taskplane 0.8.2 → 0.9.1

package/dashboard/public/app.js

@@ -530,18 +530,19 @@ function renderLanesTasks(batch, tmuxSessions) {
       // Worker stats from lane state sidecar + telemetry badges
       let workerHtml = "";
       const telemBadges = task.status !== "pending" ? telemetryBadgesHtml(tel) : "";
+      const reviewerActive = ls && ls.reviewerStatus === "running";
       if (ls && ls.workerStatus === "running" && task.status === "running") {
         const elapsed = ls.workerElapsed ? `${Math.round(ls.workerElapsed / 1000)}s` : "";
         const tools = ls.workerToolCount || 0;
         const ctx = ls.workerContextPct ? `${Math.round(ls.workerContextPct)}%` : "";
-        const lastTool = ls.workerLastTool || "";
+        const lastTool = reviewerActive ? "[awaiting review]" : (ls.workerLastTool || "");
         const tokenStr = tokenSummaryFromLaneState(ls);
         workerHtml = `<div class="worker-stats">`;
         workerHtml += `<span class="worker-stat" title="Worker elapsed">⏱ ${elapsed}</span>`;
         workerHtml += `<span class="worker-stat" title="Tool calls">🔧 ${tools}</span>`;
         if (ctx) workerHtml += `<span class="worker-stat" title="Context window used">📊 ${ctx}</span>`;
         if (tokenStr) workerHtml += `<span class="worker-stat" title="Tokens: input↑ output↓ cacheRead(R) cacheWrite(W)">🪙 ${tokenStr}</span>`;
-        if (lastTool) workerHtml += `<span class="worker-stat worker-last-tool" title="Last tool call">${escapeHtml(lastTool)}</span>`;
+        if (lastTool) workerHtml += `<span class="worker-stat worker-last-tool" title="${reviewerActive ? 'Waiting for reviewer' : 'Last tool call'}">${reviewerActive ? '<span style="color:var(--yellow)">' + escapeHtml(lastTool) + '</span>' : escapeHtml(lastTool)}</span>`;
         workerHtml += telemBadges;
         workerHtml += `</div>`;
       } else if (!ls && tel && task.status === "running") {
@@ -560,6 +561,35 @@ function renderLanesTasks(batch, tmuxSessions) {
         workerHtml = `<div class="worker-stats">${telemBadges}</div>`;
       }
 
+      // Reviewer sub-row: shown when reviewer is actively running
+      let reviewerRowHtml = "";
+      if (reviewerActive) {
+        const rElapsed = ls.reviewerElapsed ? `${Math.round(ls.reviewerElapsed / 1000)}s` : "";
+        const rTools = ls.reviewerToolCount || 0;
+        const rCtx = ls.reviewerContextPct ? `${Math.round(ls.reviewerContextPct)}%` : "";
+        const rLastTool = ls.reviewerLastTool || "";
+        const rCost = ls.reviewerCostUsd ? `$${ls.reviewerCostUsd.toFixed(2)}` : "";
+        const rType = ls.reviewerType || "review";
+        const rStep = ls.reviewerStep || "?";
+        reviewerRowHtml = `
+          <div class="task-row reviewer-sub-row">
+            <span class="task-icon"></span>
+            <span class="task-actions"></span>
+            <span class="reviewer-label">📋 Reviewer</span>
+            <span class="reviewer-type">${escapeHtml(rType)} · Step ${rStep}</span>
+            <span class="task-duration">${rElapsed}</span>
+            <span></span>
+            <span class="task-step">
+              <div class="worker-stats reviewer-stats">
+                <span class="worker-stat" title="Reviewer tool calls">🔧 ${rTools}</span>
+                ${rCtx ? `<span class="worker-stat" title="Reviewer context used">📊 ${rCtx}</span>` : ""}
+                ${rCost ? `<span class="worker-stat" title="Reviewer cost">${rCost}</span>` : ""}
+                ${rLastTool ? `<span class="worker-stat worker-last-tool" title="Reviewer last tool">${escapeHtml(rLastTool)}</span>` : ""}
+              </div>
+            </span>
+          </div>`;
+      }
+
       const isViewingStatus = viewerMode === 'status-md' && viewerTarget === task.taskId;
       const eyeHtml = task.status !== 'pending'
         ? `<button class="viewer-eye-btn${isViewingStatus ? ' active' : ''}" onclick="viewStatusMd('${escapeHtml(task.taskId)}')" title="View STATUS.md">👁</button>`
@@ -575,6 +605,7 @@ function renderLanesTasks(batch, tmuxSessions) {
           <span>${progressHtml}</span>
           <span class="task-step">${stepHtml}${workerHtml}</span>
         </div>`;
+      html += reviewerRowHtml;
     }
 
