omegon 0.6.0

package/extensions/effort/tiers.ts

@@ -0,0 +1,146 @@
+/**
+ * effort/tiers — Pure functions for effort tier configuration.
+ *
+ * No side effects, no imports beyond types. Safe to call from any context.
+ */
+
+import type { EffortConfig, EffortLevel } from "./types.ts";
+import { EFFORT_NAMES } from "./types.ts";
+
+/**
+ * Static tier configuration table.
+ * Each entry is a complete EffortConfig for levels 1-7.
+ *
+ *  1 = Servitor        — all local
+ *  2 = Average    — local driver, local background
+ *  3 = Substantial — victory driver, local background (daily default)
+ *  4 = Ruthless   — victory + medium thinking
+ *  5 = Lethal     — victory + high thinking, gloriana review
+ *  6 = Absolute   — gloriana driver, victory background
+ *  7 = Omnissiah  — all gloriana
+ */
+const TIERS: Record<EffortLevel, EffortConfig> = {
+  1: {
+    level: 1,
+    name: "Servitor",
+    driver: "local",
+    thinking: "off",
+    extraction: "local",
+    compaction: "local",
+    cleavePreferLocal: true,
+    cleaveFloor: "local",
+    reviewModel: "local",
+  },
+  2: {
+    level: 2,
+    name: "Average",
+    driver: "local",
+    thinking: "minimal",
+    extraction: "local",
+    compaction: "local",
+    cleavePreferLocal: false,
+    cleaveFloor: "local",
+    reviewModel: "local",
+  },
+  3: {
+    level: 3,
+    name: "Substantial",
+    driver: "victory",
+    thinking: "low",
+    extraction: "local",
+    compaction: "victory",
+    cleavePreferLocal: false,
+    cleaveFloor: "local",
+    reviewModel: "victory",
+  },
+  4: {
+    level: 4,
+    name: "Ruthless",
+    driver: "victory",
+    thinking: "medium",
+    extraction: "local",
+    compaction: "victory",
+    cleavePreferLocal: false,
+    cleaveFloor: "local",
+    reviewModel: "victory",
+  },
+  5: {
+    level: 5,
+    name: "Lethal",
+    driver: "victory",
+    thinking: "high",
+    extraction: "local",
+    compaction: "victory",
+    cleavePreferLocal: false,
+    cleaveFloor: "victory",
+    reviewModel: "gloriana",
+  },
+  6: {
+    level: 6,
+    name: "Absolute",
+    driver: "gloriana",
+    thinking: "high",
+    extraction: "victory",
+    compaction: "victory",
+    cleavePreferLocal: false,
+    cleaveFloor: "victory",
+    reviewModel: "gloriana",
+  },
+  7: {
+    level: 7,
+    name: "Omnissiah",
+    driver: "gloriana",
+    thinking: "high",
+    extraction: "gloriana",
+    compaction: "gloriana",
+    cleavePreferLocal: false,
+    cleaveFloor: "gloriana",
+    reviewModel: "gloriana",
+  },
+};
+
+/**
+ * Get the EffortConfig for a given numeric level.
+ *
+ * @param level - Effort level 1-7
+ * @returns EffortConfig for the requested level (shared reference — do not mutate)
+ * @throws RangeError if level is outside 1-7
+ */
+export function tierConfig(level: number): EffortConfig {
+  if (level < 1 || level > 7 || !Number.isInteger(level)) {
+    throw new RangeError(`Effort level must be an integer 1-7, got ${level}`);
+  }
+  return TIERS[level as EffortLevel];
+}
+
+/** Reverse lookup: tier name (case-insensitive) → level number. */
+const NAME_TO_LEVEL: ReadonlyMap<string, EffortLevel> = new Map(
+  (Object.entries(EFFORT_NAMES) as [string, string][]).map(([k, v]) => [
+    v.toLowerCase(),
+    Number(k) as EffortLevel,
+  ]),
+);
+
+/**
+ * Parse a tier name string to its numeric level.
+ *
+ * Case-insensitive. Returns `undefined` for unknown names.
