fix(fleet): watch viewer-session leak + workdir test settle-race

The 2 deferred F3 items from the #595/#599 reviews:
- agent watch: wrap the interactive attach in try/finally and kill the grouped
  viewer session in finally — so an attach that throws or an interrupted process
  never leaves a stray <agent>-watch-<pid> session behind (the leak you found).
- test-start-agent-session.sh Test 1: pane_current_path briefly reflects the tmux
  server's cwd until the pane process establishes its -c start dir; poll until it
  settles. Fixes the cwd-dependent false failure that aborted the suite before the
  heartbeat tests (6/7) — those now run.

Verified: full shell suite green from /tmp (was failing); prettier clean.

Refs #588 #542

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/fleet/PRD-fleet-suite.md

@@ -0,0 +1,105 @@
+# PRD — Mosaic Fleet Suite (init, configure, operate)
+
+> **Workstream:** W-FLEET (Fleet) under mission `mvp-20260312` · **Phase:** 3→4 productization
+> **North star:** [docs/fleet/north-star.md](./north-star.md) · prior: Phase-2 observability (#579), durable launch (#581), real-agent enablement (#583/#584/#586), releases 0.0.35–0.0.37
+> **Lead:** Jarvis @ `w-jarvis`. **Collaborator:** coder agent @ `dragon-lin` (jwoltje@10.1.10.37:coder0-0).
+> Owner of this file: Fleet workstream lead. Does not modify MVP single-writer control-plane files.
+
+## Mission
+
+Turn the proven fleet primitives into a **user-installable, AI-free-configurable fleet product**:
+a user runs `mosaic fleet init`, answers a few questions (general / coding / research / hybrid),
+gets a recommended set of agents plus one always-on orchestrator wired for chat-ops, and can
+operate, mutate, re-create, and observe the fleet — over tmux today and Matrix tomorrow — from
+CLI/TUI and (designed-for) the webUI.
+
+**Immediate tangible goal:** the **"Mos"** orchestrator agent running on `w-jarvis`, reachable
+in **Discord channel `1517622518662434996`** (server `1112631390438166618`). Once the fleet is
+functional, we use the fleet itself to continue the work.
+
+## Requirements
+
+### A. Configure-without-AI CLI
+| ID | Requirement |
+|---|---|
+| R1 | `mosaic fleet` command set is functional end-to-end (init/install/start/stop/status/ps/verify + agent verbs). |
+| R2 | `mosaic fleet init` is an interactive, **AI-free** CLI wizard. |
+| R3 | Init asks the **configuration type**: `general`, `coding`, `research`, `hybrid`, … (extensible). |
+| R4 | Based on the answer, the fleet is populated with a **recommended set of agents** (a preset). |
+| R5 | **Exactly one main orchestrator agent** is always configured, regardless of type. |
+| R10 | A set of **recommended configurations (presets)** ships for easy duplication. |
+| R8 | User can **re-create** the fleet when config needs change (idempotent re-init / reconfigure). |
+| R17 | Fleet controls are **simple and intuitive**. |
+
+### B. Comms & orchestrator chat-ops
+| ID | Requirement |
+|---|---|
+| R6 | Init can wire the orchestrator to a chat connector — **Telegram / Discord / Matrix / Slack** — for command + comms. |
+| R7 | Designed with the end-goal of **Matrix comms on a locally-controlled server**. |
+| R16 | Fleet supports **tmux AND Matrix** comms, **user-configurable** at init or any time. Not all users want Matrix. |
+| R19 | **"Mos" orchestrator on Discord** (`chan 1517622518662434996` / `srv 1112631390438166618`) on `w-jarvis` — the first live target. |
+
+### C. Runtime, health, lifecycle
+| ID | Requirement |
+|---|---|
+| R9 | Fleet is **mutable by the orchestrator agent** — add/remove agents per need. |
+| R13 | Fleet **gracefully handles Pi + Claude harness updates** — keep harnesses current. |
+| R14 | The **Pi harness is customized** for proper tool usage, etc. |
+| R15 | **Agent heartbeat** properly configured for **Claude AND GPT/Pi** agents. |
+
+### D. Surfaces, testing, docs
+| ID | Requirement |
+|---|---|
+| R18 | Fleet built so the **webUI can view / monitor / terminate / butt-in** on a session. |
+| R11 | Installed and **tested on both `w-jarvis` and `dragon-lin`**. |
+| R12 | **Documentation**: how to install, configure, and use the fleet. |
+
+## Architecture / approach
+
+- **Config model:** `roster.yaml` is the source of truth (already exists). Add **presets** (`general`/`coding`/`research`/`hybrid`) as shipped example rosters; `init` selects a preset, always injects the orchestrator, and writes the roster. Re-init = regenerate roster (preserve user/site overrides — mirrors install env-merge from #567).