     html += `</div>`; // close lane-group

package/dashboard/public/style.css

@@ -631,6 +631,32 @@ body {
   text-overflow: ellipsis;
 }
 
+/* ─── Reviewer Sub-Row ─────────────────────────────────────────────── */
+
+.reviewer-sub-row {
+  background: var(--bg-surface-hover);
+  border-left: 3px solid var(--yellow, #e5c07b);
+  padding-left: 11px; /* compensate for border */
+  font-size: 0.85em;
+}
+
+.reviewer-sub-row .reviewer-label {
+  font-family: var(--font-mono);
+  font-size: 0.75rem;
+  color: var(--yellow, #e5c07b);
+  font-weight: 600;
+}
+
+.reviewer-sub-row .reviewer-type {
+  font-family: var(--font-mono);
+  font-size: 0.7rem;
+  color: var(--text-muted);
+}
+
+.reviewer-stats {
+  margin-top: 0;
+}
+
 /* ─── Telemetry Badges (retry, compaction) ─────────────────────────────── */
 
 .telem-badge {

package/dashboard/server.cjs

@@ -740,12 +740,15 @@ function computeBatchTotalCost(laneStates, telemetry) {
   let totalCost = 0;
   const coveredPrefixes = new Set();
 
-  // Primary: sum cost from lane states
+  // Primary: sum cost from lane states (worker + reviewer)
   for (const [prefix, ls] of Object.entries(laneStates)) {
     if (ls.workerCostUsd) {
       totalCost += ls.workerCostUsd;
       coveredPrefixes.add(prefix);
     }
+    if (ls.reviewerCostUsd) {
+      totalCost += ls.reviewerCostUsd;
+    }
   }
 
