@pencil-agent/nano-pencil 1.13.5 → 1.13.7

package/dist/build-meta.json

@@ -0,0 +1,6 @@
+{
+  "version": "1.13.7",
+  "commitHash": "9b46faa",
+  "branch": "main",
+  "builtAt": "2026-04-25T14:10:25.443Z"
+}

package/dist/core/model-registry.js

@@ -207,8 +207,9 @@ export class ModelRegistry {
             // Keep built-in models even if custom models failed to load
         }
         const builtInModels = this.useOnlyCustomModels
-            ? this.loadBuiltInModels(overrides, modelOverrides, new Set(["openrouter"]), {
+            ? this.loadBuiltInModels(overrides, modelOverrides, new Set(["openrouter", "zai"]), {
                 openrouter: new Set(NANOPENCIL_OPENROUTER_BUILTIN_MODEL_IDS),
+                // zai not specified = load all zai models
             })
             : this.loadBuiltInModels(overrides, modelOverrides);
         let combined = this.mergeCustomModels(builtInModels, customModels);

package/dist/core/runtime/agent-session.js

@@ -1183,8 +1183,22 @@ export class AgentSession {
         this.agent.setModel(model);
         this.sessionManager.appendModelChange(model.provider, model.id);
         this.settingsManager.setDefaultModelAndProvider(model.provider, model.id);
-        // Re-clamp thinking level for new model's capabilities
-        this.setThinkingLevel(this.thinkingLevel);
+        // Auto-select thinking level based on model capabilities
+        const currentLevel = this.thinkingLevel;
+        let newLevel;
+        if (!model.reasoning) {
+            // Model doesn't support thinking, force off
+            newLevel = "off";
+        }
+        else if (currentLevel === "off") {
+            // Model supports thinking but current level is off, default to medium
+            newLevel = "medium";
+        }
+        else {
+            // Keep current level but clamp to new model's capabilities
+            newLevel = currentLevel;
+        }
+        this.setThinkingLevel(newLevel);
         await this._emitModelSelect(model, previousModel, "set");
     }
     /**
@@ -1290,8 +1304,19 @@ export class AgentSession {
         this.agent.setModel(nextModel);
         this.sessionManager.appendModelChange(nextModel.provider, nextModel.id);
         this.settingsManager.setDefaultModelAndProvider(nextModel.provider, nextModel.id);
-        // Re-clamp thinking level for new model's capabilities
-        this.setThinkingLevel(this.thinkingLevel);
+        // Auto-select thinking level based on model capabilities
+        const currentLevel = this.thinkingLevel;
+        let newLevel;
+        if (!nextModel.reasoning) {
+            newLevel = "off";
+        }
+        else if (currentLevel === "off") {
+            newLevel = "medium";
+        }
+        else {
+            newLevel = currentLevel;
+        }
+        this.setThinkingLevel(newLevel);
         await this._emitModelSelect(nextModel, currentModel, "cycle");
         return {
             model: nextModel,

package/dist/extensions/defaults/AGENT.md

@@ -44,7 +44,7 @@ sal/weights.ts: SalWeights interface, SAL_DEFAULT_WEIGHTS, loadSalWeights() read
 sal/eval/index.ts: createEvalSink() factory + barrel re-exports; adapter selection via options.adapter or endpoint scheme inference (http(s)→insforge, file://|/|./|../→jsonl, missing→noop); ONLY entry point SAL imports from
 sal/eval/types.ts: EvalSink interface, EvalEventEnvelope/EvalEventType (run_start/run_end/turn_anchor), EvalAdapterId ("insforge"|"jsonl"|"noop"), CreateEvalSinkOptions, createEvalEvent factory; zero-dependency type surface
 sal/eval/noop-sink.ts: noopSink — silent EvalSink used when eval disabled or no adapter configured
-sal/eval/insforge-sink.ts: InsForgeEvalSink — PostgREST adapter, routes run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, run_end→eval_runs PATCH; allowSelfSigned TLS option, batching with default 2000ms interval
+sal/eval/insforge-sink.ts: InsForgeEvalSink — PostgREST adapter, routes run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, tool_trace→eval_tool_traces with PGRST204 legacy-schema fallback, memory_recalls→eval_memory_recalls batch INSERT, run_end→eval_runs PATCH; allowSelfSigned TLS option, batching with default 2000ms interval
 sal/eval/jsonl-sink.ts: JsonlEvalSink — append-only filesystem adapter, one JSON object per line, accepts file:// URLs or plain paths, auto-creates parent dir, batched writes
 sal/README.md: SAL extension usage, sidecar output layout, weights override, pluggability contract
 team/index.ts: AgentTeam extension entry, /team:/team:spawn/:send/:status/:stop/:terminate/:approve/:mode commands, TEAM_MESSAGE_TYPE renderer

