id && u.startedAt === unit.startedAt);
+    if (dupeIdx >= 0) {
+        scopedLedger.units[dupeIdx] = unit;
+    }
+    else {
+        scopedLedger.units.push(unit);
+    }
+    saveLedger(base, scopedLedger);
+    if (isUnifiedAuditEnabled()) {
+        emitUokAuditEvent(base, buildAuditEnvelope({
+            traceId: opts?.traceId ?? `metrics:${unitType}:${unitId}`,
+            turnId: opts?.turnId,
+            causedBy: opts?.causedBy,
+            category: "metrics",
+            type: "unit-metrics-snapshot",
+            payload: {
+                unitType,
+                unitId,
+                model,
+                tokens: unit.tokens,
+                cost: unit.cost,
+                toolCalls: unit.toolCalls,
+            },
+        }));
+    }
+    return unit;
+}
 function emptyTokens() {
     return { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 };
 }
@@ -423,6 +576,12 @@ export function pruneMetricsLedger(base, keepCount) {
     if (ledger) {
         ledger.units = ledger.units.slice(-keepCount);
     }
+    // Invalidate all scoped ledger cache entries. Prune is rare; clearing the
+    // entire map is simpler than tracking which entry belongs to `base`. Without
+    // this, scopedLedgers entries for the pruned workspace hold a pre-prune
+    // MetricsLedger that snapshotUnitMetricsByScope would merge back in, causing
+    // pruned units to reappear in subsequent snapshots.
+    scopedLedgers.clear();
     return removed;
 }
 /**
@@ -461,6 +620,133 @@ function deduplicateUnits(units) {
     }
     return Array.from(map.values());
 }
+// How long a lock file must be untouched (in ms) before it is considered
+// orphaned from a crashed process. Set to 2× the acquire timeout.
+export const STALE_LOCK_THRESHOLD_MS = 4000;
+// Retry interval between lock acquire attempts (ms). Caps syscall rate at
+// ~200 attempts over a 2s timeout instead of ~20,000 without any sleep.
+// Exposed for tests.
+export const LOCK_RETRY_INTERVAL_MS = 5;
+// Sync sleep via Atomics.wait — true OS-level sleep, no CPU spin.
+// Int32Array must reference a SharedArrayBuffer; we wait on index 0 which
+// will never be woken by a Atomics.notify, so the wait always times out.
+const _lockSleepBuf = new Int32Array(new SharedArrayBuffer(4));
+function syncSleep(ms) {
+    Atomics.wait(_lockSleepBuf, 0, 0, ms);
+}
+// Counts the number of sleepy retries (non-stale-evicting) made by acquireLock
+// across all calls since the last reset. Exported for test instrumentation only.
+let _lockSleepyRetries = 0;
+export function getLockSleepyRetries() { return _lockSleepyRetries; }
+export function resetLockSleepyRetries() { _lockSleepyRetries = 0; }
+/**
+ * Acquire an exclusive .lock sentinel file via O_EXCL.
+ *
+ * Improvements over the original:
+ *  - No busy spin: the inner `while (Date.now() < waitUntil) {}` spin that
+ *    burned CPU doing nothing useful is removed. Each retry attempt now makes
+ *    one `openSync` syscall and immediately re-checks the deadline, which is
+ *    orders of magnitude cheaper than a tight spin loop.
+ *  - Stale-lock detection: if the existing lock file's mtime is older than
+ *    STALE_LOCK_THRESHOLD_MS, the lock is considered orphaned (e.g. the
+ *    writing process crashed) and is forcibly removed before retrying.
+ *    A warning is logged so operators can detect crash patterns.
+ *  - PID stamp: on success, writes the acquiring process's PID and a
+ *    timestamp into the lock file so external monitors can identify orphans.
+ *  - Retry sleep: after each non-stale-evicting retry, sleeps
+ *    LOCK_RETRY_INTERVAL_MS (5ms) via Atomics.wait so the process yields to
+ *    the OS. This caps syscall rate at ~200–400/s under contention instead of
+ *    the ~20,000/s that would result from a tight openSync loop.