   // Supplementary: add cost from telemetry for uncovered lanes only

package/extensions/task-runner.ts

@@ -19,6 +19,7 @@
 
 import type { ExtensionAPI, ExtensionContext } from "@mariozechner/pi-coding-agent";
 import { DynamicBorder } from "@mariozechner/pi-coding-agent";
+import { Type } from "@mariozechner/pi-ai";
 import { Container, Text, truncateToWidth } from "@mariozechner/pi-tui";
 import { spawn, spawnSync } from "child_process";
 import {
@@ -128,8 +129,17 @@ interface TaskState {
 	workerExitDiagnostic: TaskExitDiagnostic | null;
 	reviewerStatus: "idle" | "running" | "done" | "error";
 	reviewerType: string;
+	reviewerStep: number;
+	reviewerSessionName: string;
 	reviewerElapsed: number;
 	reviewerLastTool: string;
+	reviewerToolCount: number;
+	reviewerInputTokens: number;
+	reviewerOutputTokens: number;
+	reviewerCacheReadTokens: number;
+	reviewerCacheWriteTokens: number;
+	reviewerCostUsd: number;
+	reviewerContextPct: number;
 	reviewerProc: any;
 	reviewerTimer: any;
 	reviewCounter: number;
@@ -146,8 +156,10 @@ function freshState(): TaskState {
 		workerProc: null, workerTimer: null,
 		workerRetryActive: false, workerRetryCount: 0, workerLastRetryError: "",
 		workerExitDiagnostic: null,
-		reviewerStatus: "idle", reviewerType: "", reviewerElapsed: 0,
-		reviewerLastTool: "", reviewerProc: null, reviewerTimer: null,
+		reviewerStatus: "idle", reviewerType: "", reviewerStep: 0, reviewerSessionName: "",
+		reviewerElapsed: 0, reviewerLastTool: "", reviewerToolCount: 0,
+		reviewerInputTokens: 0, reviewerOutputTokens: 0, reviewerCacheReadTokens: 0, reviewerCacheWriteTokens: 0,
+		reviewerCostUsd: 0, reviewerContextPct: 0, reviewerProc: null, reviewerTimer: null,
 		reviewCounter: 0, totalIterations: 0, stepStatuses: new Map(),
 	};
 }
@@ -162,7 +174,7 @@ const DEFAULT_CONFIG: TaskConfig = {
 	standards_overrides: {},
 	task_areas: {},
 	worker: { model: "", tools: "read,write,edit,bash,grep,find,ls", thinking: "off" },
-	reviewer: { model: "openai/gpt-5.3-codex", tools: "read,bash,grep,find,ls", thinking: "on" },
+	reviewer: { model: "", tools: "read,bash,grep,find,ls", thinking: "on" },
 	context: {
 		worker_context_window: 0, warn_percent: 85, kill_percent: 95,
 		max_worker_iterations: 20, max_review_cycles: 2, no_progress_limit: 3,
@@ -408,6 +420,18 @@ function writeLaneState(state: TaskState): void {
 			workerLastRetryError: state.workerLastRetryError,
 			workerExitDiagnostic: state.workerExitDiagnostic || undefined,
 			reviewerStatus: state.reviewerStatus || "idle",
+			reviewerSessionName: state.reviewerSessionName || "",
+			reviewerType: state.reviewerType || "",
+			reviewerStep: state.reviewerStep || 0,
+			reviewerElapsed: state.reviewerElapsed || 0,
+			reviewerContextPct: state.reviewerContextPct || 0,
+			reviewerLastTool: state.reviewerLastTool || "",
+			reviewerToolCount: state.reviewerToolCount || 0,
+			reviewerCostUsd: state.reviewerCostUsd || 0,
+			reviewerInputTokens: state.reviewerInputTokens || 0,
+			reviewerOutputTokens: state.reviewerOutputTokens || 0,
+			reviewerCacheReadTokens: state.reviewerCacheReadTokens || 0,
+			reviewerCacheWriteTokens: state.reviewerCacheWriteTokens || 0,
 			timestamp: Date.now(),
 		};
 		writeFileSync(filePath, JSON.stringify(data) + "\n");
@@ -1540,6 +1564,9 @@ function spawnAgentTmux(opts: {
 	tools: string;
 	thinking: string;
 	taskId?: string;
+	/** Optional extension paths to load in the spawned pi session (via rpc-wrapper --extensions).
+	 *  When provided, --no-extensions is NOT passed to pi (would conflict). */
+	extensions?: string[];
 	/** Called on each poll tick with accumulated telemetry from the sidecar JSONL.
 	 *  Enables the tmux poll loop to update TaskState (tokens, cost, context%, tools, retries)
 	 *  with the same signals that subprocess mode gets from onTokenUpdate/onContextPct/onToolCall. */
@@ -1652,12 +1679,20 @@ function spawnAgentTmux(opts: {
 		"--system-prompt-file", quoteArg(sysTmpFile),
 		"--prompt-file", quoteArg(promptTmpFile),
 		"--tools", quoteArg(opts.tools),
-		// Passthrough pi args: flags forwarded to the underlying pi --mode rpc process.
-		// Note: --no-session is NOT passed here — rpc-wrapper.mjs already injects it.
-		"--",
-		"--thinking", quoteArg(opts.thinking),
-		"--no-extensions", "--no-skills",
 	];
+	// When extensions are provided, pass them to rpc-wrapper (which translates to `pi -e`)
+	// and do NOT pass --no-extensions (would conflict).
+	if (opts.extensions && opts.extensions.length > 0) {
+		wrapperArgs.push("--extensions", quoteArg(opts.extensions.join(",")));
+	}
+	// Passthrough pi args: flags forwarded to the underlying pi --mode rpc process.
+	// Note: --no-session is NOT passed here — rpc-wrapper.mjs already injects it.
+	wrapperArgs.push("--");
+	wrapperArgs.push("--thinking", quoteArg(opts.thinking));
+	if (!opts.extensions || opts.extensions.length === 0) {
+		wrapperArgs.push("--no-extensions");
+	}
+	wrapperArgs.push("--no-skills");
 	const wrapperCommand = wrapperArgs.join(" ");
 
 	// ── Handle stale session ─────────────────────────────────────────
@@ -1957,6 +1992,262 @@ export default function (pi: ExtensionAPI) {
 		});
 	}
 
