@openlines/opl-pi-sht 0.2.11 → 0.3.0

package/extensions/opl-modes/index.ts

@@ -41,6 +41,7 @@ import {
   PLAN_DIR,
   buildExecutePrompt,
   buildRefinePrompt,
+  capPlan,
   USER_CONFIG,
   MODE_REGISTRY,
   getModeDefinition,
@@ -57,6 +58,7 @@ import {
   extractPlanText,
   isPlanLike,
   ensurePlanDir,
+  planFilePath,
   titleFromFilename,
   listPlanFiles,
   sanitizePlanName,
@@ -78,6 +80,10 @@ import {
   resetState,
   getRefineCount,
   incrementRefineCount,
+  publishModeGlobals,
+  clearModeGlobals,
+  getSessionCwd,
+  setSessionCwd,
 } from "./state.js";
 import { showSelectMenu } from "./menus.js";
 
@@ -260,7 +266,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
   function getPlanFilePath(): string | null {
     const file = getActivePlanFile();
     if (!file) return null;
-    return join(process.cwd(), PLAN_DIR, file);
+    return planFilePath(getSessionCwd(), file);
   }
 
   function getPlanDisplayTitle(): string | null {
@@ -269,7 +275,8 @@ export default function modeSwitcher(pi: ExtensionAPI) {
     const filePath = getPlanFilePath();
     if (filePath && existsSync(filePath)) {
       const heading = readFileSync(filePath, "utf-8").match(/^# Plan:\s*(.+)$/m);
-      if (heading) return heading[1].trim();
+      const captured = heading?.[1];
+      if (captured) return captured.trim();
     }
     return titleFromFilename(file);
   }
@@ -444,7 +451,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
   /** Open the mode picker. Handles all mode transitions including cross-mode switches. */
   async function openModePicker(ctx: ExtensionContext): Promise<void> {
     const current = getMode();
-    const planFiles = listPlanFiles();
+    const planFiles = listPlanFiles(getSessionCwd());
 
     // Build mode list dynamically from registry (skip invisible/disabled modes)
     const registryModes = Array.from(MODE_REGISTRY.entries())
@@ -551,6 +558,8 @@ export default function modeSwitcher(pi: ExtensionAPI) {
   // ─── Event: session_start ──────────────────────────────────────────────────────
 
   pi.on("session_start", async (event, ctx) => {
+    // Re-publish the footer seam: a previous session in this process cleared it on shutdown.
+    publishModeGlobals();
     if (event.reason === "startup") {
       if (pi.getFlag("chat") === true && getMode() === "off") {
         enterChatMode(ctx);
@@ -565,6 +574,12 @@ export default function modeSwitcher(pi: ExtensionAPI) {
     await syncStateFromBranch(ctx, event.reason === "resume" || event.reason === "fork");
   });
 
+  pi.on("session_shutdown", async () => {
+    // Hand the seam back: an orphaned mode global would keep rendering a mode that no
+    // longer belongs to any live session.
+    clearModeGlobals();
+  });
+
   // ─── Event: before_agent_start ──────────────────────────────────────────
 
   pi.on("before_agent_start", (event: BeforeAgentStartEvent): BeforeAgentStartEventResult => {
@@ -578,7 +593,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
       const filePath = getPlanFilePath();
       if (filePath && existsSync(filePath)) {
         const planContent = readFileSync(filePath, "utf-8");
-        return { systemPrompt: event.systemPrompt + "\n\n" + buildRefinePrompt(planContent) };
+        return { systemPrompt: event.systemPrompt + "\n\n" + buildRefinePrompt(planContent, filePath) };
       }
     }
     