package/dist/extensions/defaults/CLAUDE.md

@@ -42,14 +42,14 @@ plan/exit-plan-mode-tool.ts: createExitPlanModeTool() - ExitPlanMode tool with p
 plan/plan-agents.ts: Explore/Plan subagent definitions with read-only tools for plan mode workflow
 plan/plan-validation.ts: validatePlan() - validates plan has required sections (Context, Approach, Files, Verification)
 plan/teammate-approval.ts: isInTeammateContext(), submitPlanToLeader(), formatPlanSubmittedMessage() - teammate plan approval integration
-sal/index.ts: SAL extension entry, enabled by default, registers --nosal/--sal-rebuild-terrain flags, /sal:coverage /sal:status /sal:setup commands, before_agent_start/tool_execution_start/agent_end hooks; /sal:setup writes ~/.memory-experiments/credentials.json with adapter inference (insforge/jsonl/noop); publishes structuralAnchor via core/runtime/turn-context (no SAL-specific globals); emits run_start/turn_anchor/memory_recalls/run_end eval events through pluggable EvalSink; reads memoryRecallSnapshot from turn-context bus in agent_end; runtime no-op when --nosal is set
+sal/index.ts: SAL extension entry, enabled by default, registers --nosal/--sal-rebuild-terrain flags, /sal:coverage /sal:status /sal:setup commands, before_agent_start/tool_execution_start/tool_execution_end/agent_end hooks; /sal:setup writes ~/.memory-experiments/credentials.json with adapter inference (insforge/jsonl/noop); publishes structuralAnchor via core/runtime/turn-context (no SAL-specific globals); emits run_start/turn_anchor/memory_recalls/tool_trace/run_end eval events through pluggable EvalSink; reads memoryRecallSnapshot from turn-context bus in agent_end; runtime no-op when --nosal is set; auto-injects pencil_version from build-meta.json into run_start; emergency flush on beforeExit/SIGHUP/SIGTERM; stale run cleanup is opt-in via NANOPENCIL_EVAL_CLEANUP_STALE_RUNS / credentials cleanup_stale_runs; tool_trace is a bounded per-turn summary and includes no-tool turns
 sal/terrain.ts: TerrainSnapshot/TerrainNode/TerrainEdge model, async buildTerrainIndex()/isSnapshotStale() (fs/promises + periodic yields so TUI can flush under block terminals like Warp), checkDipCoverage(), moduleIdForPath(), parses P2 CLAUDE.md and P3 file headers
 sal/anchors.ts: StructuralAnchor/AnchorResolution model, locateTask(), locateAction(), evidence-driven scoring with tunable SalWeights, CJK bigram tokenization
 sal/weights.ts: SalWeights interface, SAL_DEFAULT_WEIGHTS, loadSalWeights() reads sal-config.json from workspace or .memory-experiments/sal/
 sal/eval/index.ts: createEvalSink() factory + barrel re-exports; adapter selection via options.adapter or endpoint scheme inference (http(s)→insforge, file://|/|./|../→jsonl, missing→noop); ONLY entry point SAL imports from
-sal/eval/types.ts: EvalSink interface, EvalEventEnvelope/EvalEventType (run_start/run_end/turn_anchor/memory_recalls), EvalAdapterId ("insforge"|"jsonl"|"noop"), CreateEvalSinkOptions, createEvalEvent factory; zero-dependency type surface
+sal/eval/types.ts: EvalSink interface, EvalEventEnvelope/EvalEventType (run_start/run_end/turn_anchor/memory_recalls/tool_trace), EvalAdapterId ("insforge"|"jsonl"|"noop"), CreateEvalSinkOptions, createEvalEvent factory; zero-dependency type surface
 sal/eval/noop-sink.ts: noopSink — silent EvalSink used when eval disabled or no adapter configured
-sal/eval/insforge-sink.ts: InsForgeEvalSink — PostgREST adapter, routes run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, memory_recalls→eval_memory_recalls batch INSERT, run_end→eval_runs PATCH; allowSelfSigned TLS option, batching with default 2000ms interval
+sal/eval/insforge-sink.ts: InsForgeEvalSink — PostgREST adapter, routes run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, tool_trace→eval_tool_traces bounded per-turn summaries (including no-tool turns and truncation counters), memory_recalls→eval_memory_recalls batch INSERT, run_end→eval_runs PATCH; allowSelfSigned TLS option, batching with default 2000ms interval
 sal/eval/jsonl-sink.ts: JsonlEvalSink — append-only filesystem adapter, one JSON object per line, accepts file:// URLs or plain paths, auto-creates parent dir, batched writes
 sal/README.md: SAL extension usage, sidecar output layout, weights override, pluggability contract
 team/index.ts: AgentTeam extension entry, /team:/team:spawn/:send/:status/:stop/:terminate/:approve/:mode commands, TEAM_MESSAGE_TYPE renderer
@@ -64,4 +64,4 @@ team/TESTING.md: Manual & smoke-test guide for Phase B AgentTeam
 