+	// ── review_step Tool (orchestrated mode only) ───────────────────
+
+	/**
+	 * Reset reviewer telemetry fields on state to idle/zero.
+	 * Called after a review completes to clear dashboard metrics.
+	 */
+	function clearReviewerState(): void {
+		state.reviewerStatus = "idle";
+		state.reviewerType = "";
+		state.reviewerStep = 0;
+		state.reviewerSessionName = "";
+		state.reviewerElapsed = 0;
+		state.reviewerLastTool = "";
+		state.reviewerToolCount = 0;
+		state.reviewerInputTokens = 0;
+		state.reviewerOutputTokens = 0;
+		state.reviewerCacheReadTokens = 0;
+		state.reviewerCacheWriteTokens = 0;
+		state.reviewerCostUsd = 0;
+		state.reviewerContextPct = 0;
+		state.reviewerProc = null;
+		if (state.reviewerTimer) clearInterval(state.reviewerTimer);
+		state.reviewerTimer = null;
+	}
+
+	if (isOrchestratedMode()) {
+		pi.registerTool({
+			name: "review_step",
+			label: "Review Step",
+			description:
+				"Spawn a reviewer agent to evaluate your work on a step. " +
+				"Returns APPROVE, REVISE, RETHINK, or UNAVAILABLE. " +
+				"Use at step boundaries based on the task's review level.",
+			promptSnippet: "review_step(step, type) — spawn reviewer for a step (plan/code review)",
+			promptGuidelines: [
+				"Call review_step at step boundaries based on the task's Review Level (from STATUS.md header).",
+				"Review Level 0: skip all reviews. Level 1: plan review before implementing. Level 2: plan + code review. Level 3: plan + code + test review.",
+				"Skip reviews for Step 0 (Preflight) and the final documentation/delivery step.",
+				"For code reviews: before starting a step, capture the current HEAD commit with `git rev-parse HEAD` and pass it as the `baseline` parameter. This lets the reviewer see only that step's changes.",
+				"On REVISE: read the review file in .reviews/ for detailed feedback, address the issues, commit fixes, then proceed.",
+				"On RETHINK: reconsider your plan approach, adjust, then implement.",
+				"On UNAVAILABLE: reviewer failed — proceed with caution.",
+			],
+			parameters: Type.Object({
+				step: Type.Number({ description: "Step number to review" }),
+				type: Type.Union(
+					[Type.Literal("plan"), Type.Literal("code")],
+					{ description: 'Review type: "plan" or "code"' },
+				),
+				baseline: Type.Optional(Type.String({
+					description: "Git commit SHA to use as the diff baseline for code reviews. " +
+						"Capture HEAD before starting a step and pass it here so the reviewer " +
+						"sees only that step's changes. If omitted, the reviewer sees the full diff against HEAD.",
+				})),
+			}),
+			async execute(_toolCallId, params, _signal, _onUpdate, ctx) {
+				const { step: stepNum, type: reviewType, baseline } = params;
+
+				if (!state.task || !state.config) {
+					return {
+						content: [{ type: "text" as const, text: "UNAVAILABLE — no task loaded" }],
+						details: undefined,
+					};
+				}
+
+				const task = state.task;
+				const config = state.config;
+				const statusPath = join(task.taskFolder, "STATUS.md");
+				const reviewsDir = join(task.taskFolder, ".reviews");
+				if (!existsSync(reviewsDir)) mkdirSync(reviewsDir, { recursive: true });
+
+				// Low-risk step check (safety net — worker template also skips)
+				if (isLowRiskStep(stepNum, task.steps.length)) {
+					const label = stepNum === 0 ? "Preflight" : "final step";
+					logExecution(statusPath, `Skip ${reviewType} review`, `Step ${stepNum} (${label}) — low-risk`);
+					return {
+						content: [{ type: "text" as const, text: `APPROVE — Step ${stepNum} is low-risk (${label}), review skipped` }],
+						details: undefined,
+					};
+				}
+
+				// Increment review counter
+				state.reviewCounter++;
+				const num = String(state.reviewCounter).padStart(3, "0");
+				const requestPath = join(reviewsDir, `request-R${num}.md`);
+				const outputPath = join(reviewsDir, `R${num}-${reviewType}-step${stepNum}.md`);
+
+				// Resolve step baseline commit for code reviews.
+				// The worker should pass the pre-step HEAD SHA as `baseline` so the
+				// reviewer sees only this step's changes (not cumulative diff).
+				// Falls back to undefined (full diff) if baseline is not provided.
+				const stepBaselineCommit: string | undefined =
+					reviewType === "code" ? (baseline || undefined) : undefined;
+
+				// Find step info for the name
+				const stepInfo = task.steps.find(s => s.number === stepNum);
+				const stepName = stepInfo?.name || `Step ${stepNum}`;
+
+				// Generate review request
+				const request = generateReviewRequest(
+					reviewType, stepNum, stepName, task, config, outputPath, stepBaselineCommit,
+				);
+				writeFileSync(requestPath, request);
+
+				// Load reviewer agent definition
+				const reviewerDef = loadAgentDef(ctx.cwd, "task-reviewer");
+				const reviewerModel = config.reviewer.model
+					|| reviewerDef?.model
+					|| (ctx.model ? `${ctx.model.provider}/${ctx.model.id}` : "anthropic/claude-sonnet-4-20250514");
+				const reviewerPrompt = reviewerDef?.systemPrompt
+					|| "You are a code reviewer. Read the request and write your review to the specified output file.";
