ultimate-pi 0.15.0 → 0.17.0

package/.pi/prompts/harness-plan.md

@@ -1,13 +1,13 @@
 ---
 description: PM-grade harness plan — scouts, implementation research, ExecutionPlan, DAG validation, selective Review Gate debate, approval.
-argument-hint: "\"<task>\" [--risk low|med|high] [--budget <amount>] [--quick]"
+argument-hint: "\"<task>\" [--risk low|med|high] [--quick]"
 ---
 
 # harness-plan
 
 You are the **planning PM** for this harness run. Produce an execution baseline (`plan-packet.yaml` + `plan-review.md`), not strategy theater. Parent owns `ask_user`, `approve_plan`, `create_plan`, debate bus commands, and YAML writes under `.pi/harness/runs/<run_id>/`.
 
-Never `write`/`edit` the final canonical packet except via **`write_harness_yaml`** for run artifacts and **`create_plan`** after approval. Do not paste JSON into `.yaml` files — subagents emit JSON; you convert via `write_harness_yaml`.
+Subagents persist artifacts via scoped **`submit_*`** tools (deterministic YAML under the run dir). Parent uses **`harness_artifact_ready`** to gate phases (no JSON parsing). Parent merges still use **`write_harness_yaml`** for `research-brief.yaml`, `plan-packet.yaml` shell, and integrator patches only.
 
 ## Allowed subagents
 
@@ -33,12 +33,12 @@ Read **harness-debate-plan** skill before Review Gate rounds.
 2. Each `subagent` call blocks until subprocesses finish — batch parallel scouts in one `tasks` array.
 3. Do **not** set `timeoutMs` unless the user explicitly requests a cap — subagents run until natural completion (optional backstop: `PI_SUBAGENT_TIMEOUT_MS`).
 4. No harness subagent spawn cap — run the full scout + research + debate pipeline without skipping lanes for budget.
-5. Compact task text: embed `HarnessSpawnContext` JSON + lane-specific instructions only.
+5. Compact task text: embed spawn context + lane instructions. Prefer `HarnessSpawnContext={"run_id":"…","plan_packet_path":"…",…}` or a JSON object with `"HarnessSpawnContext":{…}` — both parse; `run_id` is required so subprocess submit tools get `HARNESS_RUN_ID`.
 
 ## Step 0 — Parse `$ARGUMENTS`
 
 - task (required)
-- `--risk low|med|high`, `--budget`, `--quick`
+- `--risk low|med|high`, `--quick` (`--budget` is reserved/no-op; token budgets are telemetry-only unless `HARNESS_BUDGET_ENFORCE=1`)
 
 `--quick` skips **scout-semantic** and post-run adversary only — **never** skip graphify, structure, decompose, hypothesis, **Phase 3.5 implementation research**, stack research, execution plan, DAG validation, or **Review Gate debate**.
 
@@ -64,9 +64,11 @@ Do **not** run `ccc index` or `ccc search --refresh`. The harness runs increment
 
 Add `harness/planning/scout-semantic` to `tasks` unless `--quick`. Require graphify + structure success. Semantic lane uses `ccc search` only (see `scout-semantic` agent).
 
+After scouts: `harness_artifact_ready({ paths: ["artifacts/scout-graphify.yaml", "artifacts/scout-structure.yaml", ...] })`.
+
 ## Phase 2 & 3 — Decompose + hypothesis (parallel)
 
-One `subagent` call with `tasks` for `harness/planning/decompose` and `harness/planning/hypothesis`. Parse `PlanDecompositionBrief` and `PlanHypothesisBrief` from outputs. Persist with `write_harness_yaml` → `artifacts/decomposition.yaml` and `artifacts/hypothesis.yaml`.
+One `subagent` call with `tasks` for `harness/planning/decompose` and `harness/planning/hypothesis` (include scout YAML paths in task text). Gate with `harness_artifact_ready` on `artifacts/decomposition.yaml` and `artifacts/hypothesis.yaml`.
 
 Decompose **prior_art** is **internal only** (from scouts). External prior art arrives in Phase 3.5.
 