@@ -590,12 +605,14 @@ export default function modeSwitcher(pi: ExtensionAPI) {
       const customTemplate = modeDef.prompt;
       if (typeof customTemplate === "string" && customTemplate.trim().length > 0) {
         // User-configured template via modes.execute.prompt — {plan} is replaced with the plan file content.
+        // Capped here too: a custom template must not become a way around the injection budget.
+        const capped = capPlan(planContent, USER_CONFIG.plan.maxInjectBytes, filePath).text;
         const filled = customTemplate.includes("{plan}")
-          ? customTemplate.replace("{plan}", planContent)
-          : `${customTemplate}\n\nPlan:\n${planContent}`;
+          ? customTemplate.replace("{plan}", capped)
+          : `${customTemplate}\n\nPlan:\n${capped}`;
         return { systemPrompt: event.systemPrompt + "\n\n" + filled };
       }
-      return { systemPrompt: event.systemPrompt + "\n\n" + buildExecutePrompt(planContent) };
+      return { systemPrompt: event.systemPrompt + "\n\n" + buildExecutePrompt(planContent, filePath) };
     }
     
     // Generic mode prompt from registry
@@ -628,7 +645,13 @@ export default function modeSwitcher(pi: ExtensionAPI) {
     if (event.toolName !== "bash") return {};
     if (!modeDef?.safePatterns && !modeDef?.destructivePatterns) return {};
 
-    const command = event.input.command as string;
+    if (typeof event.input.command !== "string") {
+      // Uninspectable commands are refused rather than coerced: String(["rm","-rf"]) would
+      // reach the pattern matchers as "rm,-rf" and slip past a segment-based safe list.
+      return { block: true, reason: `[mode-switcher] Blocked a bash tool call whose command was not a string [${mode} mode].` };
+    }
+
+    const command = event.input.command;
     const reason = bashBlockReason(command, modeDef.safePatterns, modeDef.destructivePatterns);
     if (reason) {
       return { block: true, reason: `[mode-switcher] Command blocked — ${reason} [${mode} mode]` };
@@ -699,7 +722,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
     setRefining(false);
 
     if (planText) {
-      const planDir = ensurePlanDir();
+      const planDir = ensurePlanDir(getSessionCwd());
       const activeFile = getActivePlanFile();
       let filename = activeFile;
       if (!filename) {
@@ -850,17 +873,20 @@ export default function modeSwitcher(pi: ExtensionAPI) {
       ctx.ui.notify(USER_CONFIG.labels.plan.notifyLoaded.replace("{title}", displayName), "info");
     }
 
-    const filePath = join(process.cwd(), PLAN_DIR, filename);
+    const filePath = planFilePath(getSessionCwd(), filename);
     if (existsSync(filePath)) {
       const planContent = readFileSync(filePath, "utf-8");
-      pi.appendEntry("plan-mode", { title: displayName, plan: planContent });
+      // The entry copy is display/session only (plain custom entries never reach the
+      // model), but it is written to the session file forever, so it gets its own budget.
+      const entry = capPlan(planContent, USER_CONFIG.plan.maxEntryBytes, filePath);
+      pi.appendEntry("plan-mode", { title: displayName, plan: entry.text, truncated: entry.truncated, file: filePath });
     }
   }
 
   // ─── Entry renderer for TUI-only plan cards (never sent to the model) ──────
 
   pi.registerEntryRenderer("plan-mode", (entry, _options, theme) => {
-    const data = (entry.data ?? {}) as { title?: string; plan?: string };
+    const data = (entry.data ?? {}) as { title?: string; plan?: string; truncated?: boolean; file?: string };
     const border = new DynamicBorder((s: string) => theme.fg("border", s));
     const container = new Container();
     container.addChild(border);
@@ -870,6 +896,9 @@ export default function modeSwitcher(pi: ExtensionAPI) {
       container.addChild(new Spacer());
       container.addChild(new Text(body));
     }
+    if (data.truncated) {
+      container.addChild(new Text(theme.fg("muted", `✂ truncated copy — full plan: ${data.file ?? PLAN_DIR}`)));
+    }
     container.addChild(border);
     return container;
   });
@@ -982,7 +1011,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
       if (!input) {
         const current = getMode();
         if (current === "off") {
-          const files = listPlanFiles();
+          const files = listPlanFiles(getSessionCwd());
           if (files.length === 0) {
             await promptNameAndEnterPlanMode(ctx);
           } else {
@@ -1025,7 +1054,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
         return;
       }
       const filename = `${PLAN_FILE_PREFIX}${sanitized}.md`;
-      const filePath = join(process.cwd(), PLAN_DIR, filename);
+      const filePath = planFilePath(getSessionCwd(), filename);
 