 Rule: Members complete, one item per line, parent links valid, precise terms first
 
-[COVENANT]: Update this file header on changes and verify against parent CLAUDE.md
+[COVENANT]: Update this file header on changes and verify against parent CLAUDE.md

package/dist/extensions/defaults/sal/eval/insforge-sink.d.ts

@@ -2,7 +2,7 @@
  * [WHO]: Provides InsForgeEvalSink (PostgREST-backed adapter)
  * [FROM]: Depends on node:https, node:http, node:url; ./types.js for EvalSink/EvalEventEnvelope/CreateEvalSinkOptions
  * [TO]: Constructed by eval/index.ts factory when adapter resolves to "insforge"
- * [HERE]: extensions/defaults/sal/eval/insforge-sink.ts - InsForge-specific routing: run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, run_end→eval_runs PATCH
+ * [HERE]: extensions/defaults/sal/eval/insforge-sink.ts - InsForge-specific routing: run_start→eval_runs INSERT (merge-duplicates, includes pencil_version), turn_anchor→eval_turns + eval_sal_anchors×2, tool_trace→eval_tool_traces with legacy-schema fallback, memory_recalls→eval_memory_recalls, run_end→eval_runs PATCH
  *
  * Pluggable: nothing in this file may be imported from outside the eval/ directory.
  * To add a new backend, write a sibling file with the same EvalSink interface.
@@ -29,6 +29,7 @@ export declare class InsForgeEvalSink implements EvalSink {
     private handleTurnAnchor;
     private handleRunEnd;
     private handleMemoryRecalls;
+    private handleToolTrace;
     private postJson;
     private patchJson;
     private httpJson;