+				const systemPrompt = reviewerPrompt + "\n\n" + buildProjectContext(config, task.taskFolder);
+
+				// Update state for dashboard visibility
+				const sessionName = `${getTmuxPrefix()}-reviewer`;
+				state.reviewerStatus = "running";
+				state.reviewerType = `${reviewType} review`;
+				state.reviewerStep = stepNum;
+				state.reviewerSessionName = sessionName;
+				state.reviewerElapsed = 0;
+				state.reviewerLastTool = "";
+				state.reviewerToolCount = 0;
+				state.reviewerInputTokens = 0;
+				state.reviewerOutputTokens = 0;
+				state.reviewerCacheReadTokens = 0;
+				state.reviewerCacheWriteTokens = 0;
+				state.reviewerCostUsd = 0;
+				state.reviewerContextPct = 0;
+				updateWidgets();
+
+				const startTime = Date.now();
+				state.reviewerTimer = setInterval(() => {
+					state.reviewerElapsed = Date.now() - startTime;
+					updateWidgets();
+				}, 1000);
+
+				// Read the request file content as the prompt
+				const promptContent = readFileSync(requestPath, "utf-8");
+
+				// Resolve context window for reviewer context% calculation
+				const { contextWindow } = resolveContextWindow(config, ctx);
+
+				try {
+					// Spawn reviewer via spawnAgentTmux with onTelemetry for live metrics
+					const spawned = spawnAgentTmux({
+						sessionName,
+						cwd: ctx.cwd,
+						systemPrompt,
+						prompt: promptContent,
+						model: reviewerModel,
+						tools: config.reviewer.tools || reviewerDef?.tools || "read,write,bash,grep,find,ls",
+						thinking: config.reviewer.thinking || "on",
+						taskId: task.taskId,
+						onTelemetry: (delta) => {
+							// Accumulate tokens and cost
+							state.reviewerInputTokens += delta.inputTokens;
+							state.reviewerOutputTokens += delta.outputTokens;
+							state.reviewerCacheReadTokens += delta.cacheReadTokens;
+							state.reviewerCacheWriteTokens += delta.cacheWriteTokens;
+							state.reviewerCostUsd += delta.cost;
+
+							// Tool tracking
+							state.reviewerToolCount += delta.toolCalls;
+							if (delta.lastTool) {
+								state.reviewerLastTool = delta.lastTool;
+							}
+
+							// Context %
+							if (delta.latestTotalTokens > 0 && contextWindow > 0) {
+								state.reviewerContextPct = (delta.latestTotalTokens / contextWindow) * 100;
+							}
+
+							writeLaneState(state);
+							updateWidgets();
+						},
+					});
+
+					state.reviewerProc = { kill: spawned.kill };
+
+					// Await reviewer completion
+					const result = await spawned.promise;
+
+					clearInterval(state.reviewerTimer);
+					state.reviewerElapsed = Date.now() - startTime;
+					state.reviewerStatus = result.exitCode === 0 ? "done" : "error";
+					state.reviewerProc = null;
+					writeLaneState(state);
+					updateWidgets();
+
+					// Extract verdict from review output
+					let verdict = "UNKNOWN";
+					let reviseDetails = "";
+					if (existsSync(outputPath)) {
+						const review = readFileSync(outputPath, "utf-8");
+						verdict = extractVerdict(review);
+						if (verdict === "REVISE") {
+							// Extract a brief summary from the review for the worker
+							const summaryMatch = review.match(/###?\s*Summary[:\s]*([\s\S]*?)(?=###|$)/i);
+							reviseDetails = summaryMatch
+								? summaryMatch[1].trim().slice(0, 500)
+								: "See review file for details.";
+						}
+					} else {
+						verdict = "UNAVAILABLE";
+						logExecution(statusPath, `Reviewer R${num}`,
+							`${reviewType} review — reviewer did not produce output`);
+					}
+
+					// Log the review in STATUS.md
+					logReview(statusPath, `R${num}`, reviewType, stepNum, verdict,
+						`.reviews/R${num}-${reviewType}-step${stepNum}.md`);
+					logExecution(statusPath, `Review R${num}`,
+						`${reviewType} Step ${stepNum}: ${verdict}`);
+					updateStatusField(statusPath, "Review Counter", `${state.reviewCounter}`);
+
+					// Clear reviewer state for dashboard
+					clearReviewerState();
+					writeLaneState(state);
+					updateWidgets();
+
+					// Return verdict to the worker
+					let resultText: string;
+					if (verdict === "APPROVE") {
+						resultText = "APPROVE";
+					} else if (verdict === "REVISE") {
+						resultText = `REVISE: ${reviseDetails}\n\nFull review: .reviews/R${num}-${reviewType}-step${stepNum}.md`;
+					} else if (verdict === "RETHINK") {
+						resultText = `RETHINK — reconsider your approach. See .reviews/R${num}-${reviewType}-step${stepNum}.md`;
+					} else {
+						resultText = `UNAVAILABLE — reviewer did not produce a usable verdict.`;
+					}
+
+					return {
+						content: [{ type: "text" as const, text: resultText }],
+						details: undefined,
+					};
+				} catch (err: any) {
+					// Reviewer crashed
+					clearInterval(state.reviewerTimer);
+					clearReviewerState();
+					state.reviewerStatus = "error";
+					writeLaneState(state);
+					updateWidgets();
+
+					logExecution(statusPath, `Reviewer R${num}`,
+						`${reviewType} review — reviewer crashed: ${err?.message || err}`);
+
+					return {
+						content: [{ type: "text" as const, text: `UNAVAILABLE — reviewer error: ${err?.message || err}` }],
+						details: undefined,
+					};
+				}
+			},
+		});
+	}
+
 	// ── Execution Engine ─────────────────────────────────────────────
 