@@ -84,8 +86,8 @@ Decompose **prior_art** is **internal only** (from scouts). External prior art a
 }
 ```
 
-- `write_harness_yaml` → `artifacts/implementation-research.yaml` and `artifacts/stack.yaml`.
-- Merge both into `research-brief.yaml` (`implementation:` + `stack:`).
+- Subagents write via `submit_implementation_research` / `submit_stack_brief`; gate with `harness_artifact_ready` on both paths.
+- Merge both into `research-brief.yaml` (`implementation:` + `stack:`) via parent `write_harness_yaml`.
 - **Partial failure:** if one lane fails, re-spawn that lane once; if still failing set `plan_status: partial` and `human_required` via `ask_user`. Do not proceed to Phase 4b without both artifacts or explicit human waiver.
 - **Web dedup:** implementation owns patterns/repos; stack owns libraries/versions — no overlapping queries.
 
@@ -136,11 +138,16 @@ harness_debate_open({ debate_profile, required_focuses })
 
 Profiles:
 
-| Profile | Focuses required | min_focus_rounds |
-|---------|------------------|------------------|
-| full | spec, wbs, schedule, quality | 4 |
-| standard | all four | 4 |
-| light | spec, quality only | 2 |
+| Profile | Review gate | Focuses required | min_focus_rounds |
+|---------|-------------|------------------|------------------|
+| full | threaded (4 rounds) | spec, wbs, schedule, quality | 4 |
+| standard | threaded (4 rounds) | all four | 4 |
+| light | threaded (2 rounds) | spec, quality only | 2 |
+| fast | **consolidated** (1 round) | spec, quality | 1 |
+
+Med/low non-fork plans with clear stack and no implementation `open_questions` default to **fast** (consolidated). Escalate to threaded rounds only when integrator sets `review_gate_ready: false` or records blockers.
+
+`--quick`: skip scout-semantic; cap web research (≤2 searches, ≤3 fetches); prefer **fast** eligibility when DAG passes; use consolidated Review Gate when profile is fast.
 
 ## Phase 5 — Review Gate debate (profile-aware, pi-messenger, even with `--quick`)
 
@@ -151,13 +158,26 @@ Profiles:
 
 ### Focus coverage (required before consensus)
 
-Each required focus must appear in a submitted `review-round-rN.yaml` (`debate_round_focus`). Monotonic `round_index` (cap from profile). Consensus only when:
+Each required focus must appear in submitted review artifacts (`review-round-rN.yaml` or `review-round-consolidated.yaml` with `debate_round_focus: all`). Monotonic `round_index` (cap from profile). Consensus only when:
 
 - all **required** focuses covered, **and**
 - last round `review_gate_ready: true`, **and**
 - `validate-plan-dag.mjs` still passes (re-run after patches).
 
-### Per-round state machine
+### Consolidated state machine (`review_gate_mode: consolidated`, profile fast)
+
+```
+round_index := 1
+debate_round_focus := all
+spawn hypothesis-validator (blind)
+WHILE NOT ready_for_integrator (harness_debate_round_status round_index=1):
+  follow next_tool (validation-turn, adversary-brief, sprint-audit in parallel-friendly order; one subagent per batch)
+spawn review-integrator → write artifacts/review-round-consolidated.yaml → harness_debate_submit_round
+IF review_gate_ready false OR blockers: escalate — threaded round per missing focus (spec/wbs/schedule/quality)
+harness_debate_focus_coverage → harness_debate_consensus
+```
+
+### Threaded state machine (standard/full/light)
 
 ```
 round_index := next uncovered required focus

package/.pi/prompts/harness-run.md

@@ -1,6 +1,6 @@
 ---
 description: Execute only against an approved PlanPacket with strict phase gates.
-argument-hint: "[--budget <amount>]"
+argument-hint: ""
 ---
 
 # harness-run
@@ -9,7 +9,7 @@ Orchestrator only — spawn `harness/executor`. Do **not** implement inline.
 
 ## Step 0 — Parse arguments
 
-- optional: `--budget <amount>`
+- `--budget` is reserved/no-op (telemetry-only budgets by default)
 - Do **not** use `--plan` on happy path — load from `[HarnessActivePlan]` / `plan_packet_path`.
 
