@selesai/code 0.13.29 → 0.13.30

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 All notable changes to `@selesai/code` will be documented in this file.
 
+## [0.13.30] - 2026-09-23
+
+### Added
+- **Four default skills from `addyosmani/agent-skills`.** `planning-and-task-breakdown`, `incremental-implementation`, `code-review-and-quality`, and `code-simplification` now ship as bundled skills and seed into the agent directory like the rest. Their shared reference files (`definition-of-done.md`, `security-checklist.md`, `performance-checklist.md`) ship in `src/skills/references/` with the upstream MIT license, and the links in the skills that cite them were rebased to that location.
+- **Jev advisory routing (opt-in beta).** The new bundled `jev-advisory-routing` extension can make a bounded, read-only durable-memory lookup after an explicit memory cue, and suggest one discovered skill/workflow or a proportionate verification level. Both routes are disabled by default; neither writes memory, loads a skill, starts a workflow, runs a command, or weakens required checks.
+- **Jev capability tie-breaking (opt-in beta).** The capability gateway can use Jev only to break an ambiguous lexical hint between two or three optional tools. It never receives the full catalog, tool schemas, or conversation history; unavailable, malformed, low-confidence, and `none` decisions preserve deterministic behavior.
+
+### Changed
+- **One Jev transport for every consumer.** Advisory routing and capability tie-breaking share bounded request validation and the Token-In decisions-model path. `jev-*` calls use Token-In's non-streaming endpoint and retain its rotateable-account failover.
+
 ## [0.13.29] - 2026-09-22
 
 ### Changed

package/README.md

@@ -13,13 +13,13 @@
 
 </div>
 
-Selesai is a terminal-first **AI coding agent** that takes software tasks from prompt to verified result. It is a maintained, extension-first fork of the [Pi coding agent](https://github.com/earendil-works/pi), with subagents, adaptive workflows, live research, recovery tools, a richer TUI, and **28 built-in skills** shipped together in one install.
+Selesai is a terminal-first **AI coding agent** that takes software tasks from prompt to verified result. It is a maintained, extension-first fork of the [Pi coding agent](https://github.com/earendil-works/pi), with subagents, adaptive workflows, live research, recovery tools, a richer TUI, and **32 built-in skills** shipped together in one install.
 
 Use your preferred model provider, connect a local or OpenAI-compatible endpoint, or get a ready-to-use token from **[token.in](https://token.selesai.in)**. Selesai keeps Pi's small, hackable core while reducing the setup and extension work needed before an agent can be productive.
 
 | | |
 | --- | --- |
-| **Ready out of the box** | One package bundles coordinated extensions, tools, themes, and 28 skills across 9 meaningful categories. |
+| **Ready out of the box** | One package bundles coordinated extensions, tools, themes, and 32 skills across 9 meaningful categories. |
 | **Delegates and parallelizes** | Focused subagents can research, review, build, or run independent workstreams with foreground, background, parallel, and chained execution. |
 | **Adapts to the task** | The workflow chooses the smallest useful path—from a direct edit to reconnaissance, implementation, review, fixes, and validation. |
 | **Researches with current sources** | Built-in web research, public GitHub code search, and interactive questions help the agent resolve uncertainty before it acts. |
@@ -113,6 +113,8 @@ Skills ship inside `@selesai/code` and are discovered when Selesai starts. Invok
 
 ### 📋 Planning & orchestration
 
+- `planning-and-task-breakdown` — turn a spec into ordered, implementable tasks with acceptance criteria.
+- `incremental-implementation` — land work in thin, verifiable slices with per-increment checks.
 - `pi-subagents` — delegation patterns for parallel, asynchronous, and chained work.
 
 ### 🧹 Code simplicity — the ponytail family
@@ -127,6 +129,8 @@ Skills ship inside `@selesai/code` and are discovered when Selesai starts. Invok
 ### 🛠️ Codebase quality
 
 - `improve-codebase` — inspect and improve architecture, maintainability, and overall codebase quality.
+- `code-review-and-quality` — five-axis review, severity labels, and change sizing before merge.
+- `code-simplification` — remove accumulated complexity without changing behavior.
 
 ### ✅ Completion & output discipline
 
@@ -146,6 +150,8 @@ Skills ship inside `@selesai/code` and are discovered when Selesai starts. Invok
 
 - `agent-browser` — browser and Electron automation, scraping, and QA workflows.
 