+ *    After a stale-lock eviction (lock already removed), no sleep is injected
+ *    — we retry immediately to close the short race window.
+ *
+ * Returns true on success, false on timeout.
+ */
+function acquireLock(lockPath, timeoutMs = 2000) {
+    const deadline = Date.now() + timeoutMs;
+    while (Date.now() < deadline) {
+        try {
+            const fd = openSync(lockPath, "wx"); // O_WRONLY | O_CREAT | O_EXCL
+            closeSync(fd);
+            // Write PID stamp so external monitors can identify the lock owner.
+            try {
+                writeFileSync(lockPath, `${process.pid}\n${new Date().toISOString()}\n`, "utf-8");
+            }
+            catch { /* non-fatal — stamp is diagnostic only */ }
+            return true;
+        }
+        catch {
+            // Lock held by another process — check for staleness before retrying.
+            try {
+                const st = statSync(lockPath);
+                if (Date.now() - st.mtimeMs > STALE_LOCK_THRESHOLD_MS) {
+                    logWarning("fs", `stale metrics lock at ${lockPath} (age ${Date.now() - st.mtimeMs}ms); forcibly removing and retrying`);
+                    try {
+                        unlinkSync(lockPath);
+                    }
+                    catch { /* already gone */ }
+                    // Do NOT sleep after stale-lock eviction — retry the open
+                    // immediately. The lock file was just removed; a short race window
+                    // exists and sleeping here would unnecessarily delay recovery.
+                    continue;
+                }
+            }
+            catch { /* lock file disappeared between the failed open and stat — retry */ }
+            // Sleep between retries to yield to the OS and cap syscall rate.
+            // Uses Atomics.wait for a true blocking sleep (no CPU spin).
+            _lockSleepyRetries++;
+            syncSleep(LOCK_RETRY_INTERVAL_MS);
+        }
+    }
+    return false;
+}
+function releaseLock(lockPath) {
+    try {
+        unlinkSync(lockPath);
+    }
+    catch { /* ignore */ }
+}
+/**
+ * Save the ledger with cross-process merge semantics.
+ *
+ * Acquires a .lock sentinel file, reads the current on-disk ledger,
+ * merges worker units with existing peer units (worker's entry wins on
+ * type+id+startedAt conflict since it has the latest finishedAt),
+ * then writes atomically. This prevents parallel auto-mode workers from
+ * silently discarding each other's metrics entries.
+ *
+ * Falls back to a direct write (no merge) if the lock cannot be acquired
+ * within the timeout — better to potentially overwrite than to lose data
+ * entirely.
+ */
 function saveLedger(base, data) {
-    saveJsonFile(metricsPath(base), data);
+    const path = metricsPath(base);
+    const lockPath = `${path}.lock`;
+    const acquired = acquireLock(lockPath);
+    if (acquired) {
+        try {
+            // Read current on-disk state and merge with worker's in-memory units.
+            // Worker units take precedence on conflict (by finishedAt in deduplicateUnits).
+            const onDisk = loadJsonFileOrNull(path, isMetricsLedger);
+            if (onDisk && onDisk.units.length > 0) {
+                const merged = deduplicateUnits([...onDisk.units, ...data.units]);
+                saveJsonFile(path, { ...data, units: merged });
+            }
+            else {
+                saveJsonFile(path, data);
+            }
+        }
+        finally {
+            releaseLock(lockPath);
+        }
+    }
+    else {
+        // Lock could not be acquired within the timeout. Fall back to a direct
+        // write (no cross-process merge) to avoid losing this worker's data
+        // entirely. A concurrent writer may overwrite us, but that is preferable
+        // to a torn write caused by two writers simultaneously executing the
+        // read-merge-write sequence without mutual exclusion.