 If plan not ready:

package/.pi/prompts/harness-setup.md

@@ -327,7 +327,7 @@ sentrux plugin add-standard 2>/dev/null || echo "Plugins already installed or fa
 
 ## Step 3 — Pi Extension Packages
 
-Bundled extensions load from the installed `ultimate-pi` package. **Per-turn model routing** comes from a **vendored** fork of [`yeliu84/pi-model-router`](https://github.com/yeliu84/pi-model-router) in `vendor/pi-model-router/`, wired through [`.pi/extensions/pi-model-router-harness.ts`](.pi/extensions/pi-model-router-harness.ts). The harness **gates** activation on `.pi/model-router.json` (Step **3.5** below) so `router/auto` and built-in tiers such as `openai/gpt-5.4-pro` cannot load prematurely. Attribution: see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and `vendor/pi-model-router/UPSTREAM_PIN.md`. Maintainer refresh: `npm run vendor:sync-router`.
+Bundled extensions load from the installed `ultimate-pi` package. **Session-locked model routing** comes from a **vendored** fork of [`yeliu84/pi-model-router`](https://github.com/yeliu84/pi-model-router) in `vendor/pi-model-router/`, wired through [`.pi/extensions/pi-model-router-harness.ts`](.pi/extensions/pi-model-router-harness.ts). The router picks **one concrete model** when the session starts (from the first user prompt + system prompt complexity), then changes **thinking level only** each turn. The harness **gates** activation on `.pi/model-router.json` (Step **3.5** below) so `router/auto` cannot load prematurely. Attribution: see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and `vendor/pi-model-router/UPSTREAM_PIN.md`. Maintainer refresh: `npm run vendor:sync-router`.
 
 Optionally install the companion lockfile used in development:
 
@@ -381,9 +381,9 @@ If generation prints "No authenticated Pi providers": warn in report — user sh
 
 Do NOT block setup. If no config is written, `harness-sync-model-router.mjs` clears a premature `defaultProvider: "router"` in `.pi/settings.json`.
 
-**Router onboarding** — The vendored extension starts only after `.pi/model-router.json` appears. Running the script above prepares that file plus optional Pi defaults (**`router` / `auto`**) via `harness-sync-model-router.mjs` when `defaultProvider` was unset—then **`/reload`**.
+**Router onboarding** — The vendored extension starts only after `.pi/model-router.json` appears. Running the script above prepares that file plus optional Pi defaults (**`router` / `auto`**, or whatever `defaultProfile` is) via `harness-sync-model-router.mjs` when `defaultProvider` was unset—then **`/reload`**. Generated profiles use **one model SKU per profile**; high/medium/low tiers differ in **thinking** only. Subagents resolve their subprocess model from the **agent system prompt** complexity (same lock rules).
 
-Manual override: **`/router profile auto`** anytime after reload if they changed defaults.
+Manual override: **`/router profile auto`** or **`/router profile opencode-go`** anytime after reload if they changed defaults.
 
 ## Step 3.6 — Harness agents (package-resolved)
 
@@ -677,7 +677,7 @@ Output summary table:
 | sentrux | ✓/✗ | CLI + plugins; rules via Step 4.2 bootstrap |
 | Sentrux rules.toml | ✓/✗ | `.sentrux/rules.toml` synced from manifest |
 | pi extensions | ✓/✗ | 4 packages |
-| model router | ✓/✗ | Package + config verified, activation via `/router profile auto` |
+| model router | ✓/✗ | Package + config verified, activation via `/router profile auto` (or `opencode-go`) |
 | `.env` | ✓/✗/ask | Created / keys appended / user declined |
 
 | .gitignore | ✓/✗ | entries added (incl. `.env`) |

package/.pi/scripts/harness-generate-model-router.mjs

@@ -22,9 +22,9 @@ const UP_PKG = join(SCRIPT_DIR, "..", "..");
 const OUT_PATH = join(process.cwd(), ".pi", "model-router.json");
 