package/dist/extensions/defaults/sal/eval/insforge-sink.js

@@ -2,7 +2,7 @@
  * [WHO]: Provides InsForgeEvalSink (PostgREST-backed adapter)
  * [FROM]: Depends on node:https, node:http, node:url; ./types.js for EvalSink/EvalEventEnvelope/CreateEvalSinkOptions
  * [TO]: Constructed by eval/index.ts factory when adapter resolves to "insforge"
- * [HERE]: extensions/defaults/sal/eval/insforge-sink.ts - InsForge-specific routing: run_start→eval_runs INSERT (merge-duplicates), turn_anchor→eval_turns + eval_sal_anchors×2, run_end→eval_runs PATCH
+ * [HERE]: extensions/defaults/sal/eval/insforge-sink.ts - InsForge-specific routing: run_start→eval_runs INSERT (merge-duplicates, includes pencil_version), turn_anchor→eval_turns + eval_sal_anchors×2, tool_trace→eval_tool_traces with legacy-schema fallback, memory_recalls→eval_memory_recalls, run_end→eval_runs PATCH
  *
  * Pluggable: nothing in this file may be imported from outside the eval/ directory.
  * To add a new backend, write a sibling file with the same EvalSink interface.
@@ -103,6 +103,9 @@ export class InsForgeEvalSink {
                 case "memory_recalls":
                     await this.handleMemoryRecalls(event);
                     break;
+                case "tool_trace":
+                    await this.handleToolTrace(event);
+                    break;
                 case "run_end":
                     await this.handleRunEnd(event);
                     break;
@@ -123,6 +126,7 @@ export class InsForgeEvalSink {
                 task_file: strOrNull(p.task_file),
                 model: strOrNull(p.model),
                 thinking: p.thinking === true,
+                pencil_version: strOrNull(p.pencil_version),
                 commit_hash: strOrNull(p.commit, "unknown"),
                 branch_name: strOrNull(p.branch, "unknown"),
                 workspace_root: strOrNull(p.workspace_root),
@@ -213,6 +217,39 @@ export class InsForgeEvalSink {
         }));
         await this.postJson(`${this.base}/api/database/records/eval_memory_recalls`, rows, { prefer: "resolution=ignore-duplicates" });
     }
+    // INSERT into eval_tool_traces — one row per turn with tool usage summary
+    // InsForge columns are all TEXT; JSONB fields must be serialized to strings.
+    async handleToolTrace(ev) {
+        const p = ev.payload;
+        const taskSignals = p.task_signals;
+        const row = {
+            run_id: ev.run_id,
+            turn_id: String(p.turn_id ?? 0),
+            event_id: ev.event_id,
+            tool_calls: p.tool_calls ? JSON.stringify(p.tool_calls) : null,
+            tool_sequence: p.tool_sequence ? JSON.stringify(p.tool_sequence) : null,
+            intent: strOrNull(taskSignals?.intent),
+            prompt_length: String(taskSignals?.prompt_length ?? 0),
+            has_error_trace: String(taskSignals?.has_error_trace === true),
+            has_file_reference: String(taskSignals?.has_file_reference === true),
+            has_tool_usage: String(p.has_tool_usage === true),
+            total_tool_calls: String(p.total_tool_calls ?? 0),
+            total_errors: String(p.total_errors ?? 0),
+            completed_tool_calls: String(p.completed_tool_calls ?? 0),
+            truncated_tool_calls: String(p.truncated_tool_calls ?? 0),
+            truncated_tool_summary: String(p.truncated_tool_summary ?? 0),
+            duration_ms: String(p.duration_ms ?? 0),
+            recorded_at: ev.ts,
+        };
+        const url = `${this.base}/api/database/records/eval_tool_traces`;
+        const result = await this.postJson(url, [row], {
+            prefer: "resolution=ignore-duplicates",
+            quietErrorCodes: ["PGRST204"],
+        });
+        if (!result.ok && result.errorCode === "PGRST204") {
+            await this.postJson(url, [toLegacyToolTraceRow(row)], { prefer: "resolution=ignore-duplicates" });
+        }
+    }
     // ------------------------------------------------------------------
     // HTTP helpers
     // ------------------------------------------------------------------
@@ -220,12 +257,12 @@ export class InsForgeEvalSink {
         const extraHeaders = {};
         if (extra?.prefer)
             extraHeaders["Prefer"] = extra.prefer;
-        return this.httpJson("POST", url, body, extraHeaders);
+        return this.httpJson("POST", url, body, extraHeaders, extra?.quietErrorCodes);
     }
     patchJson(url, body) {
         return this.httpJson("PATCH", url, body, {});
     }
-    httpJson(method, url, body, extraHeaders) {
+    httpJson(method, url, body, extraHeaders, quietErrorCodes = []) {
         return new Promise((resolve) => {
             const payload = JSON.stringify(body);
             let parsed;
@@ -234,7 +271,7 @@ export class InsForgeEvalSink {
             }
             catch {
                 console.error(`[sal][eval] invalid URL: ${url}`);
-                resolve(false);
+                resolve({ ok: false });
                 return;
             }
             const isHttps = parsed.protocol === "https:";
@@ -258,20 +295,21 @@ export class InsForgeEvalSink {
                 res.on("data", (chunk) => { rawBody += chunk; });
                 res.on("end", () => {
                     const ok = res.statusCode !== undefined && res.statusCode < 300;
-                    if (!ok) {
+                    const errorCode = parsePostgrestErrorCode(rawBody);
+                    if (!ok && !quietErrorCodes.includes(errorCode ?? "")) {
                         console.error(`[sal][eval] HTTP ${res.statusCode} ${method} ${parsed.pathname} — ${rawBody.slice(0, 300)}`);
                     }
-                    resolve(ok);
+                    resolve({ ok, statusCode: res.statusCode, body: rawBody, errorCode });
                 });
             });
             req.on("error", (err) => {
                 console.error(`[sal][eval] network error → ${parsed.hostname}: ${err.message}`);
-                resolve(false);
+                resolve({ ok: false });
             });
             req.on("timeout", () => {
                 console.error(`[sal][eval] timeout ${method} ${parsed.pathname}`);
                 req.destroy();
-                resolve(false);
+                resolve({ ok: false });
             });
             req.write(payload);
             req.end();
@@ -292,3 +330,16 @@ function numOrNull(v) {
     const n = Number(v);
     return isNaN(n) ? null : n;
 }
+function parsePostgrestErrorCode(rawBody) {
+    try {
+        const parsed = JSON.parse(rawBody);
+        return typeof parsed?.code === "string" ? parsed.code : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+function toLegacyToolTraceRow(row) {
+    const { has_tool_usage: _hasToolUsage, completed_tool_calls: _completedToolCalls, truncated_tool_calls: _truncatedToolCalls, truncated_tool_summary: _truncatedToolSummary, ...legacyRow } = row;
+    return legacyRow;
+}