+`planning-and-task-breakdown`, `incremental-implementation`, `code-review-and-quality`, and `code-simplification` are adapted from [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills) (MIT). Their shared reference material ships in `src/skills/references/`, with the upstream license alongside it.
+
 ---
 
 ## What Selesai adds to Pi

package/dist/core/model-registry.d.ts

@@ -1,4 +1,4 @@
-import type { Api, AuthResult, Model, Provider } from "@earendil-works/pi-ai";
+import type { Api, AssistantMessage, AssistantMessageEventStream, AuthResult, Context, Model, ModelsSimpleStreamOptions, Provider } from "@earendil-works/pi-ai";
 import type { AuthStorage } from "./auth-storage.ts";
 import type { ModelRuntime } from "./model-runtime.ts";
 import type { AuthStatus, ProviderConfigInput } from "./provider-composer.ts";
@@ -31,6 +31,18 @@ export declare class ModelRegistry {
     hasConfiguredAuth(model: Model<Api>): boolean;
     getApiKeyAndHeaders(model: Model<Api>): Promise<ResolvedRequestAuth>;
     getProviderAuthStatus(provider: string): AuthStatus;
+    /**
+     * Run a request through the composed provider layer — the same path the agent
+     * uses, so a provider's `streamSimple` override, auth resolution, base-URL
+     * overrides, and header transforms all apply.
+     *
+     * Extensions that call a model directly should use this instead of pi-ai's
+     * compat dispatch, which knows nothing about the providers this runtime
+     * composes (a provider override registered here is invisible to it).
+     */
+    streamSimple(model: Model<Api>, context: Context, options?: ModelsSimpleStreamOptions): AssistantMessageEventStream;
+    /** One non-streaming request through the composed provider layer. */
+    complete(model: Model<Api>, context: Context, options?: ModelsSimpleStreamOptions): Promise<AssistantMessage>;
     getProvider(provider: string): Provider | undefined;
     getProviderDisplayName(provider: string): string;
     getProviderAuth(provider: string): Promise<AuthResult | undefined>;

package/dist/core/model-registry.js

@@ -63,6 +63,22 @@ export class ModelRegistry {
     getProviderAuthStatus(provider) {
         return this.runtime.getProviderAuthStatus(provider);
     }
+    /**
+     * Run a request through the composed provider layer — the same path the agent
+     * uses, so a provider's `streamSimple` override, auth resolution, base-URL
+     * overrides, and header transforms all apply.
+     *
+     * Extensions that call a model directly should use this instead of pi-ai's
+     * compat dispatch, which knows nothing about the providers this runtime
+     * composes (a provider override registered here is invisible to it).
+     */
+    streamSimple(model, context, options) {
+        return this.runtime.streamSimple(model, context, options);
+    }
+    /** One non-streaming request through the composed provider layer. */
+    complete(model, context, options) {
+        return this.runtime.complete(model, context, options);
+    }
     getProvider(provider) {
         return this.runtime.getProvider(provider);
     }

package/dist/defaults/settings.json

@@ -1,21 +1,16 @@
 {
-  "autoModel": {
-    "enabled": false,
-    "classifier": {
-      "provider": "tokenin",
-      "model": "jev-1.13"
-    },
-    "tiers": {
-      "simple": "tokenin/deepseek-v4.1-flash",
-      "medium": "tokenin/celestial-pro",
-      "complex": "tokenin/celestial-max",
-      "reasoning": "tokenin/celestial-ultra"
-    }
-  },
   "autoHandoff": {
     "enabled": true,
     "thresholdTokens": 256000
   },
+  "jevAdvisory": {
+    "provider": "tokenin",
+    "model": "jev-1.13",
+    "routes": {
+      "memory": { "enabled": false },
+      "recommendations": { "enabled": false }
+    }
+  },
   "collapseChangelog": true,
   "compaction": {
     "enabled": true,

package/dist/extensions/capability-gateway/catalog.ts

@@ -29,13 +29,13 @@ export function firstSentence(text: string): string {
 	return (end === -1 ? trimmed : trimmed.slice(0, end + 1)).trim();
 }
 