+ *
+ * @param name - Tier name like "Ruthless", "low", "OMNISSIAH"
+ * @returns Numeric level 1-7, or undefined if not recognized
+ */
+export function parseTierName(name: string): EffortLevel | undefined {
+  return NAME_TO_LEVEL.get(name.toLowerCase());
+}
+
+/** Default effort level when no env var or config is present. */
+export const DEFAULT_EFFORT_LEVEL: EffortLevel = 3;
+
+/** All valid tier names for display/error messages. */
+export const TIER_NAMES: readonly string[] = Object.values(EFFORT_NAMES);
+
+/** Minimum valid effort level. */
+export const MIN_LEVEL: EffortLevel = 1;
+
+/** Maximum valid effort level. */
+export const MAX_LEVEL: EffortLevel = 7;

package/extensions/effort/types.ts

@@ -0,0 +1,105 @@
+/**
+ * effort/types — Type definitions for the effort-tiers system.
+ *
+ * Effort tiers provide 7-level global inference cost control,
+ * from fully-local (Servitor) to all-gloriana (Omnissiah).
+ */
+
+// ─── Model Tiers ─────────────────────────────────────────────
+
+/**
+ * Model tier for effort configuration — which model class to use.
+ * Includes "retribution" so that policy-upgraded extraction tiers (local→retribution) are
+ * representable without casts. Aligns with the shared ModelTier in model-routing.ts.
+ */
+export type EffortModelTier = "local" | "retribution" | "victory" | "gloriana";
+
+/** Thinking level passed to the driver model. */
+export type ThinkingLevel = "off" | "minimal" | "low" | "medium" | "high";
+
+// ─── Effort Level ────────────────────────────────────────────
+
+/** Numeric effort levels 1-7. */
+export type EffortLevel = 1 | 2 | 3 | 4 | 5 | 6 | 7;
+
+/** Human-readable tier names, indexed by level. */
+export const EFFORT_NAMES = {
+  1: "Servitor",
+  2: "Average",
+  3: "Substantial",
+  4: "Ruthless",
+  5: "Lethal",
+  6: "Absolute",
+  7: "Omnissiah",
+} as const satisfies Record<EffortLevel, string>;
+
+/**
+ * Greek letter glyphs for each effort level — α through ζ, then ω for Omnissiah.
+ * α = origin (Servitor), ω = complete arc (Omnissiah). Display alongside tier names.
+ */
+export const EFFORT_GLYPHS = {
+	1: "α",
+	2: "β",
+	3: "γ",
+	4: "δ",
+	5: "ε",
+	6: "ζ",
+	7: "ω",
+} as const satisfies Record<EffortLevel, string>;
+
+/** Union of valid tier name strings. */
+export type EffortName = (typeof EFFORT_NAMES)[EffortLevel];
+
+// ─── Effort Config ───────────────────────────────────────────
+
+/**
+ * Immutable configuration for a single effort tier.
+ * Returned by `tierConfig(level)` — pure data, no side effects.
+ */
+export interface EffortConfig {
+  /** Numeric level 1-7. */
+  level: EffortLevel;
+  /** Human-readable tier name. */
+  name: EffortName;
+
+  // ── Model selection ──
+  /** Primary driver model. */
+  driver: EffortModelTier;
+  /** Extended thinking budget for the driver. */
+  thinking: ThinkingLevel;
+
+  // ── Background tasks ──
+  /** Model for memory extraction / summarization. */
+  extraction: EffortModelTier;
+  /** Model for context compaction. */
+  compaction: EffortModelTier;
+
+  // ── Cleave dispatch ──
+  /** Whether cleave should prefer local models for child tasks. */
+  cleavePreferLocal: boolean;
+  /** Minimum model tier for cleave child tasks. */
+  cleaveFloor: EffortModelTier;
+
+  // ── Review ──
+  /** Model used for code review in cleave review loops. */
+  reviewModel: EffortModelTier;
+}
+
+// ─── Effort State ────────────────────────────────────────────
+
+/**
+ * Runtime effort state stored in SharedState.
+ * Extends EffortConfig with cap tracking and resolved model IDs.