+        logWarning("fs", "saveLedger: lock not acquired — falling back to direct write (no merge)");
+        saveJsonFile(path, data);
+    }
 }

package/dist/resources/extensions/gsd/migrate/command.js

@@ -14,6 +14,7 @@ import { gsdRoot } from "../paths.js";
 import { fileURLToPath } from "node:url";
 import { showNextAction } from "../../shared/tui.js";
 import { validatePlanningDirectory, parsePlanningDirectory, transformToGSD, generatePreview, writeGSDDirectory, } from "./index.js";
+import { homedir } from "node:os";
 /** Format preview stats for embedding in the review prompt. */
 function formatPreviewStats(preview) {
     const lines = [
@@ -50,10 +51,10 @@ export async function handleMigrate(args, ctx, pi) {
     // Default to cwd when no args given; expand ~ to HOME
     let rawPath = args.trim() || ".";
     if (rawPath.startsWith("~/")) {
-        rawPath = join(process.env.HOME ?? "~", rawPath.slice(2));
+        rawPath = join(homedir(), rawPath.slice(2));
     }
     else if (rawPath === "~") {
-        rawPath = process.env.HOME ?? "~";
+        rawPath = homedir();
     }
     let sourcePath = resolve(process.cwd(), rawPath);
     if (!sourcePath.endsWith(".planning")) {

package/dist/resources/extensions/gsd/milestone-id-reservation.js

@@ -0,0 +1,36 @@
+import { isDbAvailable, getAllMilestones } from "./gsd-db.js";
+import { getReservedMilestoneIds, milestoneIdSort, nextMilestoneId, reserveMilestoneId, } from "./milestone-ids.js";
+import { isReusableGhostMilestone } from "./state.js";
+function getDatabaseMilestoneIds() {
+    if (!isDbAvailable())
+        return [];
+    return getAllMilestones().map((milestone) => milestone.id);
+}
+/**
+ * Generate the next milestone ID, accounting for DB rows and in-process
+ * reservations, and reserve it.
+ */
+export function nextMilestoneIdReserved(existingIds, uniqueEnabled, basePath) {
+    const reservedIds = getReservedMilestoneIds();
+    const allIds = [
+        ...new Set([
+            ...existingIds,
+            ...reservedIds,
+            ...getDatabaseMilestoneIds(),
+        ]),
+    ];
+    if (basePath) {
+        const sorted = [...allIds].sort(milestoneIdSort);
+        for (const candidate of sorted) {
+            if (reservedIds.has(candidate))
+                continue;
+            if (isReusableGhostMilestone(basePath, candidate)) {
+                reserveMilestoneId(candidate);
+                return candidate;
+            }
+        }
+    }
+    const id = nextMilestoneId(allIds, uniqueEnabled);
+    reserveMilestoneId(id);
+    return id;
+}

package/dist/resources/extensions/gsd/milestone-scope-classifier.js

@@ -120,13 +120,14 @@ function mentionsWithoutNegation(text, term) {
         const start = m.index ?? 0;
         const windowStart = Math.max(0, start - 30);
         const window = lower.slice(windowStart, start);
-        // Negator anywhere in the 30-char lookback window counts as negation —
-        // covers "no backend", "without a server", "not using api", "zero
-        // dependencies on an api". If a sentence break intervenes between the
-        // negator and the term, treat as a different clause (positive mention).
-        const hasNegator = /(^|[^a-z0-9])(no|without|not|zero|skip(s|ping)?|drops?)\b/i.test(window);
-        const hasSentenceBreak = /[.;!?]/.test(window);
-        if (hasNegator && !hasSentenceBreak)
+        // Negator anywhere in the same clause counts as negation — covers "no
+        // backend", "without a server", "not using api", "zero dependencies on
+        // an api". Only inspect text after the most recent sentence break so
+        // "No backend. Add API" still treats API as a positive mention.