       if (existsSync(filePath)) {
         if (getMode() !== "off") enterOffMode(ctx, undefined, true);
@@ -1078,13 +1107,14 @@ export default function modeSwitcher(pi: ExtensionAPI) {
           pi.sendUserMessage("Execute the plan steps now.", { deliverAs: "followUp" });
           return;
         }
-        const files = listPlanFiles();
+        const files = listPlanFiles(getSessionCwd());
         if (files.length === 0) {
           if (ctx.hasUI) ctx.ui.notify("No plan files found. Run /plan first.", "warning");
           return;
         }
         if (files.length === 1) {
-          await startExecute(files[0].name);
+          const only = files[0];
+          if (only) await startExecute(only.name);
           return;
         }
         const items: SelectItem[] = files.map(f => ({ value: f.name, label: f.title }));
@@ -1100,7 +1130,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
         return;
       }
       const filename = `${PLAN_FILE_PREFIX}${sanitized}.md`;
-      const filePath = join(process.cwd(), PLAN_DIR, filename);
+      const filePath = planFilePath(getSessionCwd(), filename);
       if (!existsSync(filePath)) {
         if (ctx.hasUI) ctx.ui.notify(`Plan "${input}" not found. Check /plan for available plans.`, "warning");
         return;
@@ -1124,6 +1154,7 @@ export default function modeSwitcher(pi: ExtensionAPI) {
 
     const idx = order.indexOf(current);
     const next = idx === -1 ? order[0] : order[(idx + 1) % order.length];
+    if (!next) return; // order is non-empty above; this is what lets the branches below take a string
     if (next === current) return;
 
     if (next === "off") {

package/extensions/opl-modes/state.ts

@@ -7,12 +7,17 @@ import type { AgentMode, AgentModeBlob, ModeAppearanceConfig, ModeModelConfig }
 const state: {
   mode: AgentMode;
   activePlanFile: string | null;
+  sessionCwd: string;
   restoringModel: ModeModelConfig | null;
   refining: boolean;
   refineCount: number;
 } = {
   mode: "off",
   activePlanFile: null,
+  // Until the first `session_start` hands us the real one, the directory the harness was started
+  // in is the only candidate. Plans and the footer path must never be resolved from
+  // `process.cwd()` at the call site: a session opened elsewhere would write its plans there.
+  sessionCwd: process.cwd(),
   restoringModel: null,
   refining: false,
   refineCount: 0,
@@ -23,6 +28,23 @@ const state: {
 (globalThis as Record<string, unknown>).__planMode = { mode: "off" };
 (globalThis as Record<string, unknown>).__chatMode = { mode: "off" };
 
+/**
+ * Cross-extension seam. `opl-footer` publishes `__footerRequestRender` while a footer is
+ * mounted; `__agentMode`, `__planMode` and `__chatMode` are read by the footer's mode
+ * segment and by `opl-input`. Extensions install one directory at a time and load in any
+ * combination, so nothing here may assume the other side exists or is healthy: the value can
+ * be missing, be something else entirely, or close over a TUI that was already torn down.
+ */
+function requestFooterRender(): void {
+  const requestRender = (globalThis as Record<string, unknown>).__footerRequestRender;
+  if (typeof requestRender !== "function") return;
+  try {
+    (requestRender as () => void)();
+  } catch {
+    // A dead footer must not fail a mode transition.
+  }
+}
+
 /** Derive and sync all globalThis snapshots from unified state. */
 function syncGlobalThis(): void {
   const m = state.mode;
@@ -35,8 +57,23 @@ function syncGlobalThis(): void {
   // plan-family modes light up __planMode, so a custom mode cannot read as "Plan mode ON".
   (globalThis as Record<string, unknown>).__planMode = { mode: m === "plan" || m === "execute" ? m : "off" };
   (globalThis as Record<string, unknown>).__chatMode = { mode: m === "chat" ? "chat" : "off" };
-  const requestRender = (globalThis as Record<string, unknown>).__footerRequestRender;
-  if (typeof requestRender === "function") requestRender();
+  requestFooterRender();
+}
+
+/** Re-publish the seam after a fresh session starts in this process. */
+export function publishModeGlobals(): void {
+  syncGlobalThis();
+}
+
+/**
+ * Drop the seam when the session ends. A leftover value would let a footer in a later
+ * session render a mode that belongs to nobody, and the footer's own render trigger is
+ * removed with it (see `opl-footer`).
+ */
+export function clearModeGlobals(): void {
+  delete (globalThis as Record<string, unknown>).__agentMode;
+  delete (globalThis as Record<string, unknown>).__planMode;
+  delete (globalThis as Record<string, unknown>).__chatMode;
 }
 