+ */
+export interface EffortState extends EffortConfig {
+  /** Whether the effort level is capped (ceiling-locked by operator). */
+  capped: boolean;
+  /** If capped, the level at which the cap is set. */
+  capLevel?: EffortLevel;
+  /**
+   * Concrete model ID resolved for extraction work under the current routing policy.
+   * May differ from `extraction` when cheapCloudPreferredOverLocal upgrades local→retribution.
+   * Populated by resolveExtractionTier() on session_start and tier switches.
+   */
+  resolvedExtractionModelId?: string;
+}

package/extensions/lib/git-state.ts

@@ -0,0 +1,227 @@
+/**
+ * lib/git-state — Shared helpers for inspecting git dirty-tree state.
+ *
+ * Pure helpers only: parse porcelain output, separate tracked/untracked files,
+ * classify built-in volatile artifacts, and prepare checkpoint/stash plans
+ * without executing git mutations.
+ */
+
+export const BUILTIN_VOLATILE_ALLOWLIST = [".pi/memory/facts.jsonl"] as const;
+
+export type GitStatusCode =
+	| "modified"
+	| "added"
+	| "deleted"
+	| "renamed"
+	| "copied"
+	| "unmerged"
+	| "untracked"
+	| "ignored"
+	| "unknown";
+
+export interface GitStatusEntry {
+	path: string;
+	code: string;
+	indexStatus: string;
+	worktreeStatus: string;
+	tracked: boolean;
+	untracked: boolean;
+	staged: boolean;
+	unstaged: boolean;
+	deleted: boolean;
+	renamed: boolean;
+	copied: boolean;
+	originalPath?: string;
+	indexKind: GitStatusCode;
+	worktreeKind: GitStatusCode;
+}
+
+export interface GitStateSnapshot {
+	entries: GitStatusEntry[];
+	tracked: GitStatusEntry[];
+	untracked: GitStatusEntry[];
+	volatile: GitStatusEntry[];
+	nonVolatile: GitStatusEntry[];
+}
+
+export interface GitPathsetPlan {
+	paths: string[];
+	includesTracked: boolean;
+	includesUntracked: boolean;
+	pathspec: string[];
+}
+
+export interface GitPathsetInput {
+	tracked?: string[];
+	untracked?: string[];
+}
+
+export interface PreparedStashPlan extends GitPathsetPlan {
+	kind: "stash";
+	label: string;
+	includeUntracked: boolean;
+	command: string[];
+}
+
+export interface PreparedCheckpointPlan extends GitPathsetPlan {
+	kind: "checkpoint";
+	message: string;
+	command: string[];
+	requiresApproval: true;
+}
+
+export function parseGitStatus(output: string): GitStatusEntry[] {
+	return output
+		.split("\n")
+		.map((line) => line.trimEnd())
+		.filter(Boolean)
+		.map(parseGitStatusLine);
+}
+
+export function inspectGitState(
+	output: string,
+	volatileAllowlist: readonly string[] = BUILTIN_VOLATILE_ALLOWLIST,
+): GitStateSnapshot {
+	const entries = parseGitStatus(output);
+	const tracked = entries.filter((entry) => entry.tracked);
+	const untracked = entries.filter((entry) => entry.untracked);
+	const volatile = entries.filter((entry) => isVolatilePath(entry.path, volatileAllowlist));
+	return {
+		entries,
+		tracked,
+		untracked,
+		volatile,
+		nonVolatile: entries.filter((entry) => !isVolatilePath(entry.path, volatileAllowlist)),
+	};
+}
+
+export function buildPathsetFromEntries(entries: GitStatusEntry[]): GitPathsetInput {
+	return {
+		tracked: entries.filter((entry) => entry.tracked).map((entry) => entry.path),
+		untracked: entries.filter((entry) => entry.untracked).map((entry) => entry.path),
+	};
+}
+
+export function isVolatilePath(
+	path: string,
+	volatileAllowlist: readonly string[] = BUILTIN_VOLATILE_ALLOWLIST,
+): boolean {
+	return volatileAllowlist.some((allowed) => pathMatchesAllowlist(path, allowed));
+}
+
+export function prepareStashPlan(input: {
+	paths?: string[];
+	tracked?: string[];
+	untracked?: string[];
+	label: string;
+	includesUntracked?: boolean;
+}): PreparedStashPlan {
+	const pathset = prepareGitPathset({
+		tracked: input.paths ?? input.tracked,
+		untracked: input.untracked,
+	});
+	const includeUntracked = input.includesUntracked ?? pathset.includesUntracked;
+	const command = ["git", "stash", "push", "-m", input.label];
+	if (includeUntracked) command.push("--include-untracked");
+	command.push(...pathset.pathspec);
+	return {
+		kind: "stash",
+		label: input.label,
+		includeUntracked,
+		command,
+		...pathset,
+	};
+}
+
+export function prepareCheckpointPlan(input: {
+	paths?: string[];
+	tracked?: string[];
+	untracked?: string[];
+	message: string;
+}): PreparedCheckpointPlan {
+	const pathset = prepareGitPathset({
+		tracked: input.paths ?? input.tracked,
+		untracked: input.untracked,
+	});
+	return {
+		kind: "checkpoint",
+		message: input.message,
+		requiresApproval: true,
+		command: ["git", "commit", "-m", input.message],
+		...pathset,
+	};
+}
+
+export function prepareGitPathset(input: string[] | GitPathsetInput): GitPathsetPlan {
+	const tracked = Array.isArray(input) ? input : (input.tracked ?? []);
+	const untracked = Array.isArray(input) ? [] : (input.untracked ?? []);
+	const uniquePaths = [...new Set([...tracked, ...untracked].filter(Boolean))].sort();
+	return {
+		paths: uniquePaths,
+		includesTracked: tracked.length > 0,
+		includesUntracked: untracked.length > 0,
+		pathspec: uniquePaths.length > 0 ? ["--", ...uniquePaths] : [],
+	};
+}
+
+function parseGitStatusLine(line: string): GitStatusEntry {
+	const indexStatus = line[0] ?? " ";
+	const worktreeStatus = line[1] ?? " ";
+	const payload = line.slice(3);
+	const renameParts = payload.split(" -> ");
+	const path = renameParts[renameParts.length - 1] ?? payload;
+	const originalPath = renameParts.length === 2 ? renameParts[0] : undefined;
+	const tracked = indexStatus !== "?" && worktreeStatus !== "?";
+	const untracked = indexStatus === "?" && worktreeStatus === "?";
+	return {
+		path,
+		code: `${indexStatus}${worktreeStatus}`,
+		indexStatus,
+		worktreeStatus,
+		tracked,
+		untracked,
+		staged: isChangedStatus(indexStatus),
+		unstaged: isChangedStatus(worktreeStatus),
+		deleted: indexStatus === "D" || worktreeStatus === "D",
+		renamed: indexStatus === "R" || worktreeStatus === "R",
+		copied: indexStatus === "C" || worktreeStatus === "C",
+		originalPath,
+		indexKind: decodeStatus(indexStatus),
+		worktreeKind: decodeStatus(worktreeStatus),
+	};
+}
+
+function decodeStatus(status: string): GitStatusCode {
+	switch (status) {
+		case "M":
+			return "modified";
+		case "A":
+			return "added";
+		case "D":
+			return "deleted";
+		case "R":
+			return "renamed";
+		case "C":
+			return "copied";
+		case "U":
+			return "unmerged";
+		case "?":
+			return "untracked";
+		case "!":
+			return "ignored";
+		case " ":
+			return "unknown";
+		default:
+			return "unknown";
+	}
+}
+
+function isChangedStatus(status: string): boolean {
+	return status !== " " && status !== "?" && status !== "!";
+}
+
+function pathMatchesAllowlist(path: string, allowed: string): boolean {
+	if (path === allowed) return true;
+	if (allowed.endsWith("/")) return path.startsWith(allowed);
+	return false;
+}

package/extensions/lib/local-models.ts

@@ -0,0 +1,157 @@
+/**
+ * Canonical local model registry for Omegon.
+ *
+ * Single source of truth consumed by: offline-driver, effort, cleave, project-memory.
+ * Update this file as new models are released — all preference logic reads from here.
+ *
+ * PREFERENCE ORDERING PRINCIPLES
+ * ─────────────────────────────────────────────────────────────────────────────
+ * Quality-first: best available model wins automatically.
+ * Users install what fits their hardware; the lists handle selection.
+ *
+ *   64GB  → 70B tier   (qwen2.5:72b, llama3.3:70b)
+ *   32GB  → 32B tier   (qwen3:32b, qwen2.5:32b)
+ *   24GB  → 14B Q8 or MoE 30B Q4 (qwen3-coder:30b, qwen3:14b)
+ *   16GB  → 8B Q8     (qwen3:8b, llama3.1:8b)
+ *    8GB  → 4B Q8     (qwen3:4b, phi4-mini)
+ *
+ * MoE NOTE: qwen3-coder:30b has 30B total params but only 3.3B activated —
+ * weight file is ~18GB at Q4 but inference cost is like a 3-4B dense model.
+ * Fits on 24GB, runs fast, explicitly trained on SWE-Bench agentic tasks.
+ *
+ * CONTEXT WINDOW NOTE: phi4:14b has 16K context (too small for deep agent
+ * sessions) — excluded from main preferences. phi4-mini has 128K and is fine.
+ * mistral:7b and codestral:22b are capped at 32K; prefer Qwen/Llama at those sizes.
+ *
+ * Last reviewed: 2026-03-07
+ * Next review triggers: major Qwen4/Llama4/Gemma4 releases, new MoE coding models.
+ */
+
+export interface LocalModelMeta {
+  /** Human-readable display name */
+  label: string;
+  /** Emoji icon for UI display */
+  icon: string;
+  /** Native context window in tokens */
+  contextWindow: number;
+  /** Max generation tokens */
+  maxTokens: number;
+}
+
+// ---------------------------------------------------------------------------
+// Model registry — metadata for known models
+// ---------------------------------------------------------------------------
+
+export const KNOWN_MODELS: Record<string, LocalModelMeta> = {
+  // ── 70B tier (64GB+) ──────────────────────────────────────────────────────
+  "qwen2.5:72b":           { label: "Qwen2.5 72B",            icon: "🧠", contextWindow: 131072,  maxTokens: 32768 },
+  "llama3.3:70b":          { label: "Llama 3.3 70B",          icon: "🦙", contextWindow: 131072,  maxTokens: 32768 },
+  "llama3.1:70b":          { label: "Llama 3.1 70B",          icon: "🦙", contextWindow: 131072,  maxTokens: 32768 },
+
+  // ── 32B tier (32GB+, or 24GB at Q4) ──────────────────────────────────────
+  "qwen3:32b":             { label: "Qwen3 32B",              icon: "🐉", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5:32b":           { label: "Qwen2.5 32B",            icon: "🧠", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5-coder:32b":     { label: "Qwen2.5-Coder 32B",      icon: "💻", contextWindow: 131072,  maxTokens: 32768 },
+
+  // ── MoE 30B tier — runs lean despite total param count ───────────────────
+  // qwen3-coder: 30B total / 3.3B active, ~18GB at Q4. Fits 24GB+.
+  // Explicitly trained on SWE-Bench; best local code-agent model at its size.