+        const lastBreak = Math.max(window.lastIndexOf("."), window.lastIndexOf(";"), window.lastIndexOf("!"), window.lastIndexOf("?"));
+        const clauseWindow = window.slice(lastBreak + 1);
+        const hasNegator = /(^|[^a-z0-9])(no|without|not|zero|skip(s|ping)?|drops?)\b/i.test(clauseWindow);
+        if (hasNegator)
             continue;
         return true;
     }
@@ -220,7 +221,8 @@ export function classifyMilestoneScope(input) {
         input.verificationOperational ?? "",
         input.verificationUat ?? "",
     ].join("\n");
-    const overrideHits = containsAnyKeyword(haystack, OVERRIDE_KEYWORDS);
+    const overrideHits = containsAnyKeyword(haystack, OVERRIDE_KEYWORDS)
+        .filter((kw) => mentionsWithoutNegation(haystack, kw));
     const complexHits = containsAnyKeyword(haystack, COMPLEX_KEYWORDS);
     const trivialHits = containsAnyKeyword(haystack, TRIVIAL_KEYWORDS);
     const fileCountHint = extractFileCountHint(haystack);

package/dist/resources/extensions/gsd/model-router.js

@@ -228,11 +228,8 @@ export function getEligibleModels(tier, availableModelIds, routingConfig) {
         if (availableModelIds.includes(explicitModel))
             return [explicitModel];
         // Provider-prefix-stripped match
-        const match = availableModelIds.find(id => {
-            const bareAvail = id.includes("/") ? id.split("/").pop() : id;
-            const bareExplicit = explicitModel.includes("/") ? explicitModel.split("/").pop() : explicitModel;
-            return bareAvail === bareExplicit;
-        });
+        const bareExplicit = bareModelId(explicitModel);
+        const match = availableModelIds.find(id => bareModelId(id) === bareExplicit);
         if (match)
             return [match];
     }
@@ -323,12 +320,31 @@ export function resolveModelForComplexity(classification, phaseConfig, routingCo
     }
     // Downgrade-only: if requested tier >= configured tier, no change
     if (tierOrdinal(requestedTier) >= tierOrdinal(configuredTier)) {
+        // If the configured primary is directly available, use it
+        if (isModelAvailable(configuredPrimary, availableModelIds)) {
+            return {
+                modelId: configuredPrimary,
+                fallbacks: phaseConfig.fallbacks,
+                tier: requestedTier,
+                wasDowngraded: false,
+                reason: `tier ${requestedTier} >= configured ${configuredTier}`,
+                selectionMethod: "tier-only",
+            };
+        }
+        // Configured primary is unavailable (e.g. Anthropic model configured but
+        // running on a non-Anthropic provider). Find the best available model at
+        // the same capability tier so routing still works cross-provider.
+        const crossProviderEquivalent = findModelForTier(configuredTier, routingConfig, availableModelIds, routingConfig.cross_provider !== false);
         return {
-            modelId: configuredPrimary,
-            fallbacks: phaseConfig.fallbacks,
+            modelId: crossProviderEquivalent ?? configuredPrimary,
+            fallbacks: crossProviderEquivalent
+                ? [...phaseConfig.fallbacks.filter(f => f !== crossProviderEquivalent), configuredPrimary]
+                : phaseConfig.fallbacks,
             tier: requestedTier,
             wasDowngraded: false,
-            reason: `tier ${requestedTier} >= configured ${configuredTier}`,
+            reason: crossProviderEquivalent
+                ? `cross-provider ${configuredTier}-tier equivalent`
+                : `tier ${requestedTier} >= configured ${configuredTier}`,
             selectionMethod: "tier-only",
         };
     }
@@ -404,7 +420,84 @@ export function defaultRoutingConfig() {
         hooks: true,
     };
 }
+// ─── Tier-Based Model Resolution (for profile defaults) ─────────────────────
+/**
+ * Fallback-only canonical model IDs per tier. Returned when the
+ * available-model list is empty (e.g., preferences are loaded before the
+ * model registry is populated at bootstrap), or when a non-empty registry has
+ * no model at the requested tier.
+ *
+ * Precedence (resolveModelForTier):
+ *   1. configured `tier_models[tier]` (via getEligibleModels) — exact/bare match
+ *   2. cheapest available model whose tier matches `tier`
+ *   3. CANONICAL_TIER_MODELS[tier] as last-resort fallback
+ */
+const CANONICAL_TIER_MODELS = {
+    light: "claude-haiku-4-5",
+    standard: "claude-sonnet-4-6",
+    heavy: "claude-opus-4-6",
+};
+export function canonicalModelForTier(tier) {
+    return CANONICAL_TIER_MODELS[tier];
+}
+/**
+ * Single source of truth for tier-based model selection.