 	async function executeTask(ctx: ExtensionContext): Promise<void> {
@@ -1972,27 +2263,12 @@ export default function (pi: ExtensionAPI) {
 
 		// ── Per-task worker loop ─────────────────────────────────────
 		// Spawn one worker per iteration; each worker handles ALL remaining
-		// steps.  Reviews run after the worker exits, per newly-completed step.
+		// steps.  The worker drives reviews inline via the review_step tool
+		// (in orchestrated mode) — no deferred reviews after worker exit.
 		// If context limit is hit mid-task, the next iteration picks up from
 		// the first incomplete step via STATUS.md — same recovery mechanism.
 
-		// Collect baseline commits per step (for code review diffs).
-		// Baselines are updated at step boundaries so each code review
-		// sees only that step's changes, not cumulative diffs.
-		const stepBaselineCommits = new Map<number, string>();
-
-		// Track steps that received a REVISE verdict and need rework.
-		// This prevents the checkbox-count heuristic from re-completing them
-		// before the worker has had a chance to address reviewer feedback.
-		const needsRework = new Set<number>();
-
-		// Track which steps have already received a plan review so we don't
-		// re-run plan review on rework cycles (only code review reruns).
-		const planReviewedSteps = new Set<number>();
-
-		// Mark all incomplete steps as in-progress and capture the baseline
-		// commit for the first one. Reviews are transition-based: they run
-		// when a step newly completes after the worker exits, not up-front.
+		// Mark all incomplete steps as in-progress
 		{
 			const currentStatus = parseStatusMd(readFileSync(statusPath, "utf-8"));
 			for (const step of task.steps) {
@@ -2002,23 +2278,12 @@ export default function (pi: ExtensionAPI) {
 				// Mark step as in-progress and log its start
 				updateStepStatus(statusPath, step.number, "in-progress");
 				logExecution(statusPath, `Step ${step.number} started`, step.name);
-
-				// Capture baseline commit for the FIRST incomplete step only.
-				// Later steps get their baselines when the prior step completes
-				// (see the newlyCompleted handler in the iteration loop).
-				// This prevents cross-step diff bleeding in code reviews.
-				if (!stepBaselineCommits.size) {
-					stepBaselineCommits.set(step.number, getHeadCommitSha());
-				}
 			}
 		}
 
 		// Helper: determine if a parsed step is complete.
-		// A step in needsRework (set after REVISE) is never complete, even if
-		// all checkboxes are checked — the worker must address the feedback first.
 		function isStepComplete(ss: StepInfo | undefined): boolean {
 			if (!ss) return false;
-			if (needsRework.has(ss.number)) return false;
 			if (ss.status === "complete") return true;
 			// Fallback: infer from checkboxes (covers "in-progress" and "not-started")
 			return ss.totalChecked === ss.totalItems && ss.totalItems > 0;
@@ -2058,13 +2323,6 @@ export default function (pi: ExtensionAPI) {
 				if (isStepComplete(ss)) completedBefore.add(ss.number);
 			}
 
-			// Ensure the first remaining step has a baseline. On subsequent
-			// iterations (after context-limit recovery), the first remaining step
-			// may not have a baseline yet if it wasn't the first step originally.
-			if (remainingSteps.length > 0 && !stepBaselineCommits.has(remainingSteps[0].number)) {
-				stepBaselineCommits.set(remainingSteps[0].number, getHeadCommitSha());
-			}
-
 			await runWorker(remainingSteps, ctx);
 
 			if (state.phase === "error") return;
@@ -2090,49 +2348,17 @@ export default function (pi: ExtensionAPI) {
 			}
 
