crewly 1.20.194 → 1.20.196

package/config/constants.ts

@@ -202,6 +202,14 @@ export const SAFE_RESTART_CONSTANTS = {
 	DRAIN_TIMEOUT_MS: 120_000,
 	/** Env var overriding DRAIN_TIMEOUT_MS; `0` disables the wait */
 	DRAIN_ENV_VAR: 'CREWLY_RESTART_DRAIN_MS',
+	/**
+	 * Hard cap on the drain while an agent it waits on has a tool call or a
+	 * subagent / background task still running (ms). Such work routinely
+	 * outlasts DRAIN_TIMEOUT_MS (2026-10-02, Eve).
+	 */
+	BACKGROUND_DRAIN_TIMEOUT_MS: 600_000,
+	/** Env var overriding BACKGROUND_DRAIN_TIMEOUT_MS */
+	BACKGROUND_DRAIN_ENV_VAR: 'CREWLY_RESTART_DRAIN_BACKGROUND_MS',
 	/**
 	 * Time allowed after the drain for the rest of the shutdown (session state
 	 * save, PTY teardown, queue flush) before a supervisor may SIGKILL (ms).
@@ -358,6 +366,66 @@ export const API_SECURITY_CONSTANTS = {
 	LOOPBACK_ADDRESSES: ['127.0.0.1', '::1', '::ffff:127.0.0.1'],
 } as const;
 
+/**
+ * Owner authentication (#999, specs/2026-10-03-owner-auth.md).
+ *
+ * Who a caller is comes from credentials, never from a missing header:
+ * - agents carry a per-session **agent badge** minted by the harness at launch;
+ * - the dashboard holds an owner session cookie plus a CSRF token;
+ * - the phone / portal relay and Cloud-forwarded Slack envelopes carry an
+ *   in-memory internal credential;
+ * - remote dashboards, the phone on the LAN and the CLI present the owner API token.
+ */
+export const OWNER_AUTH_CONSTANTS = {
+	/**
+	 * Environment variable holding the agent badge in every agent process.
+	 * Deliberately free of TOKEN / KEY / SECRET: Codex's default
+	 * `shell_environment_policy` strips variables whose names contain those
+	 * words from the shells it runs, so the skills would never see it.
+	 */
+	AGENT_BADGE_ENV: 'CREWLY_AGENT_BADGE',
+	/** Request header carrying the agent badge */
+	AGENT_BADGE_HEADER: 'x-agent-badge',
+	/** Version prefix of a badge (`cab1.<base64url session>.<hmac>`) */
+	AGENT_BADGE_PREFIX: 'cab1',
+	/** Owner session cookie name prefix; the backend port is appended (cookies ignore ports) */
+	OWNER_SESSION_COOKIE_PREFIX: 'crewly_owner_',
+	/** Version prefix of an owner session value */
+	OWNER_SESSION_PREFIX: 'cos1',
+	/** Lifetime of an owner session cookie (seconds). The signing secret is in memory, so a restart ends it sooner. */
+	OWNER_SESSION_MAX_AGE_S: 30 * 24 * 60 * 60,
+	/** Header the dashboard sends the CSRF token in on every write */
+	CSRF_HEADER: 'x-crewly-csrf',
+	/** Header carrying the in-memory internal credential (relay / cloud) */
+	INTERNAL_HEADER: 'x-crewly-internal',
+	/** Path (under `/api`) that issues the owner session and returns the CSRF token */
+	SESSION_ROUTE: '/auth/session',
+	/** HTTP methods that change state and therefore need CSRF / a header-borne token */
+	MUTATING_METHODS: ['POST', 'PUT', 'PATCH', 'DELETE'] as readonly string[],
+	/**
+	 * What a browser without a session (a dashboard tab opened before the
+	 * upgrade) is told in the 401's `error`, which that old tab shows as is.
+	 */
+	RELOAD_MESSAGE: 'Reload this page — Crewly was updated.',
+	/** Response header naming the dashboard build the backend serves (its hashed entry script) */
+	BUILD_HEADER: 'x-crewly-dashboard-build',
+	/** socket.io event announcing the served dashboard build on each connection */
+	BUILD_EVENT: 'dashboard_build',
+	/** Error codes in `{ success: false, error }` */
+	ERRORS: {
+		/** No owner credential (401). The dashboard refreshes its session and retries once. */
+		OWNER_AUTH_REQUIRED: 'owner_auth_required',
+		/** An agent called an owner-only route (403) */
+		OWNER_ONLY: 'owner_only',
+	},
+	/** Give up looking up a local client's process after this long (fail open) */
+	PEER_LOOKUP_TIMEOUT_MS: 3000,
+	/** Longest parent chain walked from a client process */
+	PEER_MAX_ANCESTRY_DEPTH: 64,
+	/** `X-Agent-Session` without a badge (migration window) is logged at most this often per session */
+	LEGACY_WARN_THROTTLE_MS: 10 * 60 * 1000,
+} as const;
+
 // ========================= TIMING CONSTANTS =========================
 
 /**

package/config/hooks/agent-status/report.sh

@@ -1,19 +1,23 @@
 #!/usr/bin/env bash
 # Crewly agent-status hook — tells the backend when Claude Code is waiting on
-# the user (a permission prompt or a question) and when it stops waiting.
+# the user (a permission prompt or a question) and when it stops waiting, and
+# where it is in its turn (turn start/end, tool calls, subagents).
 #
-# Spec: specs/2026-09-26-agent-waiting-on-human.md (#815, hook ingestion).
+# Spec: specs/2026-09-26-agent-waiting-on-human.md (#815, hook ingestion);
+# specs/2026-10-02-restart-busy-and-resume.md (runtime turn state).
 #
 # Wired by the backend into the per-session settings file it already passes
 # with `claude --settings <file>` (the control-plane guard's file) for:
-#   Notification, PermissionRequest, Stop, UserPromptSubmit, PostToolUse
+#   Notification, PermissionRequest, Stop, UserPromptSubmit, PreToolUse,
+#   PostToolUse, SubagentStart, SubagentStop, SessionStart
 #
 # Usage: bash report.sh            # hook JSON on stdin
 #
 # PRIVACY: the stdin JSON can hold tool_input, file contents, prompts and a
 # transcript path, any of which may contain secrets. This script extracts
-# exactly two TOP-LEVEL fields — hook_event_name and notification_type — keeps them only
-# if they are plain identifiers ([A-Za-z_], max 64 chars), and sends those plus
+# exactly five TOP-LEVEL fields — hook_event_name, notification_type and
+# source (kept only as plain identifiers, [A-Za-z_], max 64 chars), tool_use_id and
+# agent_id (kept only as [A-Za-z0-9_-], max 128 chars) — and sends those plus
 # the session name. Nothing else from stdin is sent, printed or logged.
 #
 # It never blocks or slows the agent: every path exits 0, and the POST has a
@@ -34,8 +38,9 @@ API_URL="${CREWLY_API_URL:-http://localhost:${WEB_PORT:-8787}}"
 # tool_input (a regex cannot: `"hook_event_name":"Stop"` inside a command would
 # spoof the event). So: jq, else node's JSON.parse, else nothing — the event is
 # dropped and the backend's screen detection still covers the state.
+NAME_PATTERN='^[A-Za-z_]{1,64}$'
 field() {
-	local name="$1" value=""
+	local name="$1" pattern="${2:-$NAME_PATTERN}" value=""
 	if command -v jq >/dev/null 2>&1; then
 		value="$(printf '%s' "$INPUT" | jq -r --arg k "$name" 'if type == "object" then (.[$k] // empty | select(type == "string")) else empty end' 2>/dev/null)"
 	elif command -v node >/dev/null 2>&1; then
@@ -51,7 +56,7 @@ field() {
 			});' "$name" 2>/dev/null)"
 	fi
 	# Keep only a plain identifier; anything else is discarded, not sanitised.
-	if printf '%s' "$value" | grep -Eq '^[A-Za-z_]{1,64}$'; then
+	if printf '%s' "$value" | grep -Eq "$pattern"; then
 		printf '%s' "$value"
 	fi
 }
@@ -59,17 +64,31 @@ field() {
 EVENT="$(field hook_event_name)"
 [ -z "$EVENT" ] && exit 0
 NOTIFICATION_TYPE="$(field notification_type)"
+ID_PATTERN='^[A-Za-z0-9_-]{1,128}$'
+TOOL_USE_ID=""
+AGENT_ID=""
+SOURCE=""
+case "$EVENT" in
+	PreToolUse|PostToolUse) TOOL_USE_ID="$(field tool_use_id "$ID_PATTERN")" ;;
+	SubagentStart|SubagentStop) AGENT_ID="$(field agent_id "$ID_PATTERN")" ;;
+	SessionStart) SOURCE="$(field source)" ;;
+esac
 
-if [ -n "$NOTIFICATION_TYPE" ]; then
-	BODY="{\"event\":\"$EVENT\",\"notificationType\":\"$NOTIFICATION_TYPE\"}"
-else
-	BODY="{\"event\":\"$EVENT\"}"
-fi
+BODY="{\"event\":\"$EVENT\""
+[ -n "$NOTIFICATION_TYPE" ] && BODY="$BODY,\"notificationType\":\"$NOTIFICATION_TYPE\""
+[ -n "$TOOL_USE_ID" ] && BODY="$BODY,\"toolUseId\":\"$TOOL_USE_ID\""
+[ -n "$AGENT_ID" ] && BODY="$BODY,\"agentId\":\"$AGENT_ID\""
+[ -n "$SOURCE" ] && BODY="$BODY,\"source\":\"$SOURCE\""
+BODY="$BODY}"
 