 const PROVIDER_PRIORITY = [
+	"openai",
 	"opencode-go",
 	"anthropic",
-	"openai",
 	"google",
 	"openrouter",
 	"groq",
@@ -35,6 +35,7 @@ const PROVIDER_PRIORITY = [
 /** Substring hints per tier (first match in available ids wins). */
 const TIER_HINTS = {
 	high: [
+		"gpt-5.5-pro",
 		"deepseek-v4-pro",
 		"gpt-5.4-pro",
 		"claude-opus",
@@ -43,6 +44,7 @@ const TIER_HINTS = {
 		"pro",
 	],
 	medium: [
+		"gpt-5.5",
 		"qwen3.6-plus",
 		"kimi-k2.6",
 		"gpt-5.4",
@@ -98,7 +100,10 @@ function canonicalRef(provider, modelId) {
 
 function pickTierModel(models, tier) {
 	const hints = TIER_HINTS[tier];
-	const ids = models.map((m) => m.id);
+	for (const hint of hints) {
+		const exact = models.find((m) => m.id === hint);
+		if (exact) return canonicalRef(exact.provider, exact.id);
+	}
 	for (const hint of hints) {
 		const match = models.find((m) => m.id.includes(hint));
 		if (match) return canonicalRef(match.provider, match.id);
@@ -114,6 +119,10 @@ function pickTierModel(models, tier) {
 	return canonicalRef(models[0].provider, models[0].id);
 }
 
+function modelsForProvider(available, provider) {
+	return available.filter((m) => m.provider === provider);
+}
+
 function choosePrimaryProvider(available) {
 	const byProvider = new Map();
 	for (const m of available) {
@@ -129,7 +138,7 @@ function choosePrimaryProvider(available) {
 
 function buildFallbacks(available, primaryProvider, highModel) {
 	const fallbacks = [];
-	for (const p of ["anthropic", "google", "openai"]) {
+	for (const p of ["anthropic", "google", "openai", "opencode-go"]) {
 		if (p === primaryProvider) continue;
 		const alt = available.filter((m) => m.provider === p);
 		if (alt.length === 0) continue;
@@ -139,6 +148,76 @@ function buildFallbacks(available, primaryProvider, highModel) {
 	return fallbacks.slice(0, 3);
 }
 
+/** Session-locked router: one model SKU per profile; tiers vary thinking only. */
+function buildRoutedProfile(available, provider) {
+	const models = modelsForProvider(available, provider);
+	if (models.length === 0) return null;
+	const sku =
+		pickTierModel(models, "medium") ??
+		pickTierModel(models, "high") ??
+		pickTierModel(models, "low");
+	if (!sku) return null;
+	const fallbacks = buildFallbacks(available, provider, sku);
+	const high = { model: sku, thinking: "high" };
+	if (fallbacks.length) high.fallbacks = fallbacks;
+	return {
+		high,
+		medium: { model: sku, thinking: "medium" },
+		low: { model: sku, thinking: "low" },
+	};
+}
+
+function addCheapDeepProfiles(profiles, available, provider) {
+	const models = modelsForProvider(available, provider);
+	if (models.length === 0) return;
+	const sku =
+		pickTierModel(models, "medium") ??
+		pickTierModel(models, "high") ??
+		pickTierModel(models, "low");
+	if (!sku) return;
+	const fallbacks = buildFallbacks(available, provider, sku);
+	const deepHigh = { model: sku, thinking: "xhigh" };
+	if (fallbacks.length) deepHigh.fallbacks = fallbacks;
+	profiles.cheap = {
+		high: { model: sku, thinking: "low" },
+		medium: { model: sku, thinking: "off" },
+		low: { model: sku, thinking: "off" },
+	};
+	profiles.deep = {
+		high: deepHigh,
+		medium: { model: sku, thinking: "medium" },
+		low: { model: sku, thinking: "low" },
+	};
+}
+
+function resolveClassifierModel(available) {
+	const openaiModels = modelsForProvider(available, "openai");
+	if (openaiModels.length > 0) {
+		return (
+			pickTierModel(openaiModels, "low") ??
+			canonicalRef(openaiModels[openaiModels.length - 1].provider, openaiModels[openaiModels.length - 1].id)
+		);
+	}
+	const { models } = choosePrimaryProvider(available);
+	return pickTierModel(models, "medium");
+}
+
+/** OpenAI-backed default profile name exposed as `router/auto`. */
+const OPENAI_PROFILE_NAME = "auto";
+
+function routerProfileName(provider) {
+	return provider === "openai" ? OPENAI_PROFILE_NAME : provider;
+}
+
+function resolveDefaultProfile(profiles) {
+	if (profiles[OPENAI_PROFILE_NAME]) return OPENAI_PROFILE_NAME;
+	if (profiles["opencode-go"]) return "opencode-go";
+	return (
+		Object.keys(profiles).find((name) => name !== "cheap" && name !== "deep") ??
+		OPENAI_PROFILE_NAME
+	);
+}
+
 async function main() {
 	const force = process.argv.includes("--force");
 	const dryRun = process.argv.includes("--dry-run");
@@ -171,23 +250,37 @@ async function main() {
 		process.exit(0);
 	}
 
-	const { provider: primaryProvider, models: primaryModels } =
-		choosePrimaryProvider(available);
-
-	const highModel = pickTierModel(primaryModels, "high");
-	const mediumModel = pickTierModel(primaryModels, "medium");
-	const lowModel = pickTierModel(primaryModels, "low");
+	const profiles = {};
+	for (const provider of ["openai", "opencode-go"]) {
+		const profile = buildRoutedProfile(available, provider);
+		if (profile) profiles[routerProfileName(provider)] = profile;
+	}
 
-	if (!highModel || !mediumModel || !lowModel) {
-		fail("could not assign tier models from available registry");
+	if (Object.keys(profiles).length === 0) {
+		const { provider: primaryProvider, models: primaryModels } =
+			choosePrimaryProvider(available);
+		const profile = buildRoutedProfile(available, primaryProvider);
+		if (!profile) {
+			fail("could not assign tier models from available registry");
+		}
+		profiles[primaryProvider] = profile;
 	}
 
-	const fallbacks = buildFallbacks(available, primaryProvider, highModel);
+	const cheapDeepSource = profiles["opencode-go"]
+		? "opencode-go"
+		: resolveDefaultProfile(profiles);
+	addCheapDeepProfiles(profiles, available, cheapDeepSource);
+
+	const defaultProfile = resolveDefaultProfile(profiles);
+	const classifierModel = resolveClassifierModel(available);
+	if (!classifierModel) {
+		fail("could not assign classifier model from available registry");
+	}
 
 	const config = {
-		defaultProfile: "auto",
+		defaultProfile,
 		debug: false,
-		classifierModel: mediumModel,
+		classifierModel,
 		phaseBias: 0.5,
 		maxSessionBudget: 1.0,
 		largeContextThreshold: 100000,
@@ -199,27 +292,13 @@ async function main() {
 			},
 			{ matches: "changelog", tier: "low" },
 		],
-		profiles: {
-			auto: {
-				high: { model: highModel, thinking: "high", fallbacks },
-				medium: { model: mediumModel, thinking: "medium" },
-				low: { model: lowModel, thinking: "low" },
-			},
-			cheap: {
-				high: { model: mediumModel, thinking: "low" },
-				medium: { model: lowModel, thinking: "off" },
-				low: { model: lowModel, thinking: "off" },
-			},
-			deep: {
-				high: { model: highModel, thinking: "xhigh", fallbacks },
-				medium: { model: mediumModel, thinking: "medium" },
-				low: { model: lowModel, thinking: "low" },
-			},
-		},
+		profiles,
 	};
 