 			// Find newly completed steps.
-			// For steps in needsRework, the worker must have addressed the
-			// reviewer feedback — we detect this by checking if STATUS.md was
-			// updated (the worker adds/checks revision items). We temporarily
-			// remove the step from needsRework to let isStepComplete() evaluate
-			// the checkbox state, then re-add if it's still not complete.
 			const newlyCompleted: StepInfo[] = [];
 			for (const step of task.steps) {
 				if (completedBefore.has(step.number)) continue;
 				const ss = afterStatus.steps.find(s => s.number === step.number);
-				if (needsRework.has(step.number)) {
-					// For rework steps, check if the worker set status to "complete"
-					// or if the step now has all checkboxes checked (worker addressed feedback)
-					if (ss?.status === "complete" ||
-						(ss && ss.totalChecked === ss.totalItems && ss.totalItems > 0)) {
-						needsRework.delete(step.number);
-						updateStepStatus(statusPath, step.number, "complete");
-						logExecution(statusPath, `Step ${step.number} complete`, `${step.name} (rework)`);
-						newlyCompleted.push(step);
-					}
-				} else if (isStepComplete(ss)) {
+				if (isStepComplete(ss)) {
 					updateStepStatus(statusPath, step.number, "complete");
 					logExecution(statusPath, `Step ${step.number} complete`, step.name);
 					newlyCompleted.push(step);
 				}
 			}
 
-			// Update baselines for subsequent steps using step boundary commits.
-			// When the worker completes multiple steps in one iteration, each step's
-			// commit becomes the baseline for the next step's code review.
-			if (newlyCompleted.length > 1) {
-				for (let i = 0; i < newlyCompleted.length; i++) {
-					const completedStep = newlyCompleted[i];
-					const boundaryCommit = findStepBoundaryCommit(
-						completedStep.number, task.taskId,
-						stepBaselineCommits.get(completedStep.number)
-					);
-					if (boundaryCommit && i + 1 < newlyCompleted.length) {
-						// Use this step's completion commit as the next step's baseline
-						stepBaselineCommits.set(newlyCompleted[i + 1].number, boundaryCommit);
-					}
-				}
-			}
-
 			// Log iteration summary with progress delta and completed steps
 			const completedNames = newlyCompleted.map(s => `Step ${s.number}`).join(", ");
 			if (newlyCompleted.length > 0) {
@@ -2143,57 +2369,8 @@ export default function (pi: ExtensionAPI) {
 				ctx.ui.notify(`Iteration ${iter + 1}: +${progressDelta} checkboxes (no steps fully completed)`, "info");
 			}
 
-			// ── Run reviews for newly completed steps (transition-based) ──
-			// Plan reviews run once per step (first completion); code reviews
-			// run on every completion (including rework). Both respect review
-			// level and low-risk skip logic.
-			// Gate on phase !== "error" so reviews still run when paused
-			// (pause is honored after reviews, before next iteration).
-			if (state.phase !== "error") {
-				for (const step of newlyCompleted) {
-					const lowRisk = isLowRiskStep(step.number, task.steps.length);
-
-					// ── Plan review (level ≥ 1, first completion only) ──
-					if (task.reviewLevel >= 1 && !planReviewedSteps.has(step.number)) {
-						if (lowRisk) {
-							const label = step.number === 0 ? "Preflight" : "final step";
-							logExecution(statusPath, `Skip plan review`, `Step ${step.number} (${label}) — low-risk`);
-							ctx.ui.notify(`⏭️ Skipping plan review for Step ${step.number} (${label})`, "info");
-						} else {
-							const verdict = await doReview("plan", step, ctx);
-							if (verdict === "RETHINK") {
-								ctx.ui.notify(`Reviewer: RETHINK on Step ${step.number} plan. Proceeding with caution.`, "warning");
-							}
-						}
-						planReviewedSteps.add(step.number);
-					}
-
-					// ── Code review (level ≥ 2) ──
-					if (task.reviewLevel >= 2) {
-						if (lowRisk) {
-							const label = step.number === 0 ? "Preflight" : "final step";
-							logExecution(statusPath, `Skip code review`, `Step ${step.number} (${label}) — low-risk`);
-							ctx.ui.notify(`⏭️ Skipping code review for Step ${step.number} (${label})`, "info");
-						} else {
-							const baseline = stepBaselineCommits.get(step.number);
-							const verdict = await doReview("code", step, ctx, baseline);
-							if (verdict === "REVISE") {
-								ctx.ui.notify(`Reviewer: REVISE on Step ${step.number}. Will rework in next iteration.`, "warning");
-								// Mark step as needing rework — undo the "complete" status.
-								// needsRework ensures isStepComplete() won't re-complete this
-								// step based on checkbox counts alone; the worker must address
-								// the reviewer feedback first, which will update STATUS.md and
-								// clear the rework flag when the step is newly completed again.
-								needsRework.add(step.number);
-								updateStepStatus(statusPath, step.number, "in-progress");
-								// Update baseline so the next code review for this step
-								// diffs only the rework changes, not the original step work
-								stepBaselineCommits.set(step.number, getHeadCommitSha());
-							}
-						}
-					}
-				}
-			}
+			// Reviews are now driven inline by the worker via the review_step
+			// tool (orchestrated mode). No deferred review logic here.
 