+ * Returns the cheapest available model whose capability tier matches `tier`,
+ * honoring `routingConfig.tier_models[tier]` when set. Returns undefined when
+ * no available model matches the tier.
+ *
+ * `crossProvider`: when false, restricts the search to models that share the
+ * canonical (Anthropic) provider for the tier. When true, any provider is
+ * eligible.
+ */
+function findModelForTier(tier, routingConfig, availableModelIds, crossProvider) {
+    const eligible = getEligibleModels(tier, availableModelIds, routingConfig);
+    if (eligible.length === 0)
+        return undefined;
+    if (crossProvider) {
+        return eligible[0];
+    }
+    // Same-provider only: keep models whose bare ID matches a canonical
+    // Anthropic ID at this tier (i.e., a claude-* model in the tier map).
+    const sameProvider = eligible.filter(id => {
+        const bare = bareModelId(id);
+        return MODEL_CAPABILITY_TIER[bare] === tier && bare.startsWith("claude-");
+    });
+    return sameProvider[0];
+}
+export function resolveModelForTier(tier, availableModelIds, routingConfigOrCrossProvider = defaultRoutingConfig(), crossProvider) {
+    const routingConfig = typeof routingConfigOrCrossProvider === "boolean"
+        ? defaultRoutingConfig()
+        : { ...defaultRoutingConfig(), ...routingConfigOrCrossProvider };
+    const allowCrossProvider = typeof routingConfigOrCrossProvider === "boolean"
+        ? routingConfigOrCrossProvider
+        : crossProvider ?? routingConfig.cross_provider !== false;
+    // No available models known — return canonical fallback
+    if (availableModelIds.length === 0) {
+        return canonicalModelForTier(tier);
+    }
+    // Cross-provider tier search
+    const resolved = findModelForTier(tier, routingConfig, availableModelIds, allowCrossProvider);
+    return resolved
+        ? normalizeResolvedTierModelId(resolved, tier, routingConfig)
+        : canonicalModelForTier(tier);
+}
 // ─── Internal ────────────────────────────────────────────────────────────────
+/**
+ * Check whether a model ID is present in the available models list.
+ * Handles bare IDs ("claude-opus-4-6") and provider-prefixed IDs ("anthropic/claude-opus-4-6").
+ */
+function isModelAvailable(modelId, availableModelIds) {
+    if (availableModelIds.includes(modelId))
+        return true;
+    // Strip provider prefix for comparison. Treat trailing-slash IDs ("provider/")
+    // as no-bare-ID rather than empty-string match (which would erroneously match
+    // any other "provider/" ID).
+    const bare = bareModelId(modelId);
+    if (!bare)
+        return false;
+    return availableModelIds.some(id => bareModelId(id) === bare);
+}
 function getModelTier(modelId) {
     // Strip provider prefix if present
     const bareId = bareModelId(modelId);
@@ -443,8 +536,20 @@ function getModelCost(modelId) {
     // Unknown cost — assume expensive to avoid routing to unknown cheap models
     return 999;
 }
+function normalizeResolvedTierModelId(modelId, tier, routingConfig) {
+    const explicitModel = routingConfig.tier_models?.[tier];
+    if (explicitModel?.includes("/")) {
+        return modelId;
+    }
+    const bareId = bareModelId(modelId);
+    return MODEL_CAPABILITY_TIER[bareId] ? bareId : modelId;
+}
 function bareModelId(modelId) {
-    return modelId.includes("/") ? modelId.split("/").pop() : modelId;
+    if (!modelId.includes("/"))
+        return modelId;
+    // .pop() never returns undefined on a non-empty string but ?? guards future
+    // refactors and avoids the misleading non-null assertion.
+    return modelId.split("/").pop() ?? modelId;
 }
 // ─── Provider-specific Tool Limits ─────────────────────────────────────────
 /**