 export function getMode(): AgentMode { return state.mode; }
@@ -126,6 +163,7 @@ export function restore(
 ): boolean {
   for (let i = entries.length - 1; i >= 0; i--) {
     const entry = entries[i];
+    if (!entry) continue; // a hole in the persisted array is not a restore candidate
     if (entry.type !== "custom" || entry.customType !== ENTRY_TYPE) continue;
 
     const data = entry.data as AgentModeBlob | undefined;
@@ -145,6 +183,7 @@ export function restore(
   // Backward compat: restore from legacy entries written by the old extensions.
   for (let i = entries.length - 1; i >= 0; i--) {
     const entry = entries[i];
+    if (!entry) continue; // a hole in the persisted array is not a restore candidate
     if (entry.type !== "custom") continue;
 
     if (entry.customType === "plan-mode") {
@@ -181,3 +220,12 @@ export function resetState(): void {
   state.refineCount = 0;
   syncGlobalThis();
 }
+
+/** The directory the current session lives in (`.pi/plans/` is resolved against this). */
+export function getSessionCwd(): string {
+  return state.sessionCwd;
+}
+
+export function setSessionCwd(cwd: string): void {
+  if (cwd) state.sessionCwd = cwd;
+}

package/extensions/opl-modes/types.ts

@@ -154,6 +154,12 @@ export interface ModeSwitcherUserConfig {
   shortcuts?: ShortcutsUserConfig;
   /** @deprecated Use `modes` registry instead. Kept for backward compat. */
   labels?: LabelsUserConfig;
+  plan?: {
+    /** Ceiling for a plan body injected into a system prompt, in bytes (default 24000). Larger plans truncate with a marker pointing at the file. */
+    maxInjectBytes?: number;
+    /** Ceiling for the plan copy kept in the `plan-mode` session entry, in bytes (default 4096). */
+    maxEntryBytes?: number;
+  };
   bashPatterns?: BashPatternsUserConfig;
   /** Replace-only: provide a list of tool names to replace the default CHAT mode tool set. Omit to keep defaults. */
   chatAllowedTools?: string[];

package/extensions/opl-modes/utils.ts

@@ -52,6 +52,7 @@ function commandSegments(command: string): string[] {
   for (let i = 0; i < command.length; i++) {
     const ch = command[i];
     const top = stack[stack.length - 1];
+    if (!top) break; // stack drained: the rest of the command is literal text
     // Single quotes are literal: nothing expands, only the closing ' matters.
     if (top.quote === "'") {
       if (ch === "'") top.quote = "";
@@ -165,9 +166,9 @@ export function extractPlanText(message: string): string | null {
 
 // ─── Plan File I/O ────────────────────────────────────────────────────────────
 