-/** Compact summary for a tool: explicit discovery metadata wins, then first sentence, then snippet. */
+/** Compact summary for a tool: explicit discovery metadata wins, then the description's first sentence. */
 export function toolSummary(tool: ToolInfo): string {
 	const explicit = tool.discovery?.summary?.trim();
 	if (explicit) return explicit;
 	const sentence = firstSentence(tool.description);
 	if (sentence) return sentence;
-	return tool.promptSnippet?.trim() || tool.name;
+	return tool.name;
 }
 
 export function buildToolCatalog(tools: ToolInfo[], gatewayToolNames: Set<string>): CatalogEntry[] {

package/dist/extensions/capability-gateway/index.ts

@@ -16,8 +16,19 @@
  * - A deterministic router activates a uniquely matched tool before the run.
  *   Skills and ambiguous matches remain discoverable through the catalog
  *   without injecting fuzzy hints into the model context.
+ * - Opt-in Jev-assisted routing (capabilityGateway.routing.jev in settings.json)
+ *   is only a bounded tie-breaker: when the deterministic router returns an
+ *   ambiguous lexical hint among optional tools, the gateway offers just those
+ *   hinted tools (two or three) and the current prompt to the Jev decisions
+ *   model as one constrained choice question. Jev may answer `none` or one
+ *   hinted canonical tool name; the gateway revalidates the choice against the
+ *   live catalog and activates it for the current run only. No hint, a unique
+ *   activation, a skill match, or an already-activated tool never reaches Jev.
+ *   Every Jev failure is an ordinary abstention that leaves the deterministic
+ *   behavior in place, so a missing Token-In subscription is invisible.
  * - Temporary activations reset at agent_settled, restoring the baseline
- *   active-tool set.
+ *   active-tool set; a tool_execution_start for one of them records whether the
+ *   activation was actually used (content-free tool name + source only).
  * - The system-prompt skill index is replaced by a compact capability
  *   instruction; full skill instructions load only on explicit show/invoke.
  * - Telemetry events are emitted on the shared event bus.
@@ -36,6 +47,14 @@ import {
 	route,
 	type CatalogEntry,
 } from "./catalog.ts";
+import {
+	hintedToolCandidates,
+	MIN_GATEWAY_JEV_CANDIDATES,
+	readGatewayJevConfig,
+	routeToJevTool,
+	JEV_UNAVAILABLE_REASONS,
+} from "./routing.ts";
+import { confidenceBucket } from "../jev/decisions.ts";
 
 export const GATEWAY_ENV = "SELESAI_CAPABILITY_GATEWAY";
 export const GATEWAY_TOOLS = new Set(["capability_catalog", "capability_discover", "capability_skill_show"]);
@@ -118,9 +137,25 @@ function emitTelemetry(pi: ExtensionAPI, event: string, data: Record<string, unk
 	}
 }
 
+/** Where a run-local tool activation came from; recorded so use telemetry can attribute it. */
+type ActivationSource = "deterministic" | "jev" | "discover";
+
 export default function capabilityGatewayExtension(pi: ExtensionAPI): void {
 	if (!isEnabled()) return;
 
+	// Tools this gateway activated for the current run, and whether they were invoked.
+	// Cleared at agent_settled with the activations themselves.
+	const activations = new Map<string, { source: ActivationSource; used: boolean }>();
+
+	/** Activate one tool for the current run, keeping the rest of the loadout untouched. */
+	function activateTool(name: string, source: ActivationSource): void {
+		const active = pi.getActiveTools();
+		if (!active.includes(name)) {
+			pi.setActiveTools([...active, name]);
+		}
+		activations.set(name, { source, used: false });
+	}
+
 	// ------------------------------------------------------------------
 	// Session start: snapshot baseline, make extension tools dormant, and
 	// install the compact skill index (action methods are stubs until bind).
@@ -182,7 +217,8 @@ export default function capabilityGatewayExtension(pi: ExtensionAPI): void {
 								})
 							: filtered;
 			const text = formatCatalog(shown);
-			emitTelemetry(pi, "catalog", { query: params.query ?? "", kind: params.kind ?? "", results: shown.length });
+			// Content-free: catalog query and filter arguments never travel.
+			emitTelemetry(pi, "catalog", { results: shown.length });
 			return {
 				content: [{ type: "text", text }],
 				details: { count: shown.length, total: filtered.length },
@@ -220,10 +256,7 @@ export default function capabilityGatewayExtension(pi: ExtensionAPI): void {
 					details: { activated: false },
 				};
 			}
-			const active = pi.getActiveTools();
-			if (!active.includes(entry.name)) {
-				pi.setActiveTools([...active, entry.name]);
-			}
+			activateTool(entry.name, "discover");
 			emitTelemetry(pi, "discover", { tool: entry.name });
 			return {
 				content: [
@@ -280,32 +313,84 @@ export default function capabilityGatewayExtension(pi: ExtensionAPI): void {
 	});
 