+- **Orchestrator agent:** always present; carries the chat connector config (connector type + target IDs) so it can be commanded over chat. tmux is the substrate; the connector bridges chat ↔ the orchestrator session.
+- **Comms layers (R16):** (1) **tmux** inter-agent (`agent-send`, proven) — default, always available. (2) **chat connector** for human↔orchestrator (Discord now; Matrix the strategic target). (3) **Matrix** as the locally-controlled cross-agent bus (future). Connector is pluggable + reconfigurable.
+- **Heartbeat (R15):** runtime-agnostic launcher sidecar already covers pi/claude/codex (#584). Refine per-runtime (native HB) with the **custom Pi harness** (R14) + a Claude path.
+- **Updates (R13):** `mosaic update` (CLI) + a fleet-aware harness-update step that refreshes pi/claude/codex and re-launches agents safely (drain → update → relaunch via the durable launcher).
+- **webUI (R18):** the fleet exposes machine-readable state (`fleet ps --json` already carries tenant/host/heartbeat/managed) + control verbs (start/stop/watch/send); webUI consumes these (control plane rides federation per north star). Ensure a stable JSON contract + a terminate/attach(butt-in) path.
+
+## Phases (incremental, each shippable)
+
+| Phase | Deliverable | Notes |
+|---|---|---|
+| **F1 Presets + init wizard** | preset rosters (general/coding/research/hybrid) + always-orchestrator + AI-free `fleet init` selecting a preset; re-init idempotent | R1–R5, R8, R10, R17 |
+| **F2 Connector + Mos-on-Discord** | orchestrator chat-connector config (Discord first) + **Mos live on Discord `1517…`/`1112…`** on w-jarvis | R6, R19, partial R16 |
+| **F3 Heartbeat + harness** | HB confirmed for claude + pi/gpt; **custom Pi harness** (tool usage, native HB, model self-report); graceful harness updates | R13, R14, R15 |
+| **F4 Matrix + comms toggle** | Matrix connector (local server) + user toggle tmux/Matrix at init/anytime | R7, R16 |
+| **F5 Orchestrator-mutable fleet** | orchestrator can add/remove agents at runtime | R9 |
+| **F6 webUI hooks** | stable JSON contract + terminate/attach surface for webUI view/monitor/terminate/butt-in | R18 |
+| **F7 Test + docs** | install+test on w-jarvis AND dragon-lin; user docs (install/configure/use) | R11, R12 (runs alongside every phase) |
+
+## Work division (proposed — confirm with dragon-lin)
+
+- **Jarvis @ w-jarvis (Lead):** F1 presets+wizard, F2 connector+Mos-on-Discord, F5 mutability, F6 webUI hooks; merge authority + dual-engine reviews; co-testing on w-jarvis.
+- **coder @ dragon-lin:** F3 custom Pi harness + harness-update flow (pi/codex-savvy); plus its in-flight constitution P4–P6 (P4 installer rework underpins `fleet init`/updates — coordinate the install path). Co-testing on dragon-lin (R11).
+- **Shared:** F4 Matrix (whoever has bandwidth); F7 testing/docs continuous.
+
+## Immediate target: Mos on Discord (F2 first slice)
+
+The discord plugin is available (`~/.claude.json`). Path: configure the **orchestrator** as a durable
+fleet session running Claude Code with the discord plugin bridged to channel `1517622518662434996`
+(server `1112631390438166618`) on w-jarvis, with the existing Discord Bridge Protocol (ack within
+~3s, reply via `mcp__discord__reply`, no `AskUserQuestion`). Heartbeat via the launcher sidecar.
+
+## Success criteria
+
+- A non-AI user can `mosaic fleet init`, pick a type, and get a working fleet + orchestrator.
+- **Mos answers in Discord `1517…`** on w-jarvis.
+- Fleet runs + is observable (`fleet ps`) on **both** w-jarvis and dragon-lin.
+- Harness updates handled gracefully; HB healthy for claude + pi/gpt agents.
+- Docs let a new operator install/configure/use the fleet.
+- Re-init + orchestrator mutation work.
+
+## Assumptions (veto-able)
+
+- `ASSUMPTION:` presets ship as example rosters under the framework (`fleet/examples/*.yaml`), selected by `init`.
+- `ASSUMPTION:` chat connectors are pluggable; Discord first (target exists), Matrix is the strategic default later.
+- `ASSUMPTION:` "Mos" = a Claude Code orchestrator session with the discord plugin (reuses the documented Discord Bridge Protocol).
+- `ASSUMPTION:` per north star, runtimes default to Codex/pi-on-Codex for workers; the orchestrator "Mos" runs Claude Code (in Claude Code, which is allowed).

packages/mosaic/framework/tools/fleet/start-agent-session.sh

@@ -6,7 +6,7 @@ MOSAIC_TMUX_SOCKET=${MOSAIC_TMUX_SOCKET:-mosaic-factory}
 MOSAIC_AGENT_RUNTIME=${MOSAIC_AGENT_RUNTIME:-pi}
 MOSAIC_AGENT_WORKDIR=${MOSAIC_AGENT_WORKDIR:-$HOME}
 MOSAIC_AGENT_COMMAND=${MOSAIC_AGENT_COMMAND:-}
-MOSAIC_HEARTBEAT_RUN_DIR=${MOSAIC_HEARTBEAT_RUN_DIR:-$HOME/.config/mosaic/fleet/run}
+MOSAIC_HEARTBEAT_RUN_DIR=${MOSAIC_HEARTBEAT_RUN_DIR:-${MOSAIC_HOME:-$HOME/.config/mosaic}/fleet/run}
 MOSAIC_HEARTBEAT_INTERVAL=${MOSAIC_HEARTBEAT_INTERVAL:-15}
 
 if [ -z "$AGENT_NAME" ]; then

packages/mosaic/framework/tools/fleet/test-start-agent-session.sh

@@ -32,8 +32,15 @@ MOSAIC_AGENT_COMMAND='bash --noprofile --norc -i' \
   "$START" "$AGENT"
 
 tmux -L "$SOCKET" has-session -t "=$AGENT:0.0" || fail "agent session was not created"
+# Retry: pane_current_path briefly reflects the tmux server's cwd until the pane
+# process establishes its own cwd (the -c start dir). Poll until it settles.
+actual_dir=""
+for _ in $(seq 1 30); do
   actual_dir=$(tmux -L "$SOCKET" display-message -p -t "=$AGENT:0.0" '#{pane_current_path}')
-[ "$actual_dir" = "$WORKDIR" ] || fail "agent workdir mismatch: $actual_dir"
+  [ "$actual_dir" = "$WORKDIR" ] && break
+  sleep 0.1
+done
+[ "$actual_dir" = "$WORKDIR" ] || fail "agent workdir mismatch: $actual_dir (expected $WORKDIR)"
 
 # ── Test 2: idempotency (duplicate start prints 'already running') ─────────────
 MOSAIC_TMUX_SOCKET="$SOCKET" \

packages/mosaic/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mosaicstack/mosaic",
-  "version": "0.0.36",
+  "version": "0.0.37",
   "repository": {
     "type": "git",
     "url": "https://git.mosaicstack.dev/mosaicstack/stack.git",

packages/mosaic/src/commands/fleet.spec.ts

@@ -856,6 +856,23 @@ describe('fleet ps — heartbeat parsing', () => {
     expect(hb.health).toBe('unknown');
     expect(hb.ts).toBeNull();
   });
+
+  it('honors MOSAIC_HEARTBEAT_INTERVAL for the freshness threshold', () => {
+    const prev = process.env.MOSAIC_HEARTBEAT_INTERVAL;
+    try {
+      // A 60s-old beat is STALE at the default 15s interval (3x15 = 45s)...
+      const ts = new Date(NOW - 60_000).toISOString();
+      const content = `ts=${ts}\npid=1\nstatus=ok\n`;
+      delete process.env.MOSAIC_HEARTBEAT_INTERVAL;
+      expect(parseHeartbeat(content, NOW).health).toBe('stale');
+      // ...but HEALTHY when the operator widened the interval to 30s (3x30 = 90s).
+      process.env.MOSAIC_HEARTBEAT_INTERVAL = '30';
+      expect(parseHeartbeat(content, NOW).health).toBe('healthy');
+    } finally {
+      if (prev === undefined) delete process.env.MOSAIC_HEARTBEAT_INTERVAL;
+      else process.env.MOSAIC_HEARTBEAT_INTERVAL = prev;
+    }
+  });
 });
 
 describe('fleet ps — systemd show parsing', () => {

packages/mosaic/src/commands/fleet.ts

@@ -368,6 +368,16 @@ export function buildAgentTailCommand(
 // ---------------------------------------------------------------------------
 
 export const HEARTBEAT_INTERVAL_MS = 15_000;
+
+/**
+ * Heartbeat interval in ms, honoring MOSAIC_HEARTBEAT_INTERVAL (seconds) so the
+ * `fleet ps` freshness threshold matches the writer sidecar's actual cadence
+ * (start-agent-session.sh). Falls back to HEARTBEAT_INTERVAL_MS (15s).
+ */
+export function heartbeatIntervalMs(): number {
+  const sec = Number.parseInt(process.env.MOSAIC_HEARTBEAT_INTERVAL ?? '', 10);
+  return Number.isFinite(sec) && sec > 0 ? sec * 1000 : HEARTBEAT_INTERVAL_MS;
+}
 export const HEARTBEAT_HEALTHY_MULTIPLIER = 3;
 
 export interface HeartbeatInfo {
@@ -496,7 +506,7 @@ export function parseHeartbeat(content: string | null, nowMs = Date.now()): Hear
       status = val;
     }
   }
-  const thresholdMs = HEARTBEAT_INTERVAL_MS * HEARTBEAT_HEALTHY_MULTIPLIER;
+  const thresholdMs = heartbeatIntervalMs() * HEARTBEAT_HEALTHY_MULTIPLIER;
   let health: 'healthy' | 'stale' | 'unknown' = 'unknown';
   let ageMs: number | null = null;
   if (ts !== null) {
@@ -1428,15 +1438,19 @@ export function registerFleetAgentCommands(
 
       await runChecked(runner, buildAgentWatchCreateViewerCommand(agent, viewerName, socketName));
 
+      let exitCode = 0;
+      try {
         const [bin, args] = splitCommand(buildAgentWatchAttachCommand(viewerName, socketName));
-      const exitCode = await iRunner(bin, args);
+        exitCode = await iRunner(bin, args);
-
+      } finally {
-      // Best-effort cleanup of the viewer session regardless of how the user detached.
+        // ALWAYS clean up the viewer session — even if attach threw or the process was
-      // Errors here are intentionally suppressed — the agent session is unaffected.
+        // interrupted — so stale grouped *-watch-* sessions never accumulate. Errors here
+        // are intentionally suppressed; the agent session is unaffected.
         const killResult = await runner(
           ...splitCommand(buildAgentWatchKillViewerCommand(viewerName, socketName)),
         );
-      void killResult; // result is intentionally ignored
+        void killResult;
+      }
 
       if (exitCode !== 0) {
         process.exitCode = exitCode;