-/** Ensure .pi/plans/ exists, return its absolute path. */
-export function ensurePlanDir(): string {
-  const dir = join(process.cwd(), PLAN_DIR);
+/** Ensure <sessionCwd>/.pi/plans/ exists, return its absolute path. */
+export function ensurePlanDir(cwd: string): string {
+  const dir = join(cwd, PLAN_DIR);
   mkdirSync(dir, { recursive: true });
   return dir;
 }
@@ -181,7 +182,12 @@ export function titleFromFilename(filename: string): string {
   // Detect timestamp pattern: YYYY-MM-DD-HH-MM
   const tsMatch = stem.match(/^(\d{4})-(\d{2})-(\d{2})-(\d{2})-(\d{2})$/);
   if (tsMatch) {
-    const [, y, m, d, h, min] = tsMatch;
+    // Five digit groups by the shape test above; the explicit checks are what let the compiler see
+    // that, rather than a cast.
+    const [y, m, d, h, min] = tsMatch.slice(1);
+    if (y === undefined || m === undefined || d === undefined || h === undefined || min === undefined) {
+      return stem.replace(/-/g, " ").replace(/\b\w/g, (c) => c.toUpperCase());
+    }
     const months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec"];
     const month = months[parseInt(m, 10) - 1] ?? m;
     return `${month} ${parseInt(d, 10)}, ${y} ${h}:${min}`;
@@ -203,8 +209,8 @@ export function sanitizePlanName(name: string): string | null {
 }
 
 /** List available plan files in .pi/plans/ with titles from # Plan: heading. */
-export function listPlanFiles(): PlanFileSummary[] {
-  const dir = join(process.cwd(), PLAN_DIR);
+export function listPlanFiles(cwd: string): PlanFileSummary[] {
+  const dir = join(cwd, PLAN_DIR);
   if (!existsSync(dir)) return [];
 
   const files = readdirSync(dir)
@@ -214,7 +220,7 @@ export function listPlanFiles(): PlanFileSummary[] {
   return files.map((filename) => {
     const content = readFileSync(join(dir, filename), "utf-8");
     const titleMatch = content.match(/^# Plan:\s*(.+)$/m);
-    const title = titleMatch ? titleMatch[1].trim() : titleFromFilename(filename);
+    const title = titleMatch?.[1] ? titleMatch[1].trim() : titleFromFilename(filename);
     return { name: filename, title };
   });
 }
@@ -265,4 +271,8 @@ export function applyLabelColor(theme: Theme, color: string, text: string): stri
   } catch {
     return text;
   }
-}
+}
+/** Absolute path of a plan file inside the session directory. */
+export function planFilePath(cwd: string, filename: string): string {
+  return join(cwd, PLAN_DIR, filename);
+}

package/extensions/opl-questionnaire/index.ts

@@ -22,6 +22,7 @@ import {
 import {
 	type Answer,
 	type Question,
+	type QuestionCheck,
 	QuestionnaireParams,
 	type QuestionnaireResult,
 	type RenderOption,
@@ -42,7 +43,7 @@ export function errorResult(
 /** Validate the question list. Returns the first error message, or null when valid.
  *  Blank ids are rejected (answers are keyed by id and "" would produce an empty key),
  *  as are duplicate ids and questions with nothing selectable. */
-export function validateQuestions(questions: Question[]): string | null {
+export function validateQuestions(questions: readonly QuestionCheck[]): string | null {
 	const seenIds = new Set<string>();
 	for (const q of questions) {
 		if (q.id.trim() === "") {
@@ -303,8 +304,10 @@ export default function questionnaire(pi: ExtensionAPI) {
 						const tabs: string[] = ["← "];
 						for (let i = 0; i < questions.length; i++) {
 							const isActive = i === currentTab;
-							const isAnswered = answers.has(questions[i].id);
-							const lbl = questions[i].label;
+							const question = questions[i];
+							if (!question) continue;
+							const isAnswered = answers.has(question.id);
+							const lbl = question.label;
 							const box = isAnswered ? "■" : "□";
 							const color = isAnswered ? "success" : "muted";
 							const text = ` ${box} ${lbl} `;
@@ -326,6 +329,7 @@ export default function questionnaire(pi: ExtensionAPI) {
 					function renderOptions() {
 						for (let i = 0; i < opts.length; i++) {
 							const opt = opts[i];
+							if (!opt) continue; // a hole in the option list renders nothing rather than throwing mid-frame
 							const selected = i === optionIndex;
 							const isOther = opt.isOther === true;
 							const prefix = selected ? theme.fg("accent", "> ") : "  ";
@@ -435,7 +439,9 @@ export default function questionnaire(pi: ExtensionAPI) {
 		},
 