 			// Update local cache
 			const refreshed = parseStatusMd(readFileSync(statusPath, "utf-8"));
@@ -2727,7 +2904,7 @@ export default function (pi: ExtensionAPI) {
 		writeFileSync(requestPath, request);
 
 		const reviewerDef = loadAgentDef(ctx.cwd, "task-reviewer");
-		const reviewerModel = config.reviewer.model || reviewerDef?.model || "openai/gpt-5.3-codex";
+		const reviewerModel = config.reviewer.model || reviewerDef?.model || (ctx.model ? `${ctx.model.provider}/${ctx.model.id}` : "anthropic/claude-sonnet-4-20250514");
 		const reviewerPrompt = reviewerDef?.systemPrompt || "You are a code reviewer. Read the request and write your review to the specified output file.";
 		const systemPrompt = reviewerPrompt + "\n\n" + buildProjectContext(config, task.taskFolder);
 
@@ -2866,7 +3043,7 @@ export default function (pi: ExtensionAPI) {
 		const reviewModel = config.quality_gate.review_model
 			|| config.reviewer.model
 			|| reviewerDef?.model
-			|| "openai/gpt-5.3-codex";
+			|| (ctx.model ? `${ctx.model.provider}/${ctx.model.id}` : "anthropic/claude-sonnet-4-20250514");
 
 		const reviewerPrompt = reviewerDef?.systemPrompt
 			|| "You are a quality gate reviewer. Read the review request and write your JSON verdict to the specified file.";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "taskplane",
-  "version": "0.8.2",
+  "version": "0.9.1",
   "description": "AI agent orchestration for pi — parallel task execution with checkpoint discipline",
   "keywords": [
     "pi-package",

package/templates/agents/local/task-worker.md

@@ -16,6 +16,7 @@ name: task-worker
   - Multi-step execution (worker handles all remaining steps per invocation)
   - Iteration recovery (context limit → next invocation resumes from STATUS.md)
   - Git commit conventions (per-step commits) and .DONE file creation
+  - Review protocol (inline reviews via review_step tool when available)
   - Review response handling
 
   Add project-specific rules below. Common examples:

package/templates/agents/task-worker.md

@@ -158,6 +158,45 @@ When a reviewer returns REVISE with specific feedback items:
 - Do NOT expand task scope beyond what the steps require
 - If you discover something out of scope, note it in STATUS.md Discoveries table
 
+## Review Protocol
+
+If you have access to a `review_step` tool, use it at step boundaries to spawn
+a reviewer agent. The tool takes two parameters: `step` (number) and `type`
+("plan" or "code"). It returns a verdict string.
+
+**When to call reviews** (based on Review Level from STATUS.md header):
+
+- **Review Level 0 (None):** Skip all reviews.
+- **Review Level 1 (Plan Only):** Before implementing each step, call
+  `review_step(step=N, type="plan")` to get plan feedback.
+- **Review Level 2 (Plan + Code):** Plan review before implementing, then code
+  review after implementing and committing.
+- **Review Level 3 (Full):** Plan + code + test reviews.
+
+**Always skip reviews for:** Step 0 (Preflight) and the final step (typically
+documentation/delivery). These are low-risk steps where review overhead exceeds
+value.
+
+**Handling verdicts:**
+- **APPROVE** → proceed to next step
+- **RETHINK** → reconsider your plan approach, adjust, then implement
+- **REVISE** → read the review file in `.reviews/` for detailed feedback,
+  address the issues, commit fixes, then proceed
+- **UNAVAILABLE** → reviewer failed, proceed with caution
+
+**Example flow for a Review Level 2 task, Step 3:**
+1. Read Step 3 requirements
+2. Call `review_step(step=3, type="plan")` → get plan feedback
+3. Capture baseline: run `git rev-parse HEAD` and save the SHA
+4. Implement Step 3
+5. Commit changes
+6. Call `review_step(step=3, type="code", baseline="<saved SHA>")` → get code feedback
+7. If REVISE: fix issues, commit again
+8. Move to Step 4
+
+If the `review_step` tool is not available (e.g., non-orchestrated mode), skip
+this protocol entirely — the task-runner handles reviews externally.
+
 ## Self-Documentation
 
 You have standing permission to: