taskplane 0.24.2 → 0.24.4

package/extensions/taskplane/config-schema.ts

@@ -610,7 +610,7 @@ export const DEFAULT_ORCHESTRATOR_SECTION: OrchestratorSection = {
 		onTaskFailure: "skip-dependents",
 		onMergeFailure: "pause",
 		stallTimeout: 30,
-		maxWorkerMinutes: 30,
+		maxWorkerMinutes: 120,
 		abortGracePeriod: 60,
 	},
 	monitoring: {

package/extensions/taskplane/execution.ts

@@ -2214,7 +2214,7 @@ export async function executeLaneV2(
 			projectName: config.project?.name || "project",
 			maxIterations: 20,
 			noProgressLimit: 3,
-			maxWorkerMinutes: config.failure?.maxWorkerMinutes || 30,
+			maxWorkerMinutes: config.failure?.maxWorkerMinutes || 120,
 			warnPercent: 85,
 			killPercent: 95,
 			onSupervisorAlert,

package/extensions/taskplane/lane-runner.ts

@@ -286,30 +286,32 @@ export async function executeTaskV2(
 		let iterationTelemetry: Partial<AgentHostResult> = {};
 
 		const spawned = spawnAgent(hostOpts, undefined, (telemetry) => {
-			// Context pressure check
-			if (telemetry.contextUsage) {
-				const pct = telemetry.contextUsage.percent;
-				if (pct >= config.warnPercent) {
-					const msg = `Wrap up (context ${Math.round(pct)}%)`;
-					if (!existsSync(wrapUpFile)) writeFileSync(wrapUpFile, msg);
-				}
-				if (pct >= config.killPercent) {
-					workerKillReason = "context";
-					spawned.kill();
+			try {
+				// Context pressure check
+				if (telemetry.contextUsage) {
+					const pct = telemetry.contextUsage.percent;
+					if (pct >= config.warnPercent) {
+						const msg = `Wrap up (context ${Math.round(pct)}%)`;
+						if (!existsSync(wrapUpFile)) writeFileSync(wrapUpFile, msg);
+					}
+					if (pct >= config.killPercent) {
+						workerKillReason = "context";
+						spawned.kill();
+					}
 				}
-			}
 
-			iterationTelemetry = telemetry;
-			lastTelemetry = telemetry;
-			// Emit lane snapshot
-			emitSnapshot(config, taskId, "running", telemetry, statusPath);
+				iterationTelemetry = telemetry;
+				lastTelemetry = telemetry;
+				// Emit lane snapshot
+				emitSnapshot(config, taskId, "running", telemetry, statusPath);
+			} catch { /* non-fatal: telemetry callback must never crash the engine */ }
 		});
 
 		// Reviewer telemetry is written by the worker bridge during review_step.
 		// Poll snapshot refresh independently from worker message_end cadence so
 		// the dashboard sees reviewer activity while tool calls are in-flight.
 		const reviewerRefresh = setInterval(() => {
-			emitSnapshot(config, taskId, "running", iterationTelemetry, statusPath);
+			try { emitSnapshot(config, taskId, "running", iterationTelemetry, statusPath); } catch { /* non-fatal */ }
 		}, 1000);
 
 		let workerResult: AgentHostResult;
@@ -605,6 +607,12 @@ export function readReviewerTelemetrySnapshot(
 	}
 }
 
+/**
+ * Emit a lane snapshot to disk. NON-THROWING by contract — all errors are
+ * caught and logged. This function is called from setInterval callbacks
+ * and onTelemetry callbacks where an unhandled throw would trigger
+ * uncaughtException and crash the engine-worker process.
+ */
 function emitSnapshot(
 	config: LaneRunnerConfig,
 	taskId: string,
@@ -612,51 +620,56 @@ function emitSnapshot(
 	telemetry: Partial<AgentHostResult>,
 	statusPath: string,
 ): void {
-	// Parse progress from STATUS.md
-	let progress: RuntimeTaskProgress | null = null;
 	try {
-		const content = readFileSync(statusPath, "utf-8");
-		const parsed = parseStatusMd(content);
-		const currentStepMatch = content.match(/\*\*Current Step:\*\*\s*(.+)/);
-		const checked = parsed.steps.reduce((sum, s) => sum + s.totalChecked, 0);
-		const total = parsed.steps.reduce((sum, s) => sum + s.totalItems, 0);
-		progress = {
-			currentStep: currentStepMatch?.[1]?.trim() || "Unknown",
-			checked,
-			total,
-			iteration: parsed.iteration,
-			reviews: parsed.reviewCounter,
+		// Parse progress from STATUS.md
+		let progress: RuntimeTaskProgress | null = null;
+		try {
+			const content = readFileSync(statusPath, "utf-8");
+			const parsed = parseStatusMd(content);
+			const currentStepMatch = content.match(/\*\*Current Step:\*\*\s*(.+)/);
+			const checked = parsed.steps.reduce((sum, s) => sum + s.totalChecked, 0);
+			const total = parsed.steps.reduce((sum, s) => sum + s.totalItems, 0);
+			progress = {
+				currentStep: currentStepMatch?.[1]?.trim() || "Unknown",
+				checked,
+				total,
+				iteration: parsed.iteration,
+				reviews: parsed.reviewCounter,
+			};
+		} catch { /* best effort */ }
+
+		const reviewerSnapshot = readReviewerTelemetrySnapshot(config, statusPath);
+
+		const snapshot: RuntimeLaneSnapshot = {
+			batchId: config.batchId,
+			laneNumber: config.laneNumber,
+			laneId: `lane-${config.laneNumber}`,
+			repoId: config.repoId,
+			taskId,
+			segmentId: null,
+			status,
+			worker: {
+				agentId: buildRuntimeAgentId(config.agentIdPrefix, config.laneNumber, "worker"),
+				status: mapLaneSnapshotStatusToWorkerStatus(status),
+				elapsedMs: telemetry.durationMs ?? 0,
+				toolCalls: telemetry.toolCalls ?? 0,
+				contextPct: telemetry.contextUsage?.percent ?? 0,
+				costUsd: telemetry.costUsd ?? 0,
+				lastTool: telemetry.lastTool ?? "",
+				inputTokens: telemetry.inputTokens ?? 0,
+				outputTokens: telemetry.outputTokens ?? 0,
+				cacheReadTokens: telemetry.cacheReadTokens ?? 0,
+				cacheWriteTokens: telemetry.cacheWriteTokens ?? 0,
+			},
+			reviewer: reviewerSnapshot,
+			progress,
+			updatedAt: Date.now(),
 		};
-	} catch { /* best effort */ }
-
-	const reviewerSnapshot = readReviewerTelemetrySnapshot(config, statusPath);
-
-	const snapshot: RuntimeLaneSnapshot = {
-		batchId: config.batchId,
-		laneNumber: config.laneNumber,
-		laneId: `lane-${config.laneNumber}`,
-		repoId: config.repoId,
-		taskId,
-		segmentId: null,
-		status,
-		worker: {
-			agentId: buildRuntimeAgentId(config.agentIdPrefix, config.laneNumber, "worker"),
-			status: mapLaneSnapshotStatusToWorkerStatus(status),
-			elapsedMs: telemetry.durationMs ?? 0,
-			toolCalls: telemetry.toolCalls ?? 0,
-			contextPct: telemetry.contextUsage?.percent ?? 0,
-			costUsd: telemetry.costUsd ?? 0,
-			lastTool: telemetry.lastTool ?? "",
-			inputTokens: telemetry.inputTokens ?? 0,
-			outputTokens: telemetry.outputTokens ?? 0,
-			cacheReadTokens: telemetry.cacheReadTokens ?? 0,
-			cacheWriteTokens: telemetry.cacheWriteTokens ?? 0,
-		},
-		reviewer: reviewerSnapshot,
-		progress,
-		updatedAt: Date.now(),
-	};
 
-	writeLaneSnapshot(config.stateRoot, config.batchId, config.laneNumber, snapshot as any);
+		writeLaneSnapshot(config.stateRoot, config.batchId, config.laneNumber, snapshot as any);
+	} catch {
+		// Non-fatal: snapshot is telemetry, not execution-critical.
+		// Swallow to prevent uncaughtException crash in setInterval/callback contexts.
+	}
 }
 

package/extensions/taskplane/supervisor-primer.md

@@ -23,7 +23,7 @@ reviewers), or merge branches (that's merge agents). You supervise all of them.
 
 **Your tools:** `read`, `write`, `edit`, `bash`, `grep`, `find`, `ls`. You have
 full filesystem and command-line access. Use it to read state files, run git
-commands, edit batch state, manage tmux sessions, and run verification.
+commands, edit batch state, inspect worker/merge agent execution, and run verification.
 
 **Your system prompt** is built from `templates/agents/supervisor.md` (base
 template, ships with the package) composed with `.pi/agents/supervisor.md`
@@ -43,13 +43,13 @@ You (supervisor) ← operator talks to you
 │   ├── Computes waves (topological sort)
 │   ├── Assigns tasks to lanes (parallel execution slots)
 │   ├── Provisions git worktrees per lane
-│   ├── Spawns worker sessions in tmux
+│   ├── Spawns worker processes/sessions per lane
 │   ├── Polls for .DONE files and STATUS.md progress
 │   ├── Merges lane branches into orch branch after each wave
 │   └── Advances to next wave after successful merge
 │
 ├── Worker Agents (LLM, one per task)
-│   ├── Run in tmux sessions inside git worktrees
+│   ├── Run as subprocess agents inside git worktrees
 │   ├── Read PROMPT.md for requirements, STATUS.md for state
 │   ├── Write code, run tests, check STATUS.md boxes
 │   ├── Commit at step boundaries
@@ -216,7 +216,7 @@ Wave N starts
 ├── 1. Provision: Create lane worktrees from orch branch
 │   └── git worktree add .worktrees/{opId}-{batchId}/lane-{N} -b task/{opId}-lane-{N}-{batchId} orch/{opId}-{batchId}
 │
-├── 2. Execute: Spawn tmux sessions for each lane
+├── 2. Execute: Spawn worker sessions for each lane
 │   ├── Each session runs the task-runner extension
 │   ├── Task-runner iterates through task steps
 │   ├── Workers write code, check STATUS.md boxes, commit
@@ -225,7 +225,7 @@ Wave N starts
 │
 ├── 3. Monitor: Poll loop checks every 5 seconds
 │   ├── Check .DONE file existence → task succeeded
-│   ├── Check tmux session alive → still running
+│   ├── Check lane process/session alive → still running
 │   ├── Check STATUS.md → track progress for dashboard
 │   └── Check stall timeout → no STATUS.md change for too long
 │
@@ -250,7 +250,7 @@ Wave N starts
 | Stage | Failure | Symptom |
 |-------|---------|---------|
 | Provision | Stale worktree from previous run | `git worktree add` fails |
-| Execute | Worker session crashes | tmux session disappears without .DONE |
+| Execute | Worker session crashes | lane process exits without .DONE |
 | Execute | Worker makes no progress | STATUS.md unchanged for `stallTimeout` minutes |
 | Execute | API error (rate limit, overload) | Session exits, pi handles retry internally |
 | Merge | Merge agent times out | No result JSON within `merge.timeoutMinutes` |
@@ -342,8 +342,8 @@ git log --oneline orch/{branch}..task/{lane-branch}  # empty = already merged
 or "🔒 Merge agent on lane N appears stuck (no output for 20 min)."
 
 **How it works:** The merge health monitor (TP-056) actively polls merge agent
-tmux sessions every 2 minutes during the merge phase. It checks:
-- **Session liveness:** `tmux has-session` — is the session alive?
+processes every 2 minutes during the merge phase. It checks:
+- **Process liveness:** process registry + PID checks — is the agent still alive?
 - **Activity detection:** Captures the last 10 lines of pane output and compares
   with the previous snapshot. If output hasn't changed, the session may be stalled.
 
@@ -354,9 +354,9 @@ tmux sessions every 2 minutes during the merge phase. It checks:
 - **Stuck** (20 min no output): `merge_health_stuck` event → recommendation to kill
 
 **Recovery:**
-1. Attach to the session to inspect: `tmux attach -t {sessionName}`
-2. If truly stuck, kill the session: `tmux kill-session -t {sessionName}`
-3. The engine detects the dead session and applies the `on_merge_failure` policy
+1. Inspect lane/merge diagnostics: `read_lane_logs(lane)` and recent merge alerts
+2. If truly stuck, stop the batch/merge path via orchestrator tools (`orch_abort(hard=true)` when required)
+3. The engine detects the dead agent and applies the `on_merge_failure` policy
 4. Resume with `/orch-resume` if needed
 
 **Note:** The monitor does NOT kill sessions autonomously — it emits events for
@@ -427,7 +427,7 @@ batch terminal.
 
 ### Pattern 5: Worker Session Crash
 
-**Symptom:** Task shows failed, tmux session is gone, no `.DONE`.
+**Symptom:** Task shows failed, worker process is gone, no `.DONE`.
 
 **Diagnosis:**
 ```bash
@@ -567,12 +567,12 @@ rm -rf .worktrees/{path}
 git worktree prune
 ```
 
-### Check tmux sessions
-```bash
-tmux ls                           # list all sessions
-tmux has-session -t {name} 2>&1   # check specific session
-tmux kill-session -t {name}       # kill specific session
-tmux capture-pane -t {name} -p    # see what's on screen
+### Check active agents
+```text
+list_active_agents()      # list running worker/reviewer/merge agents
+read_agent_status()       # summarize STATUS.md + telemetry for all lanes
+read_lane_logs(<lane>)    # inspect stderr/crash diagnostics for a lane
+trigger_wrap_up(<lane>)   # graceful stop signal for a worker lane
 ```
 
 ---
@@ -774,7 +774,7 @@ If the batch is actively running, call `orch_pause()` first.
 - `read_agent_status(lane?)` — Read STATUS.md + telemetry for a lane (step, progress, context %, cost, elapsed). Omit lane for all lanes.
 - `trigger_wrap_up(lane)` — Write `.task-wrap-up` signal to gracefully stop a worker on a lane.
 - `read_lane_logs(lane)` — Read stderr/crash logs and exit diagnostics for a lane.
-- `list_active_agents()` — List all tmux sessions with role, lane, task, context %, elapsed, cost.
+- `list_active_agents()` — List active worker/reviewer/merge agents with role, lane, task, context %, elapsed, cost.
 
 Plus general tools: `read`, `write`, `edit`, `bash`, `grep`, `find`, `ls`
 for inspecting files, running git commands, and editing batch state.
@@ -1010,7 +1010,7 @@ When you activate at the start of a batch:
 2. Note the `orchBranch`, `baseBranch`, `wavePlan`, `totalWaves`
 3. Check that the orch branch exists: `git branch | grep orch/`
 4. Verify worktrees are provisioned for the current wave
-5. Confirm tmux sessions are alive for active lanes
+5. Confirm active worker lanes are alive (agent status + lane logs)
 6. Read configuration for key values: `merge.timeoutMinutes`, `maxLanes`,
    review levels, verification commands
 7. Report to operator: "Batch {batchId} active. {N} waves, {M} tasks.
@@ -1317,8 +1317,7 @@ before writing** — if files already exist (partial setup), read and merge.
       "worktreeLocation": "subdirectory",
       "worktreePrefix": ".worktrees",
       "batchIdFormat": "timestamp",
-      "spawnMode": "tmux",
-      "tmuxPrefix": "tp",
+      "spawnMode": "subprocess",
       "operatorId": ""
     },
     "dependencies": { "source": "prompt", "cache": true },
@@ -1340,7 +1339,7 @@ before writing** — if files already exist (partial setup), read and merge.
 - `project.name`: Use the actual project name (from package.json, README, etc.)
 - `paths.tasks` and `taskAreas`: Match what was agreed in the task area discussion
 - `testing.commands`: Use the detected test command as a named object (e.g., `{"test": "cd extensions && node --experimental-strip-types --experimental-test-module-mocks --no-warnings --import ./tests/loader.mjs --test tests/*.test.ts"}`)
-- `orchestrator.spawnMode`: Use `"tmux"` if tmux is available, `"subprocess"` otherwise
+- `orchestrator.spawnMode`: Use `"subprocess"` (default, recommended runtime mode)
 - `orchestrator.maxLanes`: Start with 2 for first-time users (safe default)
 - `merge.verify`: Add the project's test command for post-merge verification
 
@@ -1495,8 +1494,8 @@ when the supervisor suggests it.
 5. **Orphaned batch state**: Read `.pi/batch-state.json` — if it exists and
    phase is terminal (`completed`, `failed`, `stopped`), check if it's old
    (> 7 days since `endedAt`) and suggest cleanup
-6. **tmux availability**: Run `which tmux` — if unavailable, warn that
-   orchestrator will use subprocess mode (less observable)
+6. **Agent observability tools**: Confirm supervisor tool connectivity by checking
+   `orch_status()` and `list_active_agents()` respond without errors
 7. **Disk space**: Run `df -h .` (Unix) or `wmic logicaldisk get size,freespace`
    (Windows) — warn if less than 5GB free (worktrees use space)
 8. **Supervisor lockfile**: Check `.pi/supervisor/lock.json` — if it exists
@@ -1517,7 +1516,7 @@ Infrastructure:
   ✅ Config valid (3 task areas configured)
   ✅ Git clean, on 'develop'
   ⚠️ 2 stale worktree directories from batch 20260315T093012
-  ✅ tmux available
+  ✅ agent observability tools available
   ✅ No orphaned batch state
   ❌ Stale supervisor lockfile found (no active batch)
 

package/extensions/taskplane/supervisor.ts

@@ -104,19 +104,19 @@ export const ACTION_CLASSIFICATION_EXAMPLES: Readonly<Record<RecoveryActionClass
 		"Reading batch-state.json, STATUS.md, events.jsonl, merge results",
 		"Running git status, git log, git diff",
 		"Running test suites (node --experimental-strip-types --experimental-test-module-mocks --no-warnings --import ./tests/loader.mjs --test ..., etc.)",
-		"Listing tmux sessions (tmux list-sessions)",
+		"Inspecting active agents and lane status (list_active_agents, read_agent_status)",
 		"Checking worktree health (git worktree list)",
 		"Reading any file for diagnostics",
 	],
 	tier0_known: [
-		"Restarting a crashed tmux worker session",
+		"Triggering graceful wrap-up/retry flow for a stalled worker lane",
 		"Cleaning up stale worktrees for retry",
 		"Retrying a timed-out merge",
 		"Resetting a session name collision",
 		"Clearing a git lock file (.git/index.lock)",
 	],
 	destructive: [
-		"Killing a tmux session (tmux kill-session)",
+		"Forcing lane/batch termination paths (for example orch_abort(hard=true))",
 		"Editing batch-state.json fields",
 		"Running git reset, git merge, git checkout -B",
 		"Removing worktrees (git worktree remove)",
@@ -2053,7 +2053,7 @@ Use these to:
 - Read batch state, STATUS.md files, merge results, event logs
 - Run git commands for diagnostics and manual merge recovery
 - Edit batch-state.json for state repairs (when needed)
-- Manage tmux sessions (list, kill, attach)
+- Manage worker lane execution state (agent status, wrap-up, diagnostics)
 - Run verification commands (tests)
 
 ## Standing Orders
@@ -2084,19 +2084,19 @@ Every action you take falls into one of three categories:
 - Reading batch-state.json, STATUS.md, events.jsonl, merge results
 - Running \`git status\`, \`git log\`, \`git diff\`
 - Running test suites (\`node --experimental-strip-types --experimental-test-module-mocks --no-warnings --import ./tests/loader.mjs --test ...\`, etc.)
-- Listing tmux sessions (\`tmux list-sessions\`)
+- Inspecting active agents and lane status (\`list_active_agents\`, \`read_agent_status\`)
 - Checking worktree health (\`git worktree list\`)
 - Reading any file for diagnostics
 
 ### Tier 0 Known (known recovery patterns)
-- Restarting a crashed tmux worker session
+- Triggering graceful wrap-up/retry flow for a stalled worker lane
 - Cleaning up stale worktrees for retry
 - Retrying a timed-out merge
 - Resetting a session name collision
 - Clearing a git lock file (\`.git/index.lock\`)
 
 ### Destructive (state mutations, irreversible operations)
-- Killing a tmux session (\`tmux kill-session\`)
+- Forcing lane/batch termination paths (for example \`orch_abort(hard=true)\`)
 - Editing batch-state.json fields
 - Running \`git reset\`, \`git merge\`, \`git checkout -B\`
 - Removing worktrees (\`git worktree remove\`)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "taskplane",
-  "version": "0.24.2",
+  "version": "0.24.4",
   "description": "AI agent orchestration for pi — parallel task execution with checkpoint discipline",
   "keywords": [
     "pi-package",

package/templates/agents/supervisor.md

@@ -38,7 +38,7 @@ Use these to:
 - Read batch state, STATUS.md files, merge results, event logs
 - Run git commands for diagnostics and manual merge recovery
 - Edit batch-state.json for state repairs (when needed)
-- Manage tmux sessions (list, kill, attach)
+- Manage worker lane execution state (agent status, wrap-up, diagnostics)
 - Run verification commands (tests)
 
 ## Standing Orders
@@ -69,19 +69,19 @@ Every action you take falls into one of three categories:
 - Reading batch-state.json, STATUS.md, events.jsonl, merge results
 - Running `git status`, `git log`, `git diff`
 - Running test suites (`node --experimental-strip-types --experimental-test-module-mocks --no-warnings --import ./tests/loader.mjs --test ...`, etc.)
-- Listing tmux sessions (`tmux list-sessions`)
+- Inspecting active agents and lane status (`list_active_agents`, `read_agent_status`)
 - Checking worktree health (`git worktree list`)
 - Reading any file for diagnostics
 
 ### Tier 0 Known (known recovery patterns)
-- Restarting a crashed tmux worker session
+- Triggering graceful wrap-up/retry flow for a stalled worker lane
 - Cleaning up stale worktrees for retry
 - Retrying a timed-out merge
 - Resetting a session name collision
 - Clearing a git lock file (`.git/index.lock`)
 
 ### Destructive (state mutations, irreversible operations)
-- Killing a tmux session (`tmux kill-session`)
+- Forcing lane/batch termination paths (for example `orch_abort(hard=true)`)
 - Editing batch-state.json fields
 - Running `git reset`, `git merge`, `git checkout -B`
 - Removing worktrees (`git worktree remove`)