 		renderResult(result, _options, theme, _context) {
-			if (result.isError) {
+			// `isError` on tool results is newer than the declared floor: pi 0.87.x does not carry
+			// it in its types, and an older host leaves it undefined, which reads as "not an error".
+			if ((result as { isError?: boolean }).isError) {
 				const text = result.content[0];
 				return new Text(theme.fg("error", text?.type === "text" ? text.text : ""), 0, 0);
 			}

package/extensions/opl-questionnaire/types.ts

@@ -18,6 +18,10 @@ export interface Question {
 	allowOther: boolean;
 }
 
+/** The fields `validateQuestions` reads: the same shape on raw tool params (where `label`
+ *  still has to be defaulted) and on normalized questions. */
+export type QuestionCheck = Pick<Question, "id" | "options"> & Partial<Pick<Question, "allowOther">>;
+
 export interface Answer {
 	id: string;
 	value: string;

package/extensions/opl-simplebench/README.md

@@ -32,7 +32,7 @@ simplebench({ model: "global.openai.gpt-5.6-terra", coding_lite: true })
 simplebench({ model: "global.openai.gpt-5.6-terra", test_all: true })
 ```
 
-`--no-artifact` and `no_artifact: true` suppress the JSON file. Default runs use the provider's sampling and reasoning defaults and write one JSON report per model to Pi's current working directory. `--thinking-max` and `thinking_max: true` request `reasoning_effort: "max"` for OpenAI-compatible providers, or Pi's model-aware Bedrock max-thinking path for direct Bedrock models that advertise `reasoning: true` and `thinkingLevelMap.max`. Other providers reject that mode rather than silently using defaults.
+`--no-artifact` and `no_artifact: true` suppress the JSON file. Default runs use the provider's sampling and reasoning defaults and write one JSON report per model to the current Pi session directory (`ExtensionContext.cwd`). Without an active session context, artifacts fall back to the harness working directory. `--thinking-max` and `thinking_max: true` request `reasoning_effort: "max"` for OpenAI-compatible providers, or Pi's model-aware Bedrock max-thinking path for direct Bedrock models that advertise `reasoning: true` and `thinkingLevelMap.max`. Other providers reject that mode rather than silently using defaults.
 
 ## Extension features
 
@@ -172,7 +172,7 @@ The stats endpoint is metadata-only and does not change inference routing. Its s
 
 ## Artifacts and privacy
 
-Baseline and coding-lite artifacts are written to `process.cwd()` as JSON. `--test-all` writes a bundle:
+Baseline and coding-lite artifacts use the current Pi session directory (`ExtensionContext.cwd`), passed down explicitly; without a Pi session, they fall back to the harness working directory. `--test-all` writes a bundle:
 
 ```text
 simplebench--test-all-<sanitized-model>-<thinking>-<UTC-timestamp>/
@@ -211,6 +211,8 @@ The model receives weather and calculation tools. Simplebench validates expected
 
 Coding-lite is the execution-backed coding-agent suite. It runs model-directed edits in temporary directories. The model can list, search, read, write, and run the fixture's public tests, but cannot execute arbitrary commands, access the real repository, read hidden verification code, or use the network. The host runs hidden verification after the model stops.
 
+**The verifier executes model-authored code.** Verification runs `node --eval` over the model's files, so that process is started with an allowlisted environment (`util/exec-env.ts`): `PATH`, `HOME`, temp-directory and locale variables, TLS trust paths, terminal capability flags, and the Windows equivalents. Everything else is absent — every `*_API_KEY`, `AWS_*`, `GH_*`, `PI_*`, and any other secret your shell happens to export — and `NODE_OPTIONS`/`NODE_PATH` are never forwarded because both can make Node load attacker-chosen files. Anything the model's code spawns inherits that environment, so a nested `node src/cli.mjs` sees the same allowlist. This limits what generated code can *read out of the process*; it is not sandboxing. The child still runs with your user's filesystem permissions and network reach, and a `HOME` it can see points at files it could read by path anyway. Use `--no-artifact` and a dedicated shell if that is not acceptable for your environment.
+
 The agent loop defaults to **5 turns** per task (reduced from 12). This tighter budget forces the model to converge on a solution quickly and makes turn count a meaningful quality signal. Results are scored on three dimensions:
 
 - **STRONG** — solved in 1–2 turns (first or second attempt correct)
@@ -230,17 +232,17 @@ Use `--coding-lite` for coding tasks only. Use `--test-all` to run the existing
 
 ### Recommendation policy
 
-The recommendation is based on the three (or four) displayed categories:
+Four categories can be in play: the closed-answer contract, instruction following, tool usage, and (with `--coding-lite`) coding. A category counts as passed when its own rule succeeds - the closed-answer contract counts as passed at `STRONG` **or** `MODERATE`, and coding counts as passed on its efficiency-weighted average. The label then depends on how many passed, and on how many categories actually ran:
 
-| Result | Recommendation |
-|---|---|
-| Strong reasoning, JSON pass, tool pass, all coding STRONG | STRONG |
-| Strong reasoning, JSON pass, tool pass (coding not all STRONG) | GOOD |
-| Any three categories pass | USABLE |
-| Two categories pass | LIMITED |
-| One or zero categories pass | WEAK |
+| Categories passed | Baseline (3 categories) | With coding (4 categories) |
+|---|---|---|
+| All | STRONG if the contract is `STRONG` and, when coding ran, every task is `STRONG`; GOOD otherwise | |
+| All but one | GOOD | GOOD |
+| Two | USABLE | LIMITED |
+| One | LIMITED | LIMITED |
+| None | WEAK | WEAK |
 
-When coding-lite is included, the coding category uses efficiency-weighted scoring (STRONG=1.0, MODERATE=0.7, WEAK=0.4, FAIL=0). The category counts as passing when the weighted average ≥ 0.5 across all tasks.
+`WEAK` therefore requires that **nothing** passed; one passing category is `LIMITED`. A run that omits coding can report `USABLE` where the same results with coding report `LIMITED`, because the denominator changes. For runs with no clean category counts (an error or a cancelled run), `report.ts` falls back to the average `tests[].score`: `>= 0.8` `GOOD`, `>= 0.5` `USABLE`, `>= 0.2` `LIMITED`, otherwise `WEAK`. The coding category is scored efficiency-weighted (STRONG=1.0, MODERATE=0.7, WEAK=0.4, FAIL=0) and passes when that weighted average is >= 0.5 across all tasks.
 
 ## Provider authentication
 
@@ -260,6 +262,8 @@ Simplebench derives the region from the selected model's Bedrock base URL, signs
 aws configure export-credentials --profile "$AWS_PROFILE"
 ```
 
+The CLI is invoked with an argument vector, never through a shell, so the profile value is never interpolated into a command string; and it runs asynchronously, so a slow SSO or assume-role chain cannot stall the TUI. A profile name that is not a plain token (`^[A-Za-z0-9][A-Za-z0-9_.:@+-]{0,127}$`) is refused before any process is spawned — `AWS_PROFILE` set by repo tooling such as `.envrc` or a Makefile is not trusted to reach the CLI. Set `AWS_ACCESS_KEY_ID`/`AWS_SECRET_ACCESS_KEY` (plus `AWS_SESSION_TOKEN`) to bypass the CLI entirely.
+
 AWS CLI v2 is therefore required when the selected profile uses SSO or an assume-role chain.
 
 ## Troubleshooting
@@ -270,7 +274,7 @@ AWS CLI v2 is therefore required when the selected profile uses SSO or an assume
 | OpenRouter returns 401 | Set `OPENROUTER_API_KEY`, configure the provider key, or authenticate Pi so `auth.json` has an API key. |
 | Bedrock says credentials are unavailable | Run `aws configure export-credentials --profile "$AWS_PROFILE"` and resolve profile login/role issues first. |
 | Bedrock rejects `--thinking-max` | The selected model does not advertise max thinking in Pi metadata, or the provider rejected its model-specific reasoning request. |
-| Artifact is absent | Check the report for `--no-artifact` or an artifact-write warning; the file is written to Pi's cwd. |
+| Artifact is absent | Check the report for `--no-artifact` or an artifact-write warning; the file lands in the session directory, not where the shell was standing. |
 | Token or TTFT values are null | The provider response did not expose authoritative usage/timing values. |
 
 ## Architecture
@@ -279,13 +283,15 @@ AWS CLI v2 is therefore required when the selected profile uses SSO or an assume
 index.ts       Pi registration, command parsing, tool schema
 benchmark.ts   provider adapters and benchmark orchestration
 llama-server.ts direct llama-server /props and /metrics capture
+coding.ts      the six execution-backed coding tasks and their agent loop
+research.ts    grounded research fixtures and the deterministic verifier
 tests.ts       benchmark fixtures
 scoring.ts     extraction and scoring
 metrics.ts     request/usage aggregation
-artifact.ts    cwd JSON artifact writer
+artifact.ts    session-directory JSON artifact writer
 report.ts      terminal output and recommendation policy
 types.ts       shared benchmark types
-util/          local configuration, formatting, debugging, and provider helpers
+util/          local configuration, formatting, debugging, exec-environment scrubbing, and provider helpers
 ```
 