 ARGS=(-s -o /dev/null --max-time 2 -X POST "$API_URL/api/agent-hooks"
 	-H "Content-Type: application/json"
 	-H "User-Agent: crewly-agent-status-hook/1"
 	-H "X-Agent-Session: $SESSION")
+# The agent badge (#999): the credential behind X-Agent-Session.
+if [ -n "${CREWLY_AGENT_BADGE:-}" ]; then
+	ARGS+=(-H "X-Agent-Badge: $CREWLY_AGENT_BADGE")
+fi
 if [ -n "${CREWLY_AGENT_AUTHORIZATION:-}" ]; then
 	ARGS+=(-H "X-Agent-Authorization: b64:$(printf '%s' "$CREWLY_AGENT_AUTHORIZATION" | base64 | tr -d '\n')")
 fi

package/config/hooks/subagent-guard/subagent.sh

@@ -108,6 +108,30 @@ process.stdin.on("end", () => {
 JS
 )"
 
-printf '%s' "$INPUT" | MARKER_DIR="$MARKER_DIR" node -e "$PROGRAM" 2>/dev/null
+OUT="$(printf '%s' "$INPUT" | MARKER_DIR="$MARKER_DIR" node -e "$PROGRAM" 2>/dev/null)"
+printf '%s' "$OUT"
+
+# A subagent was sent back: tell the backend so the run's trace records it
+# (autonomy metrics, #984). Only a fixed event name and the session leave the
+# machine's hook; nothing from the transcript or stdin is sent. Backgrounded
+# (2-second ceiling, failures ignored), so the stop never waits on it.
+case "$OUT" in
+	*'"decision":"block"'*)
+		SESSION="${CREWLY_SESSION_NAME:-}"
+		if [ -n "$SESSION" ] && command -v curl >/dev/null 2>&1; then
+			API_URL="${CREWLY_API_URL:-http://localhost:${WEB_PORT:-8787}}"
+			ARGS=(-s -o /dev/null --max-time 2 -X POST "$API_URL/api/agent-hooks"
+				-H "Content-Type: application/json"
+				-H "User-Agent: crewly-subagent-guard-hook/1"
+				-H "X-Agent-Session: $SESSION")
+			if [ -n "${CREWLY_AGENT_AUTHORIZATION:-}" ]; then
+				ARGS+=(-H "X-Agent-Authorization: b64:$(printf '%s' "$CREWLY_AGENT_AUTHORIZATION" | base64 | tr -d '\n')")
+			fi
+			# In the background: a subagent stop never waits on the backend.
+			curl "${ARGS[@]}" --data '{"event":"SubagentSendBack"}' >/dev/null 2>&1 &
+			disown 2>/dev/null || true
+		fi
+		;;
+esac
 
 exit 0

package/config/roles/auditor/prompt.md

@@ -1,5 +1,3 @@
-# Crewly Auditor — System Prompt
-
 ## Crewly Operating Principles
 
 1. Outcome over activity.
@@ -11,6 +9,8 @@
 
 ---
 
+# Crewly Auditor — System Prompt
+
 You are the **Crewly Auditor**, an autonomous quality observer for the Crewly multi-agent orchestration platform.
 
 ## Your Mission

package/config/roles/orchestrator/fragments/communication.md

@@ -140,5 +140,5 @@ Always reach for `{{ORCHESTRATOR_SKILLS_PATH}}/<skill>/` first. Orc-namespaced e
 
 **Bash invocation example (orc-namespaced send-message):**
 ```bash
-bash {{ORCHESTRATOR_SKILLS_PATH}}/send-message/execute.sh '{"to":"<session>","message":"<msg>"}'
+bash {{ORCHESTRATOR_SKILLS_PATH}}/send-message/execute.sh '{"sessionName":"<session>","message":"<msg>"}'
 ```

package/config/roles/orchestrator/fragments/lifecycle.md

@@ -12,7 +12,7 @@ Do NOT create monitoring WorkItems for yourself. Delegations are already covered
 |----------|---------------|----------|
 | 🔴 **Critical** — Notify IMMEDIATELY | Agent crash, task failure, blocked, error | Runtime exited, build failed, agent stuck >15min |
 | 🟡 **Important** — Notify within 1 min | Task completed, needs user decision, milestone reached | Agent finished feature, needs review approval |
-| ⚪ **Info** — Log only, include in next summary | Agent started working, routine status change, heartbeat | idle→in_progress, scheduled check with no changes |
+| ⚪ **Info** — Log only, do not notify (mention it only if the owner asks, or in the daily summary when Onboarding Mode is on) | Agent started working, routine status change, heartbeat | idle→in_progress, scheduled check with no changes |
 
 ### Decision Rules for Events