 	const json = `${JSON.stringify(config, null, 2)}\n`;
 	const providerSet = [...new Set(available.map((m) => m.provider))].sort();
+	const autoProfile = profiles[OPENAI_PROFILE_NAME];
+	const opencodeProfile = profiles["opencode-go"];
 
 	if (dryRun) {
 		process.stdout.write(json);
@@ -230,13 +309,16 @@ async function main() {
 	writeFileSync(OUT_PATH, json, "utf8");
 
 	console.log("✓ Generated .pi/model-router.json from Pi authenticated providers:");
-	console.log(`  Primary provider: ${primaryProvider}`);
+	console.log(`  Default profile: ${defaultProfile}`);
+	console.log(`  Classifier: ${classifierModel}`);
 	console.log(`  Authenticated providers: ${providerSet.join(", ")}`);
 	console.log(`  Available models: ${available.length}`);
-	console.log(`  High tier: ${highModel}`);
-	console.log(`  Medium tier: ${mediumModel}`);
-	console.log(`  Low tier: ${lowModel}`);
-	if (fallbacks.length) console.log(`  Fallbacks: ${fallbacks.join(", ")}`);
+	if (autoProfile) {
+		console.log(`  auto (openai) high: ${autoProfile.high.model}`);
+	}
+	if (opencodeProfile) {
+		console.log(`  opencode-go high: ${opencodeProfile.high.model}`);
+	}
 }
 
 main().catch((err) => {

package/.pi/scripts/harness-model-router-routing.test.mjs

@@ -0,0 +1,97 @@
+#!/usr/bin/env node
+/**
+ * Unit tests for session-locked pi-model-router routing (no LLM).
+ * Run: npx tsx .pi/scripts/harness-model-router-routing.test.mjs
+ */
+
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import {
+	decideSessionLock,
+	applyThinkingToDecision,
+	buildRoutingDecision,
+	decideRouting,
+} from "../../vendor/pi-model-router/extensions/routing.js";
+
+const ROOT = join(dirname(fileURLToPath(import.meta.url)), "..", "..");
+
+const sampleProfile = {
+	high: { model: "openai/gpt-5.5", thinking: "high" },
+	medium: { model: "openai/gpt-5.5", thinking: "medium" },
+	low: { model: "openai/gpt-5.5", thinking: "low" },
+};
+
+const planningContext = {
+	systemPrompt: "You are a harness architect. Design tradeoffs and migration strategy.",
+	messages: [
+		{
+			role: "user",
+			content:
+				"Plan a multi-phase refactor across modules with architecture review.",
+			timestamp: 1,
+		},
+	],
+};
+
+const shortContext = {
+	systemPrompt: "Summarize briefly.",
+	messages: [{ role: "user", content: "changelog", timestamp: 1 }],
+};
+
+const lockHigh = decideSessionLock(
+	planningContext,
+	"auto",
+	sampleProfile,
+	undefined,
+	undefined,
+	0.5,
+	[{ matches: "changelog", tier: "low" }],
+);
+assert.equal(lockHigh.tier, "high", "planning prompt locks high tier");
+
+const lockLow = decideSessionLock(shortContext, "auto", sampleProfile);
+assert.equal(lockLow.tier, "low", "short summary locks low tier");
+
+const locked = buildRoutingDecision(
+	"auto",
+	sampleProfile,
+	lockHigh.tier,
+	"planning",
+	lockHigh.reasoning,
+);
+const thinkingTurn = decideRouting(
+	{
+		...planningContext,
+		messages: [
+			...planningContext.messages,
+			{ role: "user", content: "changelog only", timestamp: 2 },
+		],
+	},
+	"auto",
+	sampleProfile,
+	locked,
+);
+const merged = applyThinkingToDecision(locked, thinkingTurn, sampleProfile);
+assert.equal(merged.targetLabel, locked.targetLabel, "model stays locked");
+assert.equal(merged.tier, thinkingTurn.tier, "thinking tier follows turn");
+assert.equal(merged.thinking, "low", "low thinking from turn tier config");
+
+const examplePath = join(ROOT, ".pi", "model-router.example.json");
+const example = JSON.parse(readFileSync(examplePath, "utf8"));
+for (const [name, profile] of Object.entries(example.profiles ?? {})) {
+	const { high, medium, low } = profile;
+	assert.equal(
+		high.model,
+		medium.model,
+		`example profile ${name}: medium/high same model`,
+	);
+	assert.equal(
+		medium.model,
+		low.model,
+		`example profile ${name}: low/medium same model`,
+	);
+}
+
+console.log("harness-model-router-routing.test: PASS");

package/.pi/scripts/harness-sync-model-router.mjs

@@ -29,11 +29,24 @@ function saveSettings(settingsPath, data) {
 	);
 }
 
+function readDefaultRouterProfile(configPath) {
+	if (!existsSync(configPath)) return "auto";
+	try {
+		const data = JSON.parse(readFileSync(configPath, "utf8"));
+		const profile =
+			typeof data.defaultProfile === "string" ? data.defaultProfile.trim() : "";
+		return profile || "auto";
+	} catch {
+		return "auto";
+	}
+}
+
 function main() {
 	const root = process.cwd();
 	const configPath = join(root, ".pi", "model-router.json");
 	const settingsPath = join(root, ".pi", "settings.json");
 	const hasConfig = existsSync(configPath);
+	const defaultRouterProfile = readDefaultRouterProfile(configPath);
 
 	const settings = loadSettings(settingsPath);
 	if (!settings) {
@@ -67,14 +80,14 @@ function main() {
 
 	if (noProjectDefault) {
 		settings.defaultProvider = "router";
-		settings.defaultModel = "auto";
+		settings.defaultModel = defaultRouterProfile;
 		changed = true;
 	}
 
 	if (changed) {
 		saveSettings(settingsPath, settings);
 		console.log(
-			"✓ Router defaults set (`router` / `auto`) — run /reload in pi when ready",
+			`✓ Router defaults set (\`router\` / \`${defaultRouterProfile}\`) — run /reload in pi when ready`,
 		);
 	} else {
 		console.log("[harness-model-router] Defaults unchanged (user set defaultProvider)");

package/.pi/scripts/harness-verify.mjs

@@ -37,6 +37,8 @@ const REQUIRED_ADRS = [
 	"0009-sentrux-rules-lifecycle.md",
 	"0031-harness-run-context.md",
 	"0032-harness-command-orchestration.md",
+	"0037-subagent-submit-tools.md",
+	"0038-budget-telemetry-only.md",
 ];
 
 const REQUIRED_EXTENSIONS = [
@@ -143,6 +145,34 @@ async function checkSentruxRules() {
 	ok(".sentrux/rules.toml present");
 }
 
+async function checkModelRouterThinkingOnly() {
+	const path = join(ROOT, ".pi", "model-router.json");
+	if (!(await fileExists(path))) {
+		ok("model-router.json absent (skip thinking-only tier check)");
+		return;
+	}
+	let raw;
+	try {
+		raw = JSON.parse(await readFile(path, "utf-8"));
+	} catch {
+		fail("invalid .pi/model-router.json");
+	}
+	const profiles = raw.profiles ?? {};
+	for (const [name, profile] of Object.entries(profiles)) {
+		const high = profile?.high?.model;
+		const medium = profile?.medium?.model;
+		const low = profile?.low?.model;
+		if (
+			!(high && medium && low && high === medium && medium === low)
+		) {
+			fail(
+				`model-router profile "${name}" must use the same model on high/medium/low (thinking-only tiers)`,
+			);
+		}
+	}
+	ok("model-router.json thinking-only (same model per profile)");
+}
+
 async function checkSentruxGate() {
 	await checkSentruxRules();
 
@@ -286,6 +316,7 @@ async function main() {
 	ok("test-diff-golden.json");
 
 	await checkSentruxGate();
+	await checkModelRouterThinkingOnly();
 
 	if (!(await fileExists(AGENTS_MANIFEST))) {
 		fail(

package/CHANGELOG.md

@@ -4,6 +4,27 @@ All notable changes to this project are documented in this file.
 
 ## [Unreleased]
 
+## [v0.17.0] — 2026-05-22
+
+### ✨ Features
+
+- **Model router:** Session-locked model SKU at start (initial prompt + system prompt); per-turn routing adjusts thinking tier only; subagents lock from agent `systemPrompt` complexity.
+- **Harness:** Thinking-only profile shape in generator/verify; plan review gate, debate eligibility, and smoke fixture updates.
+
+### ✅ Tests
+
+- Add `harness-model-router-routing` and plan-debate eligibility coverage.
+
+## [v0.16.0] — 2026-05-19
+
+### ✨ Features
+
+- add submit pipeline and planning/debate updates
+
+### 🔧 Chores
+
+- refresh graph artifacts after harness updates
+
 ## [v0.15.0] — 2026-05-19
 
 ### ✨ Features

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "ultimate-pi",
-	"version": "0.15.0",
+	"version": "0.17.0",
 	"description": "Ultimate AI coding harness for pi.dev — extensible skills, Obsidian wiki knowledge layer, compressed context, deterministic output",
 	"keywords": [
 		"pi-package",
@@ -84,7 +84,7 @@
 		"format": "biome format --write",
 		"format:check": "biome format",
 		"prepare": "lefthook install",
-		"test": "node --test test/harness-verify.test.mjs test/harness-ask-user.test.mjs test/harness-subagents-loader.test.mjs test/harness-subagent-precheck.test.mjs test/sentrux-rules-sync.test.mjs test/harness-budget-guard.test.mjs && node .pi/harness/evals/smoke/smoke-harness-plan.mjs --fixture && npx -y tsx --test test/harness-vcc-settings.test.ts test/harness-live-widget-status.test.ts test/harness-plan-phase-policy.test.mjs test/harness-subagent-policy.test.mjs test/harness-spawn-budget.test.mjs test/harness-turn-routing.test.mjs test/plan-approval-format.test.mjs test/plan-approval-dialog.test.mjs test/plan-approval-sync.test.mjs test/plan-create-plan.test.mjs test/plan-review-format.test.mjs test/debate-plan-phase.test.mjs test/plan-debate-eligibility.test.mjs test/plan-messenger-gate.test.mjs test/plan-debate-lane-apply.test.mjs",
+		"test": "node --test test/harness-verify.test.mjs test/harness-ask-user.test.mjs test/harness-subagents-loader.test.mjs test/harness-subagent-precheck.test.mjs test/sentrux-rules-sync.test.mjs test/harness-budget-guard.test.mjs && node .pi/harness/evals/smoke/smoke-harness-plan.mjs --fixture && npx -y tsx --test test/harness-vcc-settings.test.ts test/harness-live-widget-status.test.ts test/harness-plan-phase-policy.test.mjs test/harness-subagent-policy.test.mjs test/harness-spawn-budget.test.mjs test/harness-spawn-parse.test.mjs test/harness-schema-validate.test.mjs test/harness-turn-routing.test.mjs test/harness-budget-enforce.test.mjs test/harness-submit-policy.test.mjs test/plan-approval-format.test.mjs test/plan-approval-dialog.test.mjs test/plan-approval-sync.test.mjs test/plan-create-plan.test.mjs test/plan-review-format.test.mjs test/debate-plan-phase.test.mjs test/plan-debate-eligibility.test.mjs test/plan-messenger-gate.test.mjs test/plan-debate-lane-apply.test.mjs",
 		"test:vcc": "npx -y tsx --test vendor/pi-vcc/tests/*.test.ts",
 		"harness:sentrux-bootstrap": "node .pi/scripts/harness-sentrux-bootstrap.mjs",
 		"harness:sentrux-sync": "node .pi/scripts/sentrux-rules-sync.mjs --force",
@@ -103,6 +103,8 @@
 	},
 	"dependencies": {
 		"@posthog/pi": "latest",
+		"ajv": "^8.17.1",
+		"ajv-formats": "^3.0.1",
 		"croner": "^9.0.0",
 		"jimp": "^1.6.1",
 		"nanoid": "^5.1.5",