+  "qwen3-coder:30b":       { label: "Qwen3-Coder 30B (MoE)",  icon: "🤖", contextWindow: 262144,  maxTokens: 32768 },
+  "qwen3:30b":             { label: "Qwen3 30B",              icon: "🐲", contextWindow: 131072,  maxTokens: 32768 },
+  "nemotron-3-nano:30b":   { label: "Nemotron 3 Nano 30B",    icon: "🏔️", contextWindow: 1048576, maxTokens: 32768 },
+
+  // ── 22–27B tier (16GB+ at Q4, 32GB at Q8) ───────────────────────────────
+  "gemma3:27b":            { label: "Gemma3 27B",             icon: "♊", contextWindow: 131072,  maxTokens: 32768 },
+  // devstral:24b is the current canonical Ollama tag (replaces devstral-small-2:24b)
+  // Built on Mistral-Small-3.1; 53.6% SWE-Bench verified (top OSS as of mid-2025)
+  "devstral:24b":          { label: "Devstral 24B",           icon: "⚙️", contextWindow: 131072,  maxTokens: 32768 },
+  "devstral-small-2:24b":  { label: "Devstral Small 2 24B",   icon: "⚙️", contextWindow: 393216,  maxTokens: 32768 },
+  "mistral-small":         { label: "Mistral Small 22B",      icon: "🌬️", contextWindow: 32768,   maxTokens: 32768 },
+  "codestral:22b":         { label: "Codestral 22B",          icon: "💻", contextWindow: 32768,   maxTokens: 32768 },
+
+  // ── 14B tier (16GB+) ─────────────────────────────────────────────────────
+  "qwen3:14b":             { label: "Qwen3 14B",              icon: "🐉", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5:14b":           { label: "Qwen2.5 14B",            icon: "🧠", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5-coder:14b":     { label: "Qwen2.5-Coder 14B",      icon: "💻", contextWindow: 131072,  maxTokens: 32768 },
+  "mistral-nemo":          { label: "Mistral Nemo 12B",       icon: "🌬️", contextWindow: 131072,  maxTokens: 32768 },
+
+  // ── 7–9B tier (8GB+) ─────────────────────────────────────────────────────
+  "qwen3:8b":              { label: "Qwen3 8B",               icon: "🐉", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5:7b":            { label: "Qwen2.5 7B",             icon: "🧠", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5-coder:7b":      { label: "Qwen2.5-Coder 7B",       icon: "💻", contextWindow: 131072,  maxTokens: 32768 },
+  "llama3.1:8b":           { label: "Llama 3.1 8B",           icon: "🦙", contextWindow: 131072,  maxTokens: 32768 },
+  "llama3.2:11b":          { label: "Llama 3.2 11B",          icon: "🦙", contextWindow: 131072,  maxTokens: 32768 },
+  "gemma3:9b":             { label: "Gemma3 9B",              icon: "♊", contextWindow: 131072,  maxTokens: 32768 },
+  "phi4-mini":             { label: "Phi-4 Mini 3.8B",        icon: "Φ",  contextWindow: 131072,  maxTokens: 32768 },
+  "mistral:7b":            { label: "Mistral 7B",             icon: "🌬️", contextWindow: 32768,   maxTokens: 32768 },
+
+  // ── 3–4B tier (8GB, last useful resort for tool calling) ─────────────────
+  "qwen3:4b":              { label: "Qwen3 4B",               icon: "🐉", contextWindow: 131072,  maxTokens: 32768 },
+  "qwen2.5-coder:3b":      { label: "Qwen2.5-Coder 3B",       icon: "💻", contextWindow: 131072,  maxTokens: 32768 },
+  "gemma3:4b":             { label: "Gemma3 4B",              icon: "♊", contextWindow: 131072,  maxTokens: 32768 },
+  "llama3.2:3b":           { label: "Llama 3.2 3B",           icon: "🦙", contextWindow: 131072,  maxTokens: 32768 },
+
+  // ── Sub-3B (emergency fallback — too small for reliable orchestration) ────
+  "qwen3:1.7b":            { label: "Qwen3 1.7B",             icon: "🐣", contextWindow: 131072,  maxTokens: 8192  },
+  "llama3.2:1b":           { label: "Llama 3.2 1B",           icon: "🐣", contextWindow: 131072,  maxTokens: 8192  },
+  "qwen3:0.6b":            { label: "Qwen3 0.6B",             icon: "🐣", contextWindow: 131072,  maxTokens: 8192  },
+};
+
+// ---------------------------------------------------------------------------
+// Preference lists — consumed by offline-driver, effort, cleave
+// ---------------------------------------------------------------------------
+
+/**
+ * General orchestration preference — best installed model wins.
+ * Ordered 70B → 32B → MoE-30B → 14B → 8B → 4B → sub-3B.
+ * On any hardware, only installed models are found; others are skipped.