package/dist/extensions/defaults/sal/eval/types.d.ts

@@ -5,7 +5,7 @@
  * [HERE]: extensions/defaults/sal/eval/types.ts - transport-agnostic event types and the EvalSink contract; concrete adapters live in sibling files
  */
 export type EvalVariant = "sal" | "control" | "baseline";
-export type EvalEventType = "run_start" | "run_end" | "turn_anchor" | "memory_recalls";
+export type EvalEventType = "run_start" | "run_end" | "turn_anchor" | "memory_recalls" | "tool_trace";
 /** Wire format for eval events. Adapter implementations decide how to materialize. */
 export interface EvalEventEnvelope {
     run_id: string;

package/dist/extensions/defaults/sal/index.d.ts

@@ -1,12 +1,46 @@
 /**
- * [WHO]: SAL extension entry - enabled by default, registers --nosal/--sal-rebuild-terrain flags, /sal:coverage /sal:status /sal:setup commands, before_agent_start/tool_execution_start/agent_end hooks; runtime no-op when --nosal is set
- * [FROM]: Depends on core/extensions/types.ts, core/runtime/turn-context.ts (publishes structuralAnchor), extensions/defaults/sal/terrain.ts, anchors.ts, weights.ts, eval/index.ts (pluggable adapters)
+ * [WHO]: SAL extension entry - enabled by default, registers --nosal/--sal-rebuild-terrain flags, /sal:coverage /sal:status /sal:setup commands, before_agent_start/tool_execution_start/tool_execution_end/agent_end hooks; runtime no-op when --nosal is set
+ * [FROM]: Depends on core/extensions/types.ts (ToolExecutionStartEvent, ToolExecutionEndEvent), core/runtime/turn-context.ts (publishes structuralAnchor), extensions/defaults/sal/terrain.ts, anchors.ts, weights.ts, eval/index.ts (pluggable adapters)
  * [TO]: Loaded by builtin-extensions.ts as a default extension entry point
- * [HERE]: extensions/defaults/sal/index.ts - pluggable Structural Anchor Localization (SAL) extension; emits run_start/turn_anchor/run_end eval events; /sal:setup writes ~/.memory-experiments/credentials.json
+ * [HERE]: extensions/defaults/sal/index.ts - pluggable Structural Anchor Localization (SAL) extension; emits run_start/turn_anchor/tool_trace/run_end eval events; tool_trace captures per-turn tool usage profile (call counts, sequences, intent, errors) for self-awareness analytics
  */
+import { type EvalAdapterId } from "./eval/index.js";
 import type { ExtensionAPI } from "../../../core/extensions/types.js";
+import { type AnchorResolution } from "./anchors.js";
 import { SAL_DEFAULT_WEIGHTS } from "./weights.js";
+interface ToolCallRecord {
+    toolCallId: string;
+    tool: string;
+    startMs: number;
+    endMs?: number;
+    isError?: boolean;
+}
+interface TurnState {
+    turnId: number;
+    startedAtMs: number;
+    taskResolution?: AnchorResolution;
+    touchedFiles: Set<string>;
+    toolCalls: ToolCallRecord[];
+    prompt?: string;
+}
+interface EvalCredentials {
+    insforge_url?: string;
+    endpoint?: string;
+    api_key?: string;
+    anon_key?: string;
+    api_key_header?: string;
+    headers?: Record<string, string>;
+    enabled?: boolean;
+    allow_self_signed?: boolean;
+    cleanup_stale_runs?: boolean;
+    /** Adapter selector. When omitted, inferred from endpoint scheme (http→insforge, file/path→jsonl). */
+    adapter?: EvalAdapterId;
+}
+declare function resolveStaleCleanupEnabled(envValue: string | undefined, credentials: EvalCredentials | undefined): boolean;
 declare function normalizeExperimentId(experimentId?: string): string | undefined;
 declare function resolveSalSidecarDir(workspaceRoot: string, experimentId?: string): string;
+type TaskIntent = "fix" | "feat" | "refactor" | "explain" | "explore" | "unknown";
+declare function inferIntent(prompt: string): TaskIntent;
+declare function buildToolTracePayload(turn: TurnState, turnDuration: number): Record<string, unknown>;
 export default function salExtension(api: ExtensionAPI): Promise<void>;
-export { SAL_DEFAULT_WEIGHTS, normalizeExperimentId, resolveSalSidecarDir };
+export { SAL_DEFAULT_WEIGHTS, buildToolTracePayload, inferIntent, normalizeExperimentId, resolveSalSidecarDir, resolveStaleCleanupEnabled, };