-The extension is self-contained under `extensions/opl-simplebench/`, ready to copy into the `opl-pi-sht` extension stack after local verification.
+Everything lives under `extensions/opl-simplebench/` because `install.sh` installs one extension directory at a time; the only environment boundary model-authored code gets is `util/exec-env.ts`, which is where a new child process must go through.

package/extensions/opl-simplebench/artifact.ts

@@ -1,5 +1,8 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
+
+// The output directory is always passed in: artifacts land in the session directory
+// (`ExtensionContext.cwd`), not wherever the harness was started.
 import type { RunArtifact } from "./types";
 
 export function artifactFileName(model: string, suite: "baseline" | "coding-lite" | "test-all" = "baseline", thinking: "default" | "max" = "default", date = new Date(), tag?: string): string {
@@ -9,10 +12,10 @@ export function artifactFileName(model: string, suite: "baseline" | "coding-lite
   return `simplebench${tag ? `-${tag}-` : "--"}${suiteName}-${safeModel}-${thinking}-${stamp}.json`;
 }
 
-export function writeArtifactBundle(artifact: RunArtifact, files: Record<string, string>): string {
+export function writeArtifactBundle(artifact: RunArtifact, files: Record<string, string>, cwd: string): string {
   const name = artifactFileName(artifact.benchmark.model, artifact.benchmark.suite || "baseline", artifact.benchmark.thinking.requested, new Date(), artifact.benchmark.tag).replace(/\.json$/, "");
   for (let suffix = 0; ; suffix += 1) {
-    const output = path.resolve(process.cwd(), suffix ? `${name}-${suffix}` : name);
+    const output = path.resolve(cwd, suffix ? `${name}-${suffix}` : name);
     try {
       fs.mkdirSync(output);
       fs.writeFileSync(path.join(output, "result.json"), JSON.stringify(artifact, null, 2) + "\n", "utf8");
@@ -24,13 +27,13 @@ export function writeArtifactBundle(artifact: RunArtifact, files: Record<string,
   }
 }
 
-export function writeArtifact(artifact: RunArtifact): string {
+export function writeArtifact(artifact: RunArtifact, cwd: string): string {
   const name = artifactFileName(artifact.benchmark.model, artifact.benchmark.suite || "baseline", artifact.benchmark.thinking.requested, new Date(), artifact.benchmark.tag);
   const extension = path.extname(name);
   const base = name.slice(0, -extension.length);
   const contents = JSON.stringify(artifact, null, 2) + "\n";
   for (let suffix = 0; ; suffix += 1) {
-    const output = path.resolve(process.cwd(), suffix ? `${base}-${suffix}${extension}` : name);
+    const output = path.resolve(cwd, suffix ? `${base}-${suffix}${extension}` : name);
     try {
       fs.writeFileSync(output, contents, { encoding: "utf8", flag: "wx" });
       return output;