 	// ------------------------------------------------------------------
-	// Deterministic routing: high-confidence activation/recommendation,
-	// ambiguity hints, or nothing for unrelated prompts.
+	// Routing: the deterministic catalog router first; opt-in Jev routing is
+	// only a bounded tie-breaker for its ambiguous/hint result.
 	// ------------------------------------------------------------------
-	pi.on("before_agent_start", (event) => {
-		const result = routePrompt(event.prompt, catalogEntries(pi));
-		if (result.action === "none") return undefined;
+	pi.on("before_agent_start", async (event, ctx) => {
+		const entries = catalogEntries(pi);
+		const result = routePrompt(event.prompt, entries);
 		if (result.action === "activate" && result.entry) {
-			const active = pi.getActiveTools();
-			if (!active.includes(result.entry.name)) {
-				pi.setActiveTools([...active, result.entry.name]);
-			}
-			emitTelemetry(pi, "route_activate", { tool: result.entry.name });
+			activateTool(result.entry.name, "deterministic");
+			emitTelemetry(pi, "route", {
+				source: "deterministic",
+				outcome: "activated",
+				tool: result.entry.name,
+				candidates: 0,
+				durationMs: 0,
+			});
 			return undefined;
 		}
 		// Do not inject fuzzy recommendations or catalog hints into the model
 		// context. It can discover capabilities when it actually needs one.
+		// Jev is a tie-breaker, never a classifier: only the ambiguous/hint
+		// result counts. No lexical signal or unique skill recommendation
+		// reaches it, and skills are never offered as candidates.
+		if (result.action !== "hint") return undefined;
+		const config = readGatewayJevConfig();
+		if (!config.enabled) return undefined;
+		const candidates = hintedToolCandidates(result.candidates);
+		if (candidates.length < MIN_GATEWAY_JEV_CANDIDATES) return undefined;
+
+		emitTelemetry(pi, "route", { source: "jev", outcome: "attempt", candidates: candidates.length });
+		const jevRoute = await routeToJevTool(candidates, event.prompt, ctx, config);
+		if (!jevRoute.selected) {
+			emitTelemetry(pi, "route", {
+				source: "jev",
+				outcome: JEV_UNAVAILABLE_REASONS.has(jevRoute.reason) ? "unavailable" : "abstained",
+				reason: jevRoute.reason,
+				candidates: jevRoute.candidates,
+				durationMs: jevRoute.elapsedMs,
+			});
+			return undefined;
+		}
+
+		// Revalidate the accepted choice against the live catalog immediately
+		// before activation: a stale or no-longer-eligible name activates nothing.
+		const live = eligibleTools(pi).find((tool) => tool.name === jevRoute.tool);
+		if (live) activateTool(live.name, "jev");
+		emitTelemetry(pi, "route", {
+			source: "jev",
+			outcome: live ? "activated" : "abstained",
+			...(live ? { tool: live.name } : { tool: jevRoute.tool }),
+			confidence: confidenceBucket(jevRoute.confidence),
+			candidates: jevRoute.candidates,
+			durationMs: jevRoute.elapsedMs,
+		});
 		return undefined;
 	});
 
 	// ------------------------------------------------------------------
-	// Reset: restore the baseline active-tool set after the run settles.
+	// Selected -> used: a run-local activation is "used" when its tool is
+	// actually executed before the run settles. Content-free: name + source.
+	// ------------------------------------------------------------------
+	pi.on("tool_execution_start", (event) => {
+		const activation = activations.get(event.toolName);
+		if (!activation || activation.used) return;
+		activation.used = true;
+		emitTelemetry(pi, "use", { tool: event.toolName, source: activation.source });
+	});
+
+	// ------------------------------------------------------------------
+	// Reset: restore the baseline active-tool set after the run settles and
+	// report how many run-local activations were actually used.
 	// ------------------------------------------------------------------
 	pi.on("agent_settled", () => {
 		const baseline = pi.getActiveTools().filter((name) => !eligibleTools(pi).some((tool) => tool.name === name));
 		pi.setActiveTools(baseline);
-		emitTelemetry(pi, "reset", { activeCount: baseline.length });
+		let used = 0;
+		for (const activation of activations.values()) if (activation.used) used += 1;
+		emitTelemetry(pi, "reset", { activeCount: baseline.length, activated: activations.size, used });
+		activations.clear();
 	});
 
 	// ------------------------------------------------------------------