+ */
+export const PREFERRED_ORDER: string[] = [
+  // 70B tier (64GB+)
+  "qwen2.5:72b", "llama3.3:70b", "llama3.1:70b",
+  // 32B tier (32GB+)
+  "qwen3:32b", "qwen2.5:32b",
+  // MoE 30B (fits 24GB+ at Q4 despite total param count)
+  "qwen3-coder:30b", "qwen3:30b", "nemotron-3-nano:30b",
+  // 22–27B tier
+  "gemma3:27b", "devstral:24b", "devstral-small-2:24b", "mistral-small",
+  // 14B tier (16GB+)
+  "qwen3:14b", "qwen2.5:14b", "mistral-nemo",
+  // 7–9B tier (8GB+)
+  "qwen3:8b", "llama3.2:11b", "gemma3:9b", "qwen2.5:7b", "llama3.1:8b", "mistral:7b",
+  // 4B tier
+  "qwen3:4b", "phi4-mini", "gemma3:4b",
+  // Sub-3B (last resort)
+  "llama3.2:3b", "qwen3:1.7b", "llama3.2:1b", "qwen3:0.6b",
+];
+
+/**
+ * Code-focused preference for cleave leaf workers.
+ * Biases toward Coder/Devstral/Codestral variants at each size tier.
+ */
+export const PREFERRED_ORDER_CODE: string[] = [
+  // 32B code tier
+  "qwen2.5-coder:32b", "qwen3:32b", "codestral:22b",
+  // MoE 30B (excellent for agentic code tasks, lean inference)
+  "qwen3-coder:30b", "devstral:24b", "devstral-small-2:24b",
+  // 30B general
+  "qwen3:30b",
+  // 14B code tier
+  "qwen2.5-coder:14b", "qwen3:14b", "qwen2.5:14b",
+  // 7–8B code tier
+  "qwen2.5-coder:7b", "qwen3:8b", "llama3.1:8b",
+  // Small code
+  "qwen2.5-coder:3b", "qwen3:4b", "llama3.2:3b",
+  // General fallbacks
+  "nemotron-3-nano:30b", "qwen2.5:7b", "gemma3:9b",
+];
+
+/**
+ * Model family prefixes for startsWith matching (used by project-memory compaction).
+ * Catches any installed variant not listed explicitly in PREFERRED_ORDER.
+ * Ordered by general capability preference within each family.
+ */
+export const PREFERRED_FAMILIES: string[] = [
+  "qwen3-coder", "qwen3", "qwen2.5-coder", "qwen2.5",
+  "llama3", "gemma3", "devstral", "mistral", "nemotron", "phi",
+];

package/extensions/lib/model-preferences.ts

@@ -0,0 +1,51 @@
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+export interface PersistedDriverSelection {
+  provider: string;
+  modelId: string;
+}
+
+export interface PiConfig {
+  lastUsedModel?: PersistedDriverSelection;
+  operatorProfile?: unknown;
+  [key: string]: unknown;
+}
+
+function configPath(root: string): string {
+  return join(root, ".pi", "config.json");
+}
+
+export function loadPiConfig(root: string): PiConfig {
+  try {
+    const path = configPath(root);
+    if (!existsSync(path)) return {};
+    const raw = readFileSync(path, "utf-8");
+    const parsed = JSON.parse(raw);
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed as PiConfig : {};
+  } catch {
+    return {};
+  }
+}
+
+export function savePiConfig(root: string, config: PiConfig): void {
+  const dir = join(root, ".pi");
+  mkdirSync(dir, { recursive: true });
+  writeFileSync(configPath(root), JSON.stringify(config, null, 2) + "\n", "utf-8");
+}
+
+export function readLastUsedModel(root: string): PersistedDriverSelection | undefined {
+  const value = loadPiConfig(root).lastUsedModel;
+  if (!value || typeof value !== "object") return undefined;
+  const record = value as unknown as Record<string, unknown>;
+  const provider = record.provider;
+  const modelId = record.modelId;
+  if (typeof provider !== "string" || typeof modelId !== "string") return undefined;
+  return { provider, modelId };
+}
+
+export function writeLastUsedModel(root: string, selection: PersistedDriverSelection): void {
+  const config = loadPiConfig(root);
+  config.lastUsedModel = selection;
+  savePiConfig